A Multi-Stage Chained Task Allocation Method for Maritime Search and Rescue Based on Priority-Aware Two-Layer Collaborative Decision Making
By employing a priority-aware, two-layer collaborative decision-making approach, the problems of insufficient chain dependency modeling, inadequate consideration of cross-vessel handover costs, lack of urgency awareness, and insufficient load balancing in maritime search and rescue mission allocation were solved. This approach enabled efficient and rapid mission allocation and system recovery, thereby improving mission success rate and responsiveness.
Patent Information
- Application Number
- CN202611270057.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing maritime search and rescue mission allocation methods suffer from several problems in complex marine environments, including insufficient chain dependency modeling, inadequate consideration of cross-vessel handover costs, lack of urgency awareness, lack of load balancing, and insufficient systemic recovery capabilities under extreme disturbances. These issues result in low mission success rates and untimely responses.
We adopt a priority-aware two-layer collaborative decision-making method, evaluate the criticality of tasks through graph attention networks, construct a cost matrix and refine it using Hungarian centralized warm start and distributed CBBA consensus, combine superlinear load penalty and urgency decay factor to achieve global optimality and distributed fault tolerance, design chain co-localization and spatial collaborative reward redistribution mechanism, and construct a strategic hierarchical adaptive recovery protocol.
It enables efficient and rapid task allocation in complex marine environments, improves task completion rate, reduces inter-vessel handover delay, ensures load balancing and system recovery capability under extreme disturbances, and meets the response requirements of emergency missions.
Smart Images

Figure CN123117628A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned autonomous collaborative control technology, and in particular to a multi-stage chain-like task allocation method for maritime search and rescue based on priority-aware two-layer collaborative decision-making. Background Technology
[0002] Today, maritime emergencies and personnel distress incidents are becoming increasingly frequent. Compared to land-based environments, maritime search and rescue faces multiple severe challenges, including vast geographical areas, harsh hydrological conditions, and uncontrollable target drift. How to rapidly and accurately deploy extremely limited search and rescue forces to the vast area of the incident within a very short time window, and minimize the response time from the occurrence of the emergency to the first contact with the rescue team, has become a core issue in the construction of the current maritime search and rescue system and the improvement of emergency support capabilities.
[0003] In recent years, with the rapid development of unmanned system control and collaborative communication technologies, unmanned surface vessels (USVs), unmanned aerial vehicles (UAVs), and unmanned underwater vehicles (UUVs) have become important vehicles for carrying out large-scale, high-risk maritime search and rescue missions due to their advantages such as high mobility, low operating costs, long endurance, rapid large-scale deployment, and zero risk of personnel casualties. Compared with single traditional manned vessels, unmanned swarm collaborative operations can simultaneously execute multiple sub-tasks in vast and dangerous sea areas, effectively improving the breadth and efficiency of search and rescue; multi-domain collaboration, such as air-sea collaboration and surface-underwater collaboration, further breaks through the physical limitations of single equipment in terms of detection perspective and operational capabilities.
[0004] However, maritime search and rescue missions are not simply single-target access problems, but typically involve a chain of operations comprised of multiple stages. Compared to general task allocation problems, chained maritime search and rescue missions have three prominent characteristics: First, there is a strict dependency between successive stages, with downstream tasks requiring the completion of upstream tasks before execution; second, inter-vessel handover introduces physical waiting and information synchronization delays, potentially causing subsequent tasks to miss their time windows; and third, changes in sea state, communication constraints, platform failures, and unforeseen emergencies continuously alter the mission map and the status of available resources. Therefore, achieving efficient allocation, rapid response, and dynamic recovery of chained tasks within a limited time window is a key issue in the scheduling of unmanned surface vessel (USV) swarms for maritime search and rescue.
[0005] Existing research on cooperative task allocation for multi-agent systems has the following main shortcomings: (1) Existing task allocation methods for maritime search and rescue mostly focus on centralized architecture or basic biomimetic algorithms. Early maritime search and rescue task scheduling relied on centralized mathematical programming or heuristic algorithms. However, such centralized methods are prone to single-point failures and communication overload when facing weak communication environments in deep seas or when extreme interference causes damage to the central node. They are difficult to meet the stringent requirements of high survivability and high robustness for maritime search and rescue.
[0006] (2) Representative algorithms for distributed task allocation, namely the consensus binding algorithm CBBA and its improved algorithms, have made some progress in dealing with the horizontal constraints of time, space and capability. However, the above work is limited to single-layer auction logic and has not yet systematically dealt with the physical waiting and urgency decay caused by the cross-platform handover of multi-stage tasks. This is the core difficulty that distinguishes maritime search and rescue missions from general unmanned system scheduling.
[0007] (3) In the face of the dimensional disaster and communication congestion brought about by the explosion of mission scale, existing research has proposed several hierarchical architectures and adaptive redistribution methods. However, the above two-layer work mostly adopts the serial paradigm of pre-clustering dimensionality reduction plus single-layer CBBA, which is essentially a hierarchical decomposition of the scaled dimension. In the extreme scenario of sudden failure of a large proportion of nodes or superposition of multiple disturbances in maritime search and rescue, there is still a lack of systematic solutions on how to achieve the complementarity of high-quality centralized initial allocation and distributed fault tolerance in the performance-robust dimension.
[0008] (4) The adaptive redistribution mechanism under extreme dynamic disturbances is imperfect. In response to problems such as sudden failure of agents, weak underwater acoustic communication constraints, and highly dynamic complex environments, existing studies have proposed methods such as local dynamic redistribution based on virtual managers, local communication conflict resolution mechanism, and prediction-based two-stage distributed task allocation. However, these methods are mostly designed for single-type disturbances and lack a systematic recovery mechanism to deal with multiple disturbance scenarios such as high proportion of sudden USV failures, sudden changes in sea state, and sudden task injection. When a large number of orphan tasks are generated, the local redistribution method is difficult to restore the task chain structure in a timely manner.
[0009] (5) The progressive priority response mechanism for emergency tasks is insufficient. Traditional auction or allocation algorithms usually treat the time window as a hard constraint and lack an urgency perception mechanism that gradually increases as the remaining time decreases, making it difficult to meet the rapid response requirements of life-threatening tasks such as MAYDAY.
[0010] (6) Insufficient consideration of long-term load balancing and continuous deployment capabilities. Some algorithms excessively pursue local gains or single-time completion rates, which may lead to long-term overload of a few high-capacity platforms, reducing the overall endurance of the cluster and the response capability of subsequent tasks.
[0011] Furthermore, existing patent literature has also explored related issues. For example, Chinese patent application publication number CN116149370A improves the CBBA algorithm through clustering to enhance task allocation efficiency, but its clustering strategy is mainly based on spatial distance and does not comprehensively consider multi-dimensional factors such as task priority, time constraints, and execution capability, nor does it involve chain dependency modeling. Chinese patent application publication number CN116993073A introduces constraints such as time windows into the CBBA framework, improving the algorithm's ability to handle task timing requirements, but it still adopts a single optimization objective and does not establish a task redistribution mechanism under dynamic environments. Chinese patent application publication number CN114610058A proposes a task allocation method based on improved CBBA, introducing a task timing hierarchical mechanism and dynamic soft time window constraints, but it still takes minimizing the average execution time as a single objective and lacks real-time response to dynamic events during execution. Chinese patent application publication number CN121189684A uses an improved consensus binding algorithm combined with an intelligent optimization strategy to achieve task allocation. However, its HDBSCAN adaptive density clustering method is prone to causing severe overload of some clusters in scenarios with uneven task distribution. Furthermore, it uses soft constraint logic for the time window, which cannot completely eliminate physically unexecutable schemes. Its outer wrapper genetic algorithm optimization method has a large computational cost and poor real-time performance.
[0012] In summary, existing technologies still suffer from several shortcomings when dealing with the multi-stage task allocation problem in maritime search and rescue involving highly dynamic, strongly constrained, chain-dependent, and heterogeneous platforms. These include insufficient modeling of chain dependencies and cross-vessel handover costs, a lack of urgency perception mechanisms, inadequate load balancing considerations, and imperfect systemic recovery mechanisms under extreme disturbances. Therefore, there is an urgent need to research an intelligent scheduling method that integrates priority perception, two-layer collaborative decision-making, urgency attenuation, superlinear load balancing, and hierarchical failure recovery. This method aims to achieve high-quality and robust allocation of chained tasks for heterogeneous unmanned surface vessel (USV) swarms, meeting the practical application needs of maritime emergency search and rescue in complex marine environments. Summary of the Invention
[0013] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a multi-stage chained task allocation method for maritime search and rescue based on priority-aware two-layer collaborative decision-making. This method aims to solve or partially solve the problems of insufficient chain dependency modeling, inadequate consideration of cross-vessel handover costs, lack of urgency awareness, lack of load balancing, and insufficient system recovery capability under extreme disturbances in existing maritime search and rescue task allocation methods, which result in low mission success rate and untimely response in complex search and rescue scenarios.
[0014] The objective of this invention can be achieved through the following technical solutions: This invention provides a multi-stage chained task allocation method for maritime search and rescue based on priority-aware two-layer collaborative decision-making, comprising the following steps: Obtain the status data, task environment data, and chained dependencies of each USV in the heterogeneous unmanned surface vessel (USV) cluster, and construct a constrained multi-stage chained task dependency graph. Using the multidimensional attributes of task nodes as input, a pre-trained graph attention network is used to perform forward reasoning on the multi-stage chained task dependency graph, and output the global criticality score of each task node. For any USV and task pair, a cost matrix is constructed by introducing a criticality discount term, a second-order attenuation urgency factor AUD discount term, and a superlinear load penalty WEB term, with travel time as the basic cost term. Hungarian centralized warm start: Perform minimum cost bipartite graph matching on the cost matrix, repeating multiple times until no new allocation is generated, to generate an initial allocation scheme from a global perspective; Distributed CBBA consensus refinement: Based on the initial allocation scheme, each USV acts as a distributed agent, eliminating inter-vessel allocation conflicts through a bounded round-by-round win-bid consensus process, resulting in an optimized allocation scheme.
[0015] As a preferred technical solution, the calculation process of the global criticality score includes the following steps: The task dependency graph is forward propagated using a multi-dimensional node feature matrix consisting of the required capabilities, priority, time window, spatial coordinates, and the number of predecessors and successors in the task chain. The Sigmoid function is used to output the criticality score of each task node. The graph attention network is trained offline using the marginal contribution of the task obtained from Monte Carlo simulation to the global strict task completion rate as the supervision label.
[0016] As a preferred technical solution, the process of constructing the cost matrix includes the following steps: For any USV and mission pair, the Euclidean distance between them divided by the maximum speed of the USV is used as the basic travel time cost. Calculate the GNN criticality discount term, the value of which is proportional to the difference between the global criticality score and 0.5; The urgency is defined as a cutoff value minus the ratio of the remaining time of the time window to the width of the time window, and the secondary decay urgency factor AUD discount term is calculated by multiplying the square of the urgency by the priority weight. The urgency is zero when the current time is earlier than the start time of the time window, one when it reaches or exceeds the cutoff time, and monotonically increases when it is in between. The superlinear load penalty WEB term is calculated as a power of 1.5 of the current USV load ratio; Add the basic driving time cost term to the superlinear load penalty WEB term, and then subtract the GNN criticality discount term and the secondary decay urgency factor AUD discount term to obtain the cost matrix elements; If the USV does not have the capabilities required for the mission or exceeds the maximum cruising radius, the corresponding cost matrix element is set to positive infinity.
[0017] As a preferred technical solution, the process of generating an initial allocation scheme from a global perspective includes the following steps: Minimum cost bipartite graph matching is performed on the cost matrix, and the process is repeated at most until the upper limit of the USV's binding capacity is reached. In each round, the matched tasks are removed from the candidate set and the cost matrix is reconstructed. The process continues until no new allocation is generated in the current round, at which point the process exits early and outputs the matching results of each round as the initial allocation scheme.
[0018] As a preferred technical solution, the process of eliminating inter-vessel allocation conflicts through a bounded round-based consensus-building process includes the following steps: In each round of consensus, each pair of USVs exchanges their winning dictionaries. For any mission currently bundled with the first USV, if the second USV makes a higher bid and considers itself the winner, the first USV releases the mission and updates its local winning bid information to the second USV's bid; Convergence occurs after one round of no task transfer, and the maximum number of consensus rounds is set to a preset value. If conflicts still exist after reaching the maximum number of consensus rounds, a definitive decision will be made based on task priority, bid value, and vessel number.
[0019] As a preferred technical solution, after obtaining the optimized allocation scheme, the process also includes chain co-location and spatial collaborative reward redistribution, comprising the following steps: For subsequent tasks in the same task chain, if the marginal cost of the lead ship accepting the subsequent task is lower, provided that the feasibility constraints of unfulfilled load, capacity matching, and reachability are met, then the subsequent task will be migrated from its current USV to the lead ship. The marginal cost is compared by comparing the travel time from the lead ship to the subsequent task with the travel time from the current USV of the subsequent task to the subsequent task plus the basic co-location tolerance threshold plus the spatial cooperation reward. The spatial cooperation reward is proportional to the number of tasks currently bundled with the lead ship that are located within the preset neighborhood radius of the subsequent task.
[0020] As a preferred technical solution, after obtaining the optimized allocation scheme, a MAYDAY preemption process is also included: When there is an unassigned MAYDAY emergency task, iterate through the non-executing tasks in the bundle of each USV that has the capability required for the task and has a lower priority than the MAYDAY task. Select the scheme with the largest preemption benefit and a positive value according to the preemption benefit function, so as to release the low-priority task to create execution slots for the unassigned MAYDAY task. The preemption benefit function is equal to the product of the priority weight of the MAYDAY task and the basic reward value, plus twice the response time saved by preemption, minus half the product of the priority weight of the expelled task and the basic reward value, and the travel cost of the candidate USV to the MAYDAY task. The selected MAYDAY task is inserted at the beginning of the preemption boat path.
[0021] As a preferred technical solution, after obtaining the optimized allocation scheme, a strategic hierarchical adaptive recovery process is also included. When a USV failure results in an orphan task, the following steps are performed: For each combination of orphan tasks and candidate activity USVs, a comprehensive score is calculated by weighting the negative factors of travel time, negative factors of load factor, positive factors of capability specificity, positive factors of high-priority task speed matching reward, positive factors of task priority, and positive factors of time margin reward. Candidate USVs are divided into three levels based on time feasibility and cruise safety, and the candidate USV with the highest comprehensive score is selected from the highest level. Orphan tasks within the same chain are arranged in phase order and processed in batches according to the order within the chain, prioritizing the maintenance of locality within the chain; When the number of assigned tasks in a target USV reaches its bundled capacity limit, the lowest priority non-executing task in the USV bundle is actively expelled to a neighboring USV. Maintain a USV failure count counter within a preset sliding time window. When the counter is less than a first threshold, execute the normal recovery mode. When the counter is not less than the first threshold and is less than a second threshold, execute the medium rebalancing mode to migrate the lowest priority task in the overloaded USVs whose load exceeds the average load of the fleet by more than one task to the lightly loaded USVs. When the counter is not less than the second threshold, execute the aggressive release mode to release the low-priority assigned tasks in the overloaded USVs to the pending execution state.
[0022] As a preferred technical solution, after obtaining the optimized allocation scheme, the process also includes topology normalization and scheme output: Perform a topology sort that preserves order on the path of each USV, ensuring that the position index of any preceding mission is strictly less than the position index of the following mission; re-verify the feasibility of the time window and cruise radius, output the final allocation scheme; and trigger the next round of asynchronous reallocation when a new mission is injected, a USV fails, or the mission map changes.
[0023] As a preferred technical solution, in the process of constructing the multi-stage chained task dependency graph, the USV's state data includes initial position coordinates, maximum speed, maximum cruising radius, capability set, and bundled capacity limit; the task environment data includes task identifier, task type, required capabilities, task location, estimated execution duration, time window start and end times, priority level, task chain number, and stage position index. The priority levels are divided into four levels, corresponding to MAYDAY emergency rescue, PAN-PAN detection / location, towing / patrol, and general patrol / scanning, respectively. A bounded hierarchical weight mapping is adopted so that the four priority levels correspond to four priority weights that increase sequentially, and the ratio of the highest priority weight to the lowest priority weight does not exceed two orders of magnitude. The chain-like dependency relationship is modeled using a directed graph. For tasks belonging to the same chain and with a stage index earlier, there are directed dependency edges pointing to tasks with a stage index later.
[0024] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Balancing Global Optimality and Distributed Fault Tolerance: Existing methods lack effective coordination between global optimization and distributed fault tolerance, resulting in low computational efficiency and poor real-time performance in large-scale scenarios. This invention adopts a two-layer collaborative decision-making architecture combining Hungarian centralized warm-start with distributed CBBA bounded consensus refinement. The Hungarian warm-start phase provides a high-quality global perspective initial solution, and the distributed consensus phase eliminates conflicts and maintains distributed properties based on this. This architecture achieves complementary coupling of Hungarian global optimization and CBBA distributed fault tolerance in the performance-robust dimension, which is different from the existing serial paradigm of pre-clustering dimensionality reduction plus CBBA. Under the maximum test scale of 180 tasks / 45 boats, the average time of a single allocation is compressed to 146.6 milliseconds, which is far lower than the real-time computation budget allowed by the communication cycle. Compared with the standard CBBA and distributed baseline algorithms such as CCBBA with coupling constraints, the computation time is reduced by one to two orders of magnitude, while maintaining a high task completion rate.
[0025] (2) Improving Emergency Task Response by Identifying Bottleneck Tasks: Existing methods lack sufficient chain dependency modeling, making it difficult to identify key bottleneck tasks that affect the global completion rate, leading to delays in emergency task response. This invention introduces a pre-trained lightweight graph attention network (GAT) to perform global criticality scoring on the task dependency graph, identifying bottleneck task nodes that have a high marginal contribution to the global strict task completion rate. The model has only about 6,000 parameters, and the correlation coefficient between the prediction and ground truth on the validation set reaches about 0.97, enabling real-time inference on edge embedded platforms in a CPU-only manner. The GNN score, as a discount term in the cost matrix, guides the Hungarian matching stage to prioritize bottleneck task nodes, making it easier for key nodes in the task graph to be taken over by appropriate USVs. Ablation experiments show that the MAYDAY response time (MRT) increases significantly after removing the GNN, verifying the important value of criticality scoring for emergency task response efficiency.
[0026] (3) Achieving Progressive Priority Response: Existing methods lack urgency awareness, and traditional methods interfere too early when the task is far from the deadline, or the response strength is insufficient when the deadline is approaching. This invention designs a quadratic decay urgency factor (AUD) discount term with monotonically non-decreasing, marginal acceleration, and upper bound analytic properties, which has the response characteristics of being mild in the early stage and rapidly enhancing when the deadline is approaching. This mechanism satisfies three analytic properties: monotonically non-decreasing, marginal acceleration with a second derivative greater than zero, and upper bound: the discount is small when the task is far from the deadline, avoiding premature disturbance to the overall matching; the discount increases rapidly when the task is close to the deadline, so that the algorithm prioritizes high-priority tasks that are about to time out. In the static scale expansion experiment, the MRT of this invention remains in the range of 78 to 87 minutes in the static scenario, which is highly matched with the IAMSA golden window of 60 to 90 minutes reference response range.
[0027] (4) Ensuring Fleet Load Fairness: Existing technologies lack load balancing, with a few high-capacity USVs experiencing prolonged overload, increasing endurance pressure and failure risks. This invention employs a superlinear load penalty (WEB) term based on the current load ratio raised to the power of 1.5 to construct a cost matrix. Compared to linear penalties, this changes more slowly during low load phases, not affecting normal mission reception; however, the penalty increases rapidly when the load exceeds medium to high levels, thus guiding the Hungarian matching phase to allocate new missions to relatively lightly loaded USVs. In a P2 static scale special experiment with 60 missions / 15 vessels, after enabling WEB, the load Gini coefficient decreased from 0.322 to 0.254, and the load imbalance ratio η decreased from 7.76 to 6.12, representing a relative improvement of over 21%; simultaneously, the SCR increased by approximately 2.4%. WEB effectively avoids the endurance pressure and failure risks caused by prolonged overload of a few high-capacity platforms.
[0028] (5) Reduce cross-ship handover: Existing methods do not adequately consider the cost of cross-ship handover. Chained tasks frequently hand over between different USVs, introducing physical waiting and communication delays. This invention implements a chain co-location and spatial cooperative reward redistribution mechanism. It dynamically adjusts the migration threshold by utilizing the number of tasks in the neighborhood of the chain leader's path. When multiple tasks in the chain leader's bundle are located within the preset neighborhood radius of the subsequent tasks, it indicates that the chain leader's path naturally passes through this area, and the marginal cost of accepting the subsequent tasks is low, so the threshold is relaxed accordingly; otherwise, a strict threshold is maintained to avoid unnecessary path detours. In a high-chain-density special experiment with 80 tasks / 12 ships and a chain-task ratio of 1.0, the number of cross-ship handovers decreased from 41.94 to 40.28 after enabling the spatial cooperative reward, effectively reducing the physical waiting delay and information synchronization delay introduced by cross-ship cooperation.
[0029] (6) Enhancing Failure Recovery Capabilities: Existing systemic recovery capabilities under extreme disturbances are insufficient. Orphan tasks are difficult to relocate quickly and reasonably after USV failure, especially in cascading failure scenarios where task completion rates plummet. This invention constructs the SHARP strategic hierarchical adaptive recovery protocol, integrating multi-criteria intelligent scoring, Tier 3-level hierarchical screening, chain-aware batch reallocation, predictive capacity relocation, and cascading response grading based on failure counts to achieve local repair after failure. Multi-criteria scoring comprehensively considers six components: travel time, load factor, capability specificity, speed matching, task priority, and time margin; Tier 3-level screening ensures that tasks are still salvaged even when resources are scarce; chain-aware batch reallocation maintains locality within the chain; predictive relocation frees up execution slots for high-priority orphan tasks; and cascading response grading ensures that the algorithm remains stable during single, occasional failures and gradually increases intervention intensity during cascading failures. In the cascading failure scenarios of 20%, 40%, and 4 rounds multiplied by 12%, the present invention achieved a failure recovery rate significantly higher than the baseline. In particular, in the extreme scenario of 40% fleet failure, the mission completion rate reached about 2.5 times that of the second-best baseline, while the comparison algorithm had a failure recovery rate of zero because it did not set up a dedicated failure recovery process.
[0030] (7) Providing a safety net in extreme resource saturation scenarios: Existing methods often fail to allocate high-priority MAYDAY tasks in extreme resource saturation scenarios due to full capacity of all USVs. This invention establishes a MAYDAY preemption mechanism based on a preemption benefit function. By evaluating the net benefit of releasing low-priority tasks, it preempts slots and places them at the top of the path. When all candidate USVs are close to their capacity limits, it creates execution slots for unallocated MAYDAY tasks by releasing low-priority tasks. This mechanism is positioned as a safety net strategy in extreme resource saturation scenarios and is used in conjunction with the SHARP recovery protocol: the preemption mechanism releases MAYDAY execution slots, and the evicted tasks are reassigned by SHARP or a replanning process.
[0031] (8) Ensuring Dependency Order Feasibility and Supporting Asynchronous Incremental Replanning: Existing methods are prone to dependency order violations after multiple rounds of redistribution, where subsequent tasks in the path are ranked before preceding tasks. Furthermore, global recalculation is required when the task graph changes, resulting in poor real-time performance. This invention performs order-preserving Kahn topology sorting on the path of each USV before outputting the solution, ensuring that the position index of any preceding task in the path is strictly less than the position index of its successor task. After sorting, the feasibility of the time window and cumulative voyage is re-verified. For newly generated infeasible tasks, the lowest priority task is released before entering the next round of local repair, thereby eliminating allocation order violations at the mechanism level. At the same time, the SHARP protocol and incremental replanning are triggered asynchronously when USV failure events and task graph change events occur, respectively. Local consensus loops are only performed on the affected subgraphs, preserving the stable allocation of the unaffected parts, thus avoiding the computational overhead of triggering global recalculation for each disturbance. Furthermore, this embodiment establishes a multi-dimensional evaluation paradigm consisting of core indicators such as Strict Task Completion Rate (SCR), Chain Completion Rate (CCR), Mayday Response Time (MRT), and Failure Recovery Rate (FRR), as well as auxiliary indicators such as the load Gini coefficient, so that each mechanism can be independently verified on the proprietary indicators corresponding to its design objectives. Attached Figure Description
[0032] Figure 1 This is a flowchart of a multi-stage chained task allocation method for maritime search and rescue based on priority-aware two-layer collaborative decision-making in the embodiment. Figure 2 This is a diagram of the overall architecture of the PRISM-CBBA algorithm in the embodiment; Figure 3 This is a schematic diagram of the cost matrix construction including AUD and WEB and the Hungarian warm start process in the embodiment; Figure 4 This is a schematic diagram of the chain co-location and spatial collaborative reward redistribution mechanism in the embodiment; Figure 5 This is a schematic diagram of the MAYDAY emergency task preemption mechanism in the embodiment; Figure 6 This is a flowchart of the SHARP strategic hierarchical adaptive recovery protocol in the embodiment. Detailed Implementation
[0033] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0034] To address the shortcomings of existing technologies, such as insufficient chain dependency modeling, inadequate consideration of cross-vessel handover costs, lack of urgency awareness, inadequate load balancing, and insufficient systemic recovery capability under extreme disturbances, this embodiment provides a multi-stage chained task allocation method for maritime search and rescue based on priority-aware two-layer collaborative decision-making. (See [link to relevant documentation]). Figure 1 and Figure 2 This includes the following steps: S1. Data Initialization and Task Graph Construction: Obtain the status data, task environment data, and chained dependencies of each USV in the heterogeneous unmanned surface vessel (USV) cluster.
[0035] This embodiment considers that... m Search and rescue cluster composed of heterogeneous unmanned surface vessels V ={ v 1, v 2,…, v m The USV will coordinate and conduct multi-phase rescue operations within an 8 km x 8 km nearshore area. Each USV is characterized by a five-element designation: v i = ( x i , sp i , R i , C i , B max ) in, x i ∈ℝ 2 These are the initial position coordinates; sp i Maximum speed, expressed in kilometers per hour; R i Maximum cruising radius, in kilometers; C i ⊆ C ={camera, radar, liferaft, tow_cable} is USV i A set of capabilities; B max =12 is the maximum bundled capacity. The cluster contains at least 6 types of heterogeneous configurations, with each type of USV carrying different combinations of capabilities, reflecting the functional division of the actual search and rescue fleet.
[0036] Task Collection T ={ t 1, t 2,…, tn Each task in} t j ∈ T Defined by the following nine-tuple: t j = ( id j , type j , cap j , location j , hard j , [ tws j , two j ], pri j , cid j , sid j ) in, id j As a task identifier; type j For mission types, the value range includes at least eight categories: DET detection, POS positioning, RES rescue, TOW towing, PAT patrol, SCN scanning, and independent rescue S-RES and independent towing S-TOW. cap j ∈ C The capabilities required to perform this task; location j ∈ℝ 2 The location where the mission took place; hard j The estimated execution time is in minutes; tws j and two j The start and end times of the time window, in minutes; pri j ∈{2,3,4,5} represents the priority level, where 5 corresponds to MAYDAY emergency rescue, 4 corresponds to PAN-PAN detection / location, 3 corresponds to towing / patrol, and 2 corresponds to general patrol / scanning; cid j ∈{-1,0,1,…} represents the chain number to which the task belongs. cid j =-1 indicates an independent task; sid j ∈ℕ represents the stage position index of the task in the chain. In this embodiment, tThis indicates the current time of the simulation, in minutes, with a simulation time step of 1 minute.
[0037] To facilitate the differentiation of the impact magnitude of different priorities in the cost function, a priority weight mapping is defined. w :{2,3,4,5}→{5,20,50,100}, record the task. t j The priority weight is p j = w ( pri j ),Right now p j ∈{5,20,50,100}, where the MAYDAY emergency rescue mission corresponds to p j =100, corresponding to PAN-PAN detection / localization mission p j =50, corresponding to towing / patrol tasks p j =20, corresponding to typical patrol / scanning missions. p j =5. This mapping uses bounded hierarchical values instead of exponentially amplified values, keeping the weight ratio between the highest and lowest priorities at around 20 times. This ensures that high-priority tasks are given significant priority while avoiding exponential amplification that would completely overwhelm low-priority tasks in the cost function and cause numerical distortion.
[0038] Chain dependencies are modeled using directed graphs. Maritime search and rescue missions typically exhibit a multi-stage chain structure of detection-location-rescue-towing. This embodiment constructs at least six typical mission chain templates based on the typical search and rescue operation flow of the IAMSASAR Manual Vol.III, as shown in Table 1. Approximately 80% of the missions in the mission set are generated from these templates through weighted random sampling, while the remaining approximately 20% are independent missions.
[0039] Table 1. Task chain templates and their capability dependency sequences For those belonging to the same chain and cid Two identical tasks t i and t j ,like sid i < sid j Then there exists a directed edge ( t i , t j )∈ Edep ,express t j Must t i It can only be executed after completion. E dep For multi-stage chained task dependency graph G =( T , E dep The set of directed dependent edges of (). Let be the set of edges of (). t j The precursor set is before ( t j )={ t i :( t i , t j )∈ E dep The successor set is} succ ( t j )={ t k :( t j , t k )∈ E dep}
[0040] A feasible task allocation scheme must strictly meet the following six constraints: Constraint C1, namely capability matching: USV can only perform tasks covered by its set of capabilities.
[0041] Constraint C2, namely the cruising radius: the cumulative distance traveled by the USV along the execution path must not exceed its maximum cruising radius; Constraint C3, namely dependency order and handover waiting: Subsequent tasks can be pre-assigned during the planning phase, but their actual start time and successful completion determination must be constrained by the preceding task; subsequent tasks can only begin after the preceding task has been completed and the necessary handover has been completed; if the two phases are performed by the same USV, the handover waiting time mainly consists of path transfer time; if the two phases are performed by different USVs, the handover waiting time further includes additional delays such as information synchronization, location proximity, and load handover.
[0042] Constraint C4, i.e., time window: The execution of the task must be initiated and completed within the specified time window; considering factors such as sea state disturbances, communication delays and platform maneuvering errors in maritime operations, this embodiment sets a time window flexibility tolerance coefficient γ=1.5 to describe the limited tolerance range outside the publicly announced time window of the task. Constraint C5, namely the path execution order: USV must strictly execute tasks according to the path order assigned by the algorithm and must not skip tasks that are blocked due to dependencies; Constraint C6, namely, bundled capacity: the total number of missions for each USV must not exceed the capacity limit.
[0043] Based on the above six constraints, a strict task completion rate (SCR) is defined as the core optimization objective. A task is considered successfully completed if and only if constraints C1 to C6 are satisfied simultaneously. SCR is defined as follows: in y j ∈{0,1} is a binary perfect variable. Let |T| be the set of tasks defined above, and |T| be the total number of tasks; y j The rule for determining the value is: when the task t j The value is 1 if the simulation is successfully completed within the simulation period and all 6 constraints are satisfied simultaneously; otherwise, it is 0.
[0044] For chained tasks, this embodiment employs a cascading validity determination: if any stage in a task chain fails to strictly meet C1 to C6, then all tasks within that chain are not counted in the valid completion count of SCR; independent tasks are judged separately as individual tasks. This setting reflects the engineering characteristic that failure of any critical stage in a maritime search and rescue chained task can lead to the failure of the entire chain.
[0045] S2, GNN Keyness Score: Taking the multidimensional attributes of task nodes as input, the pre-trained lightweight graph attention network GAT performs forward reasoning on the task dependency graph and outputs the global keyness score for each task node.
[0046] The GNN module aims to identify bottleneck tasks with high marginal contributions to global SCR from the topology of the task dependency graph. Input node feature matrix. X ∈ℝ n×12 It includes 12 attributes such as task type, priority, time window, and spatial coordinates. The 12 node features are as follows: 2-dimensional normalized coordinates of the task location, normalized to an 8km edge length of the operational sea area; one-hot encoding of the required capabilities for the task, 4-dimensional, corresponding to camera, radar, liferaft, etc. t cable Normalized priority weights p j / 100, 1 dimension in total; normalized estimated execution time hard j / 30, 1 dimension in total; normalized time window width D j / 200, 1 dimension; normalized number of predecessor tasks | before ( t j )| / 4, total 1 dimension; normalized successor task number| succ ( t j )| / 4, 1 dimension in total; chain attribution indicator, 1 dimension in total. c i d j A value of 1 is assigned when the value is ≥0, and 0 is assigned for independent tasks. The task type is represented by a one-hot encoding of its required capabilities, and the task's stage attributes are jointly represented by the number of predecessor tasks, the number of successor tasks, and the chain affiliation indicator. The forward propagation of the two-layer GAT is as follows: in K To determine the number of attention heads, in this embodiment, the first layer is... K =4, the second layer takes K =1 indicates single-head mean aggregation. For the first l The hidden feature matrix of the layer's nodes and , For task nodes t j In the hidden feature vectors of layer l, ReLU is the modified linear unit activation function. For multi-head splicing, α ij (k,l) For the first l In the GAT standard layer, the first k Attention weights for each attention head For task nodes t i The set of neighbor nodes in the dependency graph, i.e. N ( i )={ t j :( t i , t j )∈ E dep or( t i , t j )∈ E dep}, For standard GAT network parameters, specifically the first... l Layerk Learnable linear transformation matrix for each attention head W (k,l) Determined during the offline training phase. For the Sigmoid function, , These represent the weights and biases of the output layer, respectively. In this embodiment, the output layer is specifically implemented as a two-layer perceptron containing a 16-dimensional ReLU hidden layer, and outputs via a Sigmoid function. The above equation represents its simplified form. c j ∈[0,1] will be used as a discount term in the subsequent cost matrix to guide the matcher to prioritize bottleneck tasks.
[0047] The random seed used in the task graph during the GNN training phase is completely separated from the subsequent algorithm evaluation phase. Training labels are obtained by estimating the marginal contribution of task nodes to the global SCR through offline Monte Carlo simulation. During the online assignment phase, only the trained lightweight model is used for key score inference, without accessing the global completion results of the test scenario. The GNN model has only about 6657 parameters, and the correlation coefficient between the validation set prediction and the ground truth is around 0.97. It can perform real-time inference in a CPU-only manner on edge embedded platforms such as ARM Cortex-A72, without GPU dependency.
[0048] S3. Construction of the cost matrix including AUD and WEB: For any USV and task pair, construct a cost matrix including base travel time, GNN criticality discount, AUD urgency discount, and WEB load penalty, such as... Figure 3 As shown.
[0049] In the r In round matching, for each pair ( v i , t j The construction cost is: in, r The matching round sequence number for warm-up startup. For the first USV in round matching Execute the task The cost; , , The three correction terms are: a GNN criticality discount term (centering the score to [-0.5, 0.5]), an AUD urgency discount term, and a WEB payload penalty term; weights α g =12, α u =8, αw =4, all three are dimensionless weighting coefficients. If v i Not available t j If the required capability exceeds the cruise radius, then . For travel time cost operator, its value is equal to v i Current location and t j Location of occurrence location j Divide the Euclidean distance between them by v i Maximum speed sp i The travel time is given in minutes, and the speed was converted from kilometers per hour to kilometers per minute during the calculation. For urgency, For the load ratio, the definitions of AUD and WEB are explained later.
[0050] The input for AUD is urgency. u j First, define the task. t j Time window width With remaining time for: in, , Tasks The end and start times of the time window, in minutes, are the values within the nine-tuple in S1. and ; t The current time is in minutes; For the task t j Time window width, For the current moment t Assign task t j The remaining time of the time window.
[0051] The urgency of a task is defined as follows: At the current moment hour, This indicates that the task is not yet urgent; when t Approaching hour, Gradually increase; when hour, The AUD cost discount items are: In the formula, This represents the discount applied to the cost matrix elements by AUD; the negative sign indicates a reduction in cost. In this embodiment, we take... α u =8. AUD uses a quadratic curve. The non-linear form aims to dynamically adjust the matching cost based on the remaining proportion of the task's time window. When a task is far from its deadline, the AUD discount is small to avoid prematurely disturbing the overall matching; when a task is close to its deadline, the discount increases rapidly, allowing the algorithm to prioritize high-priority tasks that are about to time out.
[0052] The input to the web is the load ratio. , For USV v i The number of currently assigned tasks, i.e., their task bundles. The number of tasks in the process; The maximum bundled capacity is defined in S1, therefore For values ∈[0,1], the penalty term uses a 1.5-power superlinear form. Compared to linear penalties, 1.5-power penalties change more slowly during low-load phases and do not affect normal task reception; when the load exceeds medium to high levels, the penalty increases rapidly, thereby guiding the Hungarian matching phase to allocate new tasks to relatively lightly loaded USVs.
[0053] S4, Hungarian centralized warm start: Perform minimum-cost bipartite graph matching on the cost matrix, such as... Figure 3 As shown.
[0054] Perform standard minimum cost matching on the cost matrix: In the formula, For the first r The optimal matching set obtained by round matching. for V × Any feasible match is a set of tuples that satisfy the condition that each USV matches at most one mission and each mission matches at most one USV. The USV set defined in S1 For the first r The set of candidate tasks participating in the matching round, i.e., the task set. T Middle in front r -1 is a subset of tasks yet to be assigned in round 1. Warm start executes at most... B maxIn each round, matched tasks are removed from the candidate set and the cost matrix is reconstructed until no new assignments are generated. After each round of assignments, newly assigned tasks are appended to the corresponding USV bundle. B i and path P i End.
[0055] S5, Refinement of Distributed CBBA Consensus: Based on the initial allocation scheme, each USV acts as a distributed intelligent agent, eliminating inter-vessel allocation conflicts through a bounded round-by-round win-bid consensus transmission process.
[0056] Hungarian warm-start provides high-quality initial allocation from a global perspective, but its centralized nature lacks robustness in the event of communication link interruption or leader node failure. Therefore, this embodiment introduces a bounded distributed consensus phase after warm-start, borrowing the win-pass mechanism from classic CBBA to eliminate inter-vessel allocation conflicts and restore the algorithm's distributed fault-tolerant properties.
[0057] In each round of the consensus phase, each pair of USVs ( v i , v j Exchange the winning dictionary. v i Tasks in a Bundle t ,like v j If the bidder with the highest bid believes themselves to be the winner, a task transfer is triggered. Formally, the transfer rules are as follows: like , but Right now v i release t And update the local winning bid information to v j The bid. In the formula, For USV v j Locally recorded tasks t The winner's number, express v j He considers himself the winner of the task; , They are respectively v i , v j The winning bid value of task τ recorded locally, which is the highest bid it currently holds for that task; for v iThe task binding converges when there is no task transfer for a consecutive round. Maximum number of rounds. K max =5. If a small number of conflicts still exist after reaching the maximum number of rounds, a deterministic decision will be made based on a triple rule of task priority, bid value, and vessel number to ensure that each task is assigned to at most one USV in the final output.
[0058] Winning the bid here W i ( t This is derived from the CBBA bidding scoring function. Unlike the standard CBBA, which only uses driving cost as the bidding basis, this embodiment incorporates GNN criticality, priority weights, and chain dependency information into the bidding scoring: in, For USV v i For the task t j Bid scoring, winning bid value That is, take this score. First item Basic returns weighted by priority. c j The critical score for GNNs allows bottleneck tasks to receive higher bids. ϕ stage As a stage factor, when t j For the first phase of the chain, i.e. s i d j When the value is 0, the value is 2; otherwise, the value is 1, to expedite the initial response. For the task t j The base reward value is determined by the task type. In this embodiment, Rescue (RES) is 200, Towing (TOW) is 180, Positioning (POS) is 110, Detection (DET) is 100, Independent Rescue (S-RES) is 150, Independent Towing (S-TOW) is 130, Scanning (SCN) is 80, and Patrol (PAT) is 70; the second item For chain affinity rewards, parameters =0.5, This indicates the event to which the task belongs, i.e., the task chain number. For chained tasks, its value is the chain number cid. This item is used for statistics. v i Currently bundled with t j The number of tasks belonging to the same event is limited; tasks within the same chain are encouraged to be grouped onto the same vessel to reduce handover. (Third item) middle, ψ Here is a priority-adaptive driving cost penalty function, where for v i to t j Travel time, In other words, the higher the priority weight, the smaller the penalty coefficient per unit of travel time and the higher the tolerance for distance, thus preventing the MAYDAY task from being rejected due to slightly longer distance.
[0059] S6. Chain Co-location and Spatial Cooperative Reward Redistribution: For subsequent tasks in the same task chain, migration is made towards the chain leader boat under the premise of satisfying hard feasibility constraints; the migration threshold is adaptively adjusted through the spatial cooperative reward function. The process is as follows: Figure 4 As shown.
[0060] Adjacent stages of a maritime search and rescue mission chain typically require different capabilities, resulting in stages of the same chain often being distributed across different vessels after initial matching. Cross-vessel handovers introduce waiting delays, which may trigger timeouts in downstream stages under strict path execution order constraints. The goal of a chain co-location mechanism is to reduce the dispersion of the same mission chain across different USVs under feasibility constraints, thereby mitigating the waiting risks associated with cross-vessel handovers.
[0061] Chain c The subsequent tasks (Currently allocated to) ), towards the chain-headed boat The conditions for migration are: The first three items , , As a hard feasibility constraint, the fourth item The criterion for determining the cost. Among them, For the chain c The USV currently assigned to the first phase of the mission is referred to as the Chainhead Vessel. For the chain c The subsequent phase tasks in for The current USV; For chain-headed boats The number of tasks currently assigned. Operators that represent the capabilities required to perform a task. For chain-headed boats A collection of capabilities This represents the same travel time operator as in S3, which is the Euclidean distance between two points divided by the maximum speed of the corresponding USV, in minutes. d colo =1 minute as the basic common-domain tolerance threshold β ( ) represents the spatial collaborative reward function: in, , They represent subsequent tasks respectively. Position coordinates and chain-head boat binding The position coordinates of task τ′ in the middle; For chain-headed boats Task bundling, The formula is the Euclidean norm, with a coefficient of 4.0 representing the unit collaborative reward value, and 0.5km representing the preset neighborhood radius. This formula counts the number of tasks located within a 0.5km neighborhood of the subsequent task in the chain-head boat bundle and multiplies it by the unit collaborative reward value.
[0062] The design concept is to transform the spatial clustering of tasks into an adaptive adjustment factor for the migration threshold. When multiple tasks in the lead boat cluster are already within a 500-meter neighborhood of the subsequent tasks, it indicates that the lead boat's path naturally passes through this area, and the marginal cost of accepting the subsequent tasks is low, so the threshold is relaxed accordingly; otherwise, a strict threshold is maintained to avoid unnecessary path detours.
[0063] S7, MAYDAY Preemption Mechanism: When the fleet is nearing resource saturation and there are unassigned MAYDAY emergency missions, the optimal preemption plan is selected according to the preemption benefit function, such as... Figure 5 As shown.
[0064] The MAYDAY preemption mechanism is positioned as a fallback strategy for extreme resource saturation scenarios, rather than a regular allocation mechanism that independently improves the overall SCR. When all candidate USVs are close to their capacity limit, the preemption mechanism creates execution slots for unallocated MAYDAY tasks by releasing low-priority tasks.
[0065] For unassigned MAYDAY tasks t In each ship with matching capabilities v i upper traversal binding B i Find the optimal preempted task The profits to be seized are: in, This indicates the rewards for the MAYDAY task. , The MAYDAY task τ Priority weights and base reward values, This indicates the cost of being expelled from the mission. , respectively the expulsion mission The priority weight and base reward value are determined, with a discount factor of 0.5. This indicates the cost of the candidate vessel traveling to the MAYDAY mission location. This represents the response time savings resulting from preemption, and its value is equal to the earliest arrival time that the MAYDAY mission could have achieved before preemption and the time saved by the candidate vessel. The difference in travel time to the mission is taken as 0 if the MAYDAY mission has not been accepted by any USV before preemption. The trigger condition is... Benefits and and Not currently in execution. After execution, the evicted task is released to the READY state, awaiting the next round of reassignment. Insert to preempt the first position on the boat's path to ensure the fastest response.
[0066] S8, SHARP Strategic Hierarchical Adaptive Recovery Protocol: When USV failure generates orphan tasks, it achieves rapid takeover of orphan tasks through multi-criteria intelligent scoring and Tier 3 hierarchical screening, such as... Figure 6 As shown.
[0067] When a USV fails, the tasks bundled with it become isolated tasks and need to be quickly reallocated. This embodiment designs a hierarchical adaptive recovery protocol, SHARP, which achieves local repair after failure through four stages: candidate vessel scoring, hierarchical screening, chain-aware batch reallocation, and capacity relocation.
[0068] First, there's multi-criteria intelligent scoring. For each isolated task... t With candidate submarine v Calculate the overall score: in, s ( v , t ) as candidate submarine v Take over isolated mission t Overall score; For candidate submarines To the mission The estimated travel time; This represents the number of missions currently bundled with the candidate vessel. The size of the candidate submarine capability set, i.e., the number of capabilities contained in the set; Speed of the candidate vessel; This is the maximum speed of the currently available fleet; This is an indicator function; it takes the value 1 when the task priority weight is not less than 50, and 0 otherwise. This provides a time margin between the recovery deadline and the expected completion time. The six components correspond to the following in order: travel time (negative), load factor (negative), capability specificity (positive), high-priority task speed matching reward (positive), task priority (positive), and time margin reward (positive).
[0069] Secondly, there is a three-tiered tiered screening process. To improve the takeover probability of orphan tasks in extreme failure scenarios, SHARP employs a relaxed recovery threshold during the candidate screening phase. c R =1.8. It should be noted that this is a relaxed recovery threshold. c R =1.8 is greater than the time window flexibility tolerance coefficient γ=1.5 used for SCR determination in constraint C4 of S1. Its purpose is to expand the candidate vessel range as much as possible during the failure recovery phase to salvage orphan missions. Correspondingly, any mission whose final execution time exceeds the determination upper limit of γ=1.5, even if successfully taken over, will not be counted in SCR, but only recorded as a diagnostic recovery attempt. Candidate vessels are divided into three levels according to recovery feasibility and cruise safety: Tier 1: Simultaneously meets time feasibility requirements ( t + T v,τ + hard ≤1.8 t we ) and cruise safety (cumulative range ≤ 0.85 R v ),in, T v,τ For candidate submarines v To the mission t The estimated travel time, hard For the task t The expected execution time, i.e., the S1 nine-tuple hard , t we For the task t The time window deadline, i.e., the twe in the S1 nine-piece set. R v For candidate submarines v The maximum cruising radius, i.e., the S1 quintuple R ; t The current time; 1.8 in the formula is the aforementioned relaxed recovery threshold. c R0.85 is the cruise safety margin coefficient, meaning that cruise safety is only determined when the expected cumulative range does not exceed 85% of the maximum cruise radius; Tier 2: Only meets the time feasibility requirement but not the cruise safety requirement; Tier 3: A backup plan that does not meet either requirement, accepting backup candidates that exceed the SCR (Scheduled Recovery Response) threshold, but if such tasks ultimately violate the time window, they will not be included in the SCR and will only be recorded as diagnostic recovery attempts. In addition, the comprehensive score of Tier 3 backup candidates is adjusted based on the aforementioned comprehensive score as follows: the penalty coefficient for the travel time component is increased from 3 to 4, the capability specificity component is reduced by 0.8, and the speed matching bonus and time margin bonus components are removed.
[0070] The algorithm selects the highest-scoring algorithm from the higher levels to ensure that it can still do its best to save the task when resources are scarce.
[0071] Third is chain-aware batch redistribution, which arranges isolated tasks on the same chain in phase order and processes them in batches, prioritizing the maintenance of locality within the chain.
[0072] Fourth is predictive capacity relocation. When the number of assigned tasks of a target vessel reaches the upper limit of its bundled capacity, it proactively removes the lowest priority non-executing task in its bundle to a neighboring vessel, freeing up slots for high-priority isolated tasks.
[0073] In addition, SHARP maintains a failure counter within a 120-minute sliding window. nf Based on this, the aggressiveness of the recovery strategy is adaptively adjusted. Let ℱ 120 ={ t '∈ T fail : t - t '≤120' represents the set of recent failure times, where Tfail is the set of all USV failure events since the start of the simulation. t The current time is represented by 120, which is the duration of the sliding window in minutes. n f =|ℱ 120 The cascade response is then classified as follows: Tier = 1 (Regular Recovery) n f < 2; Tier = 2 (medium rebalancing), 2 ≤ n f < 4; Tier = 3 (aggressive release), n f ≥ 4.
[0074] Tier 2 migrates the lowest priority tasks from overloaded USVs (load > L+1) to lightly loaded USVs, where L is the arithmetic mean of the number of tasks assigned to all active USVs, i.e., the fleet's average load. USVs with loads greater than L+1 are classified as overloaded, while those with loads less than L are classified as lightly loaded. Tier 3 directly releases low-priority assigned tasks from overloaded USVs into a pending state, freeing up global optimization space for the next warm-up. This hierarchical mechanism ensures the algorithm remains stable during single, isolated failures and gradually increases intervention intensity during cascading failures, avoiding mismatch of fixed strategies under these two extreme conditions.
[0075] S9. Topology Normalization and Scheme Output: Perform order-preserving Kahn sorting on the mission path of each USV, ensuring that the position index of any preceding mission in the path is strictly less than the position index of its successor mission.
[0076] For each ship v i path P i Perform Kahn sorting to preserve order, so that it satisfies ,in, This is the set of predecessor tasks for the task, defined in S1. The position index of the task in the path Pi of the ship, with a value of 1 to |Pi|, is a positive integer. t a , t b For any two tasks in path Pi, after topology normalization, the arrival time, cumulative distance, and time window feasibility of each boat's path are recalculated; if a new time window or distance violation occurs, the lowest priority infeasible task is released and the next round of local repair is initiated.
[0077] The final output includes the optimal allocation scheme for each USV, including the final mission sequence, detailed execution timetable, strict mission completion rate (SCR), chain completion rate (CCR), Mayday response time (MRT), failure recovery rate (FRR), load Gini coefficient, coefficient of variation (CV), load imbalance ratio, and number of inter-vessel chain handoffs (Handoffs). The SHARP protocol and incremental replanning are triggered asynchronously when USV failure events and mission map change events occur, respectively. Only the affected subgraphs are processed using a local mini-CBBA loop, preserving the stable allocation of unaffected parts.
[0078] In summary, the PRISM-CBBA algorithm in this embodiment performs well in each allocation cycle Δ t Perform the following steps within 3 minutes: Step 1, Cache determination: Calculate the state key key k =( frozenset (T ), frozenset (unassignedREADY / PENDING)), where, k To assign periodic numbers, frozenset ( ) represents an operator that constructs an immutable set from an input set, used to generate hashable state keys that are independent of element order. T This represents the set of all tasks within the current allocation period. Unassigned READY / PENDING represents the set of tasks that have not yet been allocated and are in the READY or PENDING state. If it matches the cache, the allocation result of the previous round will be returned directly to avoid duplicate calculations.
[0079] Step 2, GNN inference. Call... Obtain the key vector ,in This represents the forward inference function of the two-layer graph attention network obtained through offline pre-training in S2; , These represent the current multi-stage chained task dependency graph. G =( T , E dep ) and defined in S2 n A ×12-dimensional node feature matrix Indicates the number of task nodes, i.e. n =| T |;vector c The j Each component is a task. t j Key scores c j .
[0080] Step 3, Auction Initialization. Initialize the winning bid dictionary for each vessel, retain the IN_PROGRESS task, and selectively release low-priority ASSIGNED tasks based on the cascading failure count.
[0081] Step 4, Hungarian warm start. Repeat up to [number] cycles. B max In each round, the cost matrix is constructed and the matching is solved according to section S3. If there is no new allocation, the process exits.
[0082] Step 5, Distributed Consensus Refinement. Execute up to [number] steps as per section S5. K max =5 rounds of winning bids were passed on to eliminate conflicts between boats.
[0083] Step 6, Chain Co-location. Migrate the chain's successor tasks to the lead ship according to section S6.
[0084] Step 7, MAYDAY preemption. Perform preemption on the unassigned MAYDAY task as per section S7.
[0085] Step 8, Topology Normalization and Output. Perform Kahn sorting with order preservation on the path of each USV, update the cache, and output the allocation results.
[0086] The SHARP protocol and incremental replanning in Section S8 are triggered asynchronously when USV failure events and task graph change events occur, respectively. They only execute local mini-CBBA loops on the affected subgraphs, preserving the stable allocation of the unaffected parts.
[0087] The overall computational complexity analysis of the algorithm in this embodiment is as follows: Warm start execution up to B max Round matching, O(( m + n ) 3 Therefore, the overall complexity is O( ). B max ·( m + n ) 3 GNN inference is O( n ), chain co-location and preemption both do not exceed O( n · B max The consensus phase is O( R · m 2 · B max )( R ≤5), which can be ignored. The space complexity is O( m · n +| E dep |). In the formula, m For the number of USVs, n For the number of tasks, R The number of rounds actually executed in the consensus phase and R ≤ K max =5,| E dep | represents the number of dependent edges.
[0088] The analytical properties of AUD include three properties: Property 1: Monotonicity, the partial derivative of the AUD discount with respect to urgency is greater than or equal to zero, therefore the closer the task is to the deadline, the larger the AUD discount; Property 2: Marginal acceleration, the second partial derivative of the AUD discount with respect to urgency is greater than zero, therefore the AUD discount increases with increasing urgency, achieving a response effect of early gentleness and rapid enhancement near the deadline; Property 3: Boundedness, the upper bound of the maximum AUD discount is the coefficient. α u This can prevent urgency from completely outweighing travel time costs.
[0089] The balance guarantee of the web, function f ( x )= x 1.5 The load distribution is a convex growth function on [0,1]. According to Jensen's inequality, when the total workload is constant, a balanced load distribution corresponds to a lower average penalty value. Therefore, the web will guide tasks towards lighter USVs during the matching phase, but this property does not constitute a strict guarantee of perfect load balancing in the final result.
[0090] To fully verify the effectiveness, robustness, and scalability of the method of this invention, this embodiment designed a comparative experiment in nine stages: simulation environment configuration, GNN pre-training effect verification, static scale scalability test, computation time benchmark test, dynamic perturbation scenario test, mechanism ablation study, mechanism-specific experiment, parameter sensitivity analysis, and statistical significance test. Six representative baseline algorithms were selected as comparison objects: Hungarian algorithm, Standard-CBBA consensus binding algorithm, CCBBA with coupling constraints, CBBA+Priority with priority, SSI-Auction sequential single-item auction, and Greedy-Priority heuristic.
[0091] I. Simulation Environment and Parameter Configuration. The simulation adopts a discrete event-driven model with a time step of 1 minute. The maximum simulation duration varies between 480 and 900 minutes depending on the scale of the scenario. The operating sea area is a two-dimensional nearshore area of 8 km × 8 km, with a 5% safety margin at the edge. The USV cluster adopts six heterogeneous configurations, generated at a USV to task ratio of 0.25, and the minimum fleet size is set to 10 vessels. The task chain generation parameters are: 80% chained tasks plus 20% independent tasks; the chain space aggregation radius is approximately 50 meters, and the Gaussian noise σ = 30 to 40 meters; the time window starts randomly within [0, 30] minutes, with a width of 180 to 300 minutes, and each stage is delayed by 20 minutes to simulate physical handover delay. Each data point is run 100 times independently and repeatedly, and the mean and standard deviation are taken. All experiments are completed on a desktop workstation with Ubuntu 20.04 and Python 3.12 environment.
[0092] II. Verification of GNN Pre-training Performance. This embodiment first performs offline pre-training on the GNN key scoring module. During the pre-training phase, a total of 300 task graph samples are generated, each containing 8 to 25 events. The training set and validation set are divided in an 80 / 20 ratio. The upper limit of GAT model training is set to 200 epochs, and an early stopping strategy with a patience of 30 is adopted to save the optimal model for online inference.
[0093] During training, the training and validation losses of the GAT model decreased rapidly in the first few rounds before stabilizing. The Pearson correlation coefficient on the validation set remained stable around 0.97, indicating that the model can fit the task-critical structure reflected by the offline labels well. The GAT predicted point cloud closely followed the 45-degree diagonal relative to the ground truth, with a correlation coefficient r=0.967; while the heuristic baseline, i.e., the constant prediction of 0.5, collapsed entirely to the y=0.5 level, with zero variance and no correlation to the ground truth. This demonstrates that the GNN output can reflect the critical differences between different task nodes, providing a structured prior for subsequent cost matrix correction and bid scoring.
[0094] The GAT model has only 6657 parameters, and the entire training process is completed within 61 epochs. Its inference latency is less than 10 milliseconds, and it can perform real-time inference on edge embedded platforms such as ARM Cortex-A72 in a CPU-only manner, without requiring GPU support. These results demonstrate that the GNN key score module in this embodiment has engineering advantages such as high training efficiency, low inference overhead, and stable prediction accuracy, providing reliable feature input for key score enhancement bidding in the subsequent Hungarian matching stage.
[0095] III. Results of Static Scalability Experiments. Table 2 shows the comparison of key performance indicators of the seven algorithms under static scenarios P1 (30 tasks / 10 boats), P2 (60 tasks / 15 boats), P3 (120 tasks / 30 boats), and P4 (180 tasks / 45 boats).
[0096] Table 2 Key performance indicators of each algorithm in static scale scenarios P1-P4 As shown in Table 2, the average SCR of the method of this invention under the four static scales P1 to P4 are 0.970, 0.977, 0.982, and 0.982, respectively, remaining stable above 0.97 throughout, which is significantly better than the baseline algorithm. This result indicates that the method of this invention can maintain stable task completion quality as the task scale increases, without exhibiting the performance degradation phenomenon common in traditional CBBA-type algorithms with increasing scale. In terms of MAYDAY response, the MRT of the method of this invention remains within the range of 78 to 87 minutes, which highly matches the IAMSA golden window reference response range of 60 to 90 minutes. This indicates that the AUD urgency decay mechanism and priority weighting mechanism proposed in this invention can effectively guide high-priority tasks to be processed in advance.
[0097] IV. Comparison of Calculation Time. To verify the real-time deployability of the method of this invention after superimposing multiple mechanisms, pure allocation wall clock time measurements were performed on each algorithm. The results are summarized in Table 3, where the unit is milliseconds and the form is mean ± standard deviation, including 100 seeds.
[0098] Table 3 shows the single allocation of wall clock time for each algorithm at scales P1-P4. As shown in Table 3, the average time for a single allocation using the method of this invention is 146.6 milliseconds in the largest-scale P4 scenario, which is far lower than the real-time computing budget allowed by the 3-minute replanning cycle in this embodiment. Compared with pure distributed baselines such as Standard-CBBA and CCBBA, the computation time of the method of this invention is reduced by one to two orders of magnitude. This is because the Hungarian warm start provides a high-quality initial allocation, so that the subsequent CBBA consensus phase only requires a limited number of rounds of conflict resolution. At the same time, the time consumption of the method of this invention is on the same order of magnitude as that of pure Hungarian, indicating that mechanisms such as GNN inference, AUD / WEB cost correction, chain co-localization, and preemption detection do not introduce excessive additional computational burden. The above results show that the method of this invention has real-time running capability in engineering deployment.
[0099] V. Dynamic Disturbance Scenario Analysis. Under a baseline scale of 60 missions / 15 boats, seven types of dynamic events S1 to S7 were injected: S1 (20% boat fleet failure), S2 (40% boat fleet failure), S3 (injection of 18 independent missions), S4 (20% of missions canceled), S5 (35% of boat speed reduced to 45%), S6 (4 rounds multiplied by 12% cascading failure), and S7 (injection of 6 Mayday missions). The SCR comparison results of each algorithm are shown in Table 4.
[0100] Table 4 Comparison of SCR of each algorithm under dynamic disturbance scenarios S1-S7 As shown in Table 4, the method of this invention achieves the highest SCR in all seven dynamic disturbance scenarios from S1 to S7. Specifically, in the high failure rate scenario S2, the SCR of the method of this invention is 0.443, approximately 2.5 times that of the suboptimal baseline Hungarian. In scenarios such as new task injection, task cancellation, sea state deceleration, and Mayday surge, the method of this invention also maintains a high task completion rate. This indicates that the SHARP recovery protocol, incremental replanning mechanism, and priority-aware mechanism of this invention can effectively maintain the quality of task allocation under dynamic disturbance conditions.
[0101] Regarding failure recovery, the FRR of the method in this invention is 0.634, 0.796, and 0.343 in the three failure scenarios of S1, S2, and S6, respectively; while the comparative algorithms, due to the lack of a dedicated failure recovery process, all have an FRR of zero. This demonstrates that the orphan task takeover capability is the key source of the high SCR of the method in platform failure scenarios. Regarding Mayday response, the method in this invention maintains a low MRT in non-failure dynamic scenarios, with some scenarios approaching or falling within the 60-90 minute reference response range; in failure scenarios, due to the reduction in available fleet resources, the MRT increases but remains lower than most comparative algorithms.
[0102] VI. Mechanism Ablation Study. Under a mixed pressure scenario (80 missions / 20 submarines; 20% USV failure at t=60 minutes, 30% USV sea state deceleration at t=120 minutes, 12% USV cascade failure at t=150 minutes, and injection of 10 new missions at t=180 minutes), ten variants of the method of this invention were run, and the independent contributions of the nine core mechanisms were compared. The results are shown in Table 5, which includes a mixed pressure scenario of 80 missions and 100 seeds.
[0103] Table 5. Multidimensional ablation results of nine mechanisms in PRISM-CBBA As shown in Table 5, the mechanisms of this invention do not all aim solely to improve SCR, but rather serve different functional dimensions such as dynamic recovery, emergency response, load balancing, chained collaboration, and distributed conflict resolution. Specifically: SHARP and Replan are key modules for maintaining completion rate in dynamic disturbance scenarios; GNN has a significant impact on emergency task response efficiency, with MRT increasing significantly after removing GNN; WEB mainly improves fleet load fairness; Synergy mainly reduces cross-ship handover of chained tasks; Preempt is an emergency fallback mechanism in MAYDAY saturation scenarios; Consensus ensures distributed consistency of warm-start results. The above ablation results verify the multi-mechanism complementary design principle of this invention.
[0104] VII. Mechanism-Specific Experiments. Since the independent contributions of some mechanisms can be absorbed by stronger mechanisms in hybrid pressure ablation scenarios, this embodiment further isolates and tests the independent engineering value of preemption mechanisms, spatial collaborative rewards, and superlinear load penalties through three specific experiments.
[0105] (I) Special Experiment on Preemption Mechanism. To isolate and test the real contribution of the preemption mechanism, this embodiment constructs a resource saturation scenario, i.e., 80 tasks configured with only 6 USVs, making the average number of vessels bundled close to the bundled capacity limit of 12, and superimposing a MAYDAY surge event to inject 6 emergency MAYDAY tasks; in both variants, the failure recovery process is disabled, making the preemption mechanism the only mechanism that can create execution slots for unassigned MAYDAY tasks by evicting low-priority tasks. The experimental results under 100 independent seeds are shown in Table 6, which represents the resource saturation scenario and includes 100 seeds.
[0106] Table 6 Special Experiment on Preemption Mechanism Table 6 shows that the preemption mechanism is triggered an average of 6.5 times per simulation, verifying its high activity under resource saturation conditions. However, under isolation conditions with the SHARP recovery mechanism disabled, the SCR and MAYDAY completion rates of the Preempt Only variant are lower than those of the No Preempt variant, while the MRT also increases. This is because although the preemption operation can free up slots for MAYDAY tasks, the evicted low-priority tasks will remain unplaced for a long time if there is no subsequent recovery or reallocation mechanism, thus lowering the overall task completion rate.
[0107] Therefore, the preemption mechanism should be positioned as an extreme fallback component within the complete PRISM-CBBA system, rather than a conventional independent gain module. When SHARP and the replanning mechanism are activated simultaneously, the preemption mechanism can create temporary execution slots when the fleet is fully loaded and the MAYDAY mission is unassigned, allowing the rejected mission to be reassigned through the recovery or replanning process. This special experiment, from the opposite perspective, verifies the engineering constraint that the preemption mechanism must be used in conjunction with the recovery mechanism.
[0108] (II) Superlinear Load Penalty WEB Special Experiment. The design goal of the WEB mechanism is to ensure fleet load fairness, and its engineering value must be evaluated on the corresponding proprietary indicators (Gini coefficient, coefficient of variation CV, load imbalance ratio η). To isolate and test the independent contribution of the WEB, this embodiment compares two variants with and without the WEB in a P2 static scale (60 missions / 15 ships) with 100 seeds. The results are shown in Table 7, which represents the P2 static scenario and includes 100 seeds.
[0109] Table 7. Special Experiment on Superlinear Load Penalty (WEB) As shown in Table 7, after removing the web interface, the load Gini coefficient deteriorated from 0.254 to 0.322, and the load imbalance ratio increased. or The performance index (SCR) deteriorated from 6.12 to 7.76, representing a relative improvement of over 21%; simultaneously, the SCR also decreased by 2.4%. This indicates that the Web (WEB) can not only improve the load distribution of the fleet but also reduce the risk of subsequent mission failures caused by overloaded boats to a certain extent. Compared with the mixed pressure ablation scenario, the WEB effect is more pronounced in the static scenario because there is no interference from frequent failures and replanning, allowing the load penalty to more stably affect the Hungarian matching results. These results validate the independent engineering value of the WEB as a load fairness adjustment mechanism.
[0110] (III) Space Collaborative Reward Experiment. To isolate and test the independent contribution of the space collaborative reward mechanism, this embodiment constructs a high-chain-density scenario, i.e., 80 tasks / 12 submarines, with a chain-to-task ratio of 1.0. All tasks are distributed across 6 types of task chain templates. Under this configuration, each submarine is bundled with approximately 7 tasks, and the number of cross-submarine chain handover candidates is significant, providing a sufficient observation window for the adaptive threshold adjustment of the space collaborative reward mechanism. The experimental results with 100 seeds are shown in Table 8, which represents the high-density chain scenario with 100 seeds.
[0111] Table 8. Special Experiment on Spatial Collaborative Reward Mechanism As shown in Table 8, enabling the space collaboration reward reduced the number of inter-ship chain handovers from 41.94 to 40.28, indicating that the space collaboration reward can promote the co-location of tasks within the same chain and reduce task handovers between different USVs. However, enabling Synergy reduced the SCR and CCR by 0.016 and 0.015 respectively compared to disabling Synergy, suggesting that while reducing inter-ship handovers, this mechanism may introduce slight path detours or local resource occupation. Therefore, the engineering positioning of Synergy should be as a regulatory mechanism to improve the local continuity of chained tasks, rather than an independent allocation mechanism aimed at improving the overall completion rate; its advantage lies in reducing the physical waiting delay and information synchronization delay introduced by inter-ship collaboration.
[0112] 8. Parameter Sensitivity Analysis. To evaluate the robustness of the method of this invention to key engineering parameters, this embodiment performs a four-dimensional parameter scan in a P2 baseline scenario, with five sampling points in each dimension. The scan parameters include the USV to task ratio, time window width, USV failure rate to chain task ratio, and the scan range is shown in Table 9.
[0113] Table 9 Sampling range for parameter sensitivity analysis The parameter sensitivity analysis results can be summarized as follows: as the USV to task ratio increases from 0.15 to 0.35, the SCR of the proposed method remains stable between 0.927 and 0.981 across the entire range, reaching a peak of 0.981 at a ratio of 0.30; under the same conditions, the SCR of the Hungarian method increases from 0.511 to 0.869. Even under the most resource-constrained extreme condition of 0.15, the proposed method still outperforms the suboptimal baseline by 81.5%, indicating that the proposed method's advantage is most prominent under tight resource constraints. As the time window widens, the SCR of the proposed method monotonically converges from 0.883 to 0.981; even under the tightest (90,150) minute window, the proposed method maintains a high level of 0.883, far exceeding the 0.664 of the Hungarian method under the same conditions. These results demonstrate that the AUD urgency decay, priority weighting, and replanning mechanism can effectively maintain the task completion rate under tight time window conditions. As the USV failure rate increased from 0 to 0.4, the SCR of the proposed method decreased from 0.978 to 0.439, a decrease consistent with physical expectations; under the same conditions, the Hungarian SCR dropped sharply from 0.792 to 0.184. Notably, the advantage factor of the proposed method relative to the Hungarian SCR actually increased with the failure rate, rising from 1.23 times at a failure rate of 0% to 2.39 times at a failure rate of 0.4, verifying the robust recovery capability of the SHARP protocol under large-scale failure scenarios. With increasing chain density, the SCR of the proposed method gradually decreased from 0.990 to 0.951, remaining above 0.95 across the entire range; under the same conditions, the Hungarian SCR simultaneously decreased from 0.907 to 0.685, a decrease approximately 5.7 times that of the proposed method. This indicates that the dependency-aware bidding, topology normalization, and chain constraint handling mechanisms can alleviate the dependency propagation pressure caused by increased chain density.
[0114] Based on the above four-dimensional sensitivity scan results, it can be seen that the method of the present invention maintains stable leading performance in all parameter ranges, with no sudden inflection points in the curve. Moreover, the advantage is amplified under extreme conditions such as tight resources, tight windows, and high failure rates, indicating that the method of the present invention has good parameter robustness in engineering deployment.
[0115] IX. Statistical Significance Test. To verify whether the performance advantage of the method of this invention relative to each baseline algorithm has sufficient confidence in a statistical sense, this embodiment performs paired t-tests on the method of this invention and all six baseline algorithms on the SCR index, covering 11 scenarios (static P1-P4 and dynamic S1-S7). Each scenario uses 100 independent seeds, for a total of 66 tests. Table 10 shows the complete paired t-test results for three representative baselines (Hungarian, CCBBA, and Standard-CBBA). Each cell gives the mean difference Δ, the 95% confidence interval, and the p-value; three asterisks indicate a high significance level of p < 0.01.
[0116] Table 10. Results of paired t-tests of SCR between PRISM-CBBA and representative baselines (N=100) Note: Three asterisks in the table indicate a high significance level of p<0.01; Δ represents the difference between the SCR means (PRISM-CBBA minus the baseline); the numbers in square brackets represent the 95% confidence interval.
[0117] As shown in Table 10, in this embodiment, 66 pairs of SCR t-tests were performed on the method of this invention and all six baselines across 11 scenarios, all achieving a highly significant p < 0.01, with all lower bounds of the 95% confidence interval being positive. The most stringent test (P1 vs. Hungarian) had p = 0.0024, while the remaining 65 tests had p < 0.001. This indicates that the SCR advantage of the method of this invention relative to each baseline has sufficient statistical confidence. The difference between the method of this invention and the strongest baseline, Hungarian, exhibits a differentiated characteristic depending on the scenario type: in static scenarios, the Δ range is +0.070 (P1) to +0.207 (P3); in dynamic perturbation scenarios, the Δ difference further widens, with the advantage of S1 mainly stemming from the SHARP recovery and replanning mechanism, while the advantage of S4 mainly comes from rapid replanning after task cancellation and the maintenance of allocation stability.
[0118] Based on the above statistical test results, it can be seen that the SCR advantage of the method of the present invention over various representative baseline algorithms is statistically significant under various working conditions such as static, dynamic, failure and emergency, and the results are reliable and reproducible.
[0119] Based on the nine comparative experiments described in this embodiment, it can be concluded that the intelligent allocation method for multi-stage chained tasks in maritime search and rescue based on a priority-aware two-layer collaborative decision-making architecture of the present invention, through the complementary functions of multiple mechanisms such as the two-layer collaborative decision-making architecture, GNN criticality scoring, secondary decay urgency factor AUD, superlinear load penalty WEB, spatial collaborative reward, chain co-location redistribution, MAYDAY preemption mechanism, and SHARP three-level hierarchical adaptive recovery protocol, can maintain a stable high task completion rate, fast MAYDAY response, high failure recovery rate, and good load fairness under various working conditions such as static scale expansion, dynamic disturbance, mixed pressure, resource saturation, and high chain density. Moreover, it exhibits good robustness and statistical significance across the entire scanning range of key engineering parameters, and is particularly suitable for maritime emergency search and rescue scenarios in deep sea with weak communication, strong dynamics, and strong constraints.
[0120] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making, characterized in that, Includes the following steps: Obtain the status data, task environment data, and chained dependencies of each USV in the heterogeneous unmanned surface vessel (USV) cluster, and construct a constrained multi-stage chained task dependency graph. Using the multidimensional attributes of task nodes as input, a pre-trained graph attention network is used to perform forward reasoning on the multi-stage chained task dependency graph, and output the global criticality score of each task node. For any USV and task pair, a cost matrix is constructed by introducing a criticality discount term, a second-order attenuation urgency factor AUD discount term, and a superlinear load penalty WEB term, with travel time as the basic cost term. Hungarian centralized warm start: Perform minimum cost bipartite graph matching on the cost matrix, repeating multiple times until no new allocation is generated, to generate an initial allocation scheme from a global perspective; Distributed CBBA consensus refinement: Based on the initial allocation scheme, each USV acts as a distributed agent, eliminating inter-vessel allocation conflicts through a bounded round-by-round win-bid consensus process, resulting in an optimized allocation scheme.
2. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, The calculation process for the global criticality score includes the following steps: The task dependency graph is forward propagated using a multi-dimensional node feature matrix consisting of the required capabilities, priority, time window, spatial coordinates, and the number of predecessors and successors in the task chain. The Sigmoid function is used to output the criticality score of each task node. The graph attention network is trained offline using the marginal contribution of the task obtained from Monte Carlo simulation to the global strict task completion rate as the supervision label.
3. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, The process of constructing the cost matrix includes the following steps: For any USV and mission pair, the Euclidean distance between them divided by the maximum speed of the USV is used as the basic travel time cost. Calculate the GNN criticality discount term, the value of which is proportional to the difference between the global criticality score and 0.5; The urgency is defined as a cutoff value minus the ratio of the remaining time of the time window to the width of the time window, and the secondary decay urgency factor AUD discount term is calculated by multiplying the square of the urgency by the priority weight. The urgency is zero when the current time is earlier than the start time of the time window, one when it reaches or exceeds the cutoff time, and monotonically increases when it is in between. The superlinear load penalty WEB term is calculated as a power of 1.5 of the current USV load ratio; Add the basic driving time cost term to the superlinear load penalty WEB term, and then subtract the GNN criticality discount term and the secondary decay urgency factor AUD discount term to obtain the cost matrix elements; If the USV does not have the capabilities required for the mission or exceeds the maximum cruising radius, the corresponding cost matrix element is set to positive infinity.
4. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, The process of generating an initial allocation scheme from a global perspective includes the following steps: Minimum cost bipartite graph matching is performed on the cost matrix, and the process is repeated at most until the upper limit of the USV's binding capacity is reached. In each round, the matched tasks are removed from the candidate set and the cost matrix is reconstructed. The process continues until no new allocation is generated in the current round, at which point the process exits early and outputs the matching results of each round as the initial allocation scheme.
5. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, The process of eliminating inter-vessel allocation conflicts through the consensus-building process of winning bids in bounded rounds includes the following steps: In each round of consensus, each pair of USVs exchanges their winning dictionaries. For any mission currently bundled with the first USV, if the second USV makes a higher bid and considers itself the winner, the first USV releases the mission and updates its local winning bid information to the second USV's bid; Convergence occurs after one round of no task transfer, and the maximum number of consensus rounds is set to a preset value. If conflicts still exist after reaching the maximum number of consensus rounds, a definitive decision will be made based on task priority, bid value, and vessel number.
6. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, After obtaining the optimized allocation scheme, the process also includes chain co-location and spatial collaborative reward redistribution, which includes the following steps: For subsequent tasks in the same task chain, if the marginal cost of the lead ship accepting the subsequent task is lower, provided that the feasibility constraints of unfulfilled load, capacity matching, and reachability are met, then the subsequent task will be migrated from its current USV to the lead ship. The marginal cost is compared by comparing the travel time from the lead ship to the subsequent task with the travel time from the current USV of the subsequent task to the subsequent task plus the basic co-location tolerance threshold plus the spatial cooperation reward. The spatial cooperation reward is proportional to the number of tasks currently bundled with the lead ship that are located within the preset neighborhood radius of the subsequent task.
7. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, After obtaining the optimized allocation scheme, the process also includes a Mayday preemption process: When there is an unassigned MAYDAY emergency task, iterate through the non-executing tasks in the bundle of each USV that has the capability required for the task and has a lower priority than the MAYDAY task. Select the scheme with the largest preemption benefit and a positive value according to the preemption benefit function, so as to release the low-priority task to create execution slots for the unassigned MAYDAY task. The preemption benefit function is equal to the product of the priority weight of the MAYDAY task and the basic reward value, plus twice the response time saved by preemption, minus half the product of the priority weight of the expelled task and the basic reward value, and the travel cost of the candidate USV to the MAYDAY task. The selected MAYDAY task is inserted at the beginning of the preemption boat path.
8. The method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making as described in claim 1, characterized in that, After obtaining the optimized allocation scheme, a strategic hierarchical adaptive recovery process is also included. When a USV failure results in orphan tasks, the following steps are performed: For each combination of orphan tasks and candidate activity USVs, a comprehensive score is calculated by weighting the negative factors of travel time, negative factors of load factor, positive factors of capability specificity, positive factors of high-priority task speed matching reward, positive factors of task priority, and positive factors of time margin reward. Candidate USVs are divided into three levels based on time feasibility and cruise safety, and the candidate USV with the highest comprehensive score is selected from the highest level. Orphan tasks within the same chain are arranged in phase order and processed in batches according to the order within the chain, prioritizing the maintenance of locality within the chain; When the number of assigned tasks in a target USV reaches its bundled capacity limit, the lowest priority non-executing task in the USV bundle is actively expelled to a neighboring USV. Maintain a USV failure count counter within a preset sliding time window. When the counter is less than a first threshold, execute the normal recovery mode. When the counter is not less than the first threshold and is less than a second threshold, execute the medium rebalancing mode to migrate the lowest priority task in the overloaded USVs whose load exceeds the average load of the fleet by more than one task to the lightly loaded USVs. When the counter is not less than the second threshold, execute the aggressive release mode to release the low-priority assigned tasks in the overloaded USVs to the pending execution state.
9. A method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making, as described in claim 1, is characterized in that... After obtaining the optimized allocation scheme, the process also includes topology normalization and scheme output: Perform a topology sort that preserves order on the path of each USV, ensuring that the position index of any preceding mission is strictly less than the position index of the following mission; re-verify the feasibility of the time window and cruising radius, and output the final allocation scheme. And trigger the next round of asynchronous reallocation when a new task is injected, USV fails, or the task graph changes.
10. A method for multi-stage chained task allocation in maritime search and rescue based on priority-aware two-layer collaborative decision-making, as described in claim 1, is characterized in that... During the construction of the multi-stage chained task dependency graph, the USV's state data includes initial position coordinates, maximum speed, maximum cruising radius, capability set, and bundled capacity limit; the task environment data includes task identifier, task type, required capabilities, task location, estimated execution duration, time window start and end times, priority level, task chain number, and stage position index. The priority levels are divided into four levels, corresponding to MAYDAY emergency rescue, PAN-PAN detection / location, towing / patrol, and general patrol / scanning, respectively. A bounded hierarchical weight mapping is adopted so that the four priority levels correspond to four priority weights that increase sequentially, and the ratio of the highest priority weight to the lowest priority weight does not exceed two orders of magnitude. The chain-like dependency relationship is modeled using a directed graph. For tasks belonging to the same chain and with a stage index earlier, there are directed dependency edges pointing to tasks with a stage index later.
Citation Information
Patent Citations
Multi-task distributed task allocation method for cluster aircraft
CN114610058A
Multi-unmanned aerial vehicle distributed dynamic task allocation method based on clustering
CN116149370A
Dynamic task allocation method and system for multiple unmanned ships
CN116993073A
Heterogeneous unmanned ship task dynamic optimization method based on intelligent distributed CBBA
CN121189684A
