Unmanned ship cluster distributed cooperative path planning method and system

Abstract

The invention discloses an unmanned ship cluster distributed cooperative path planning method and system, and the core of the method is that each unmanned ship intelligent body constructs a future space-time risk map through probabilistic prediction; when potential collision conflicts are detected, related unmanned ships do not need to be intervened by a central node, and finally, the party with the lower cost actively undertakes the avoidance responsibility through a distributed negotiation mechanism by bidding the multi-target comprehensive cost paid for avoiding the conflicts, so that an optimal collaborative path for the whole cluster is achieved. According to the method, through probabilistic prediction, distributed negotiation and dynamic multi-objective optimization, the safety, the adaptability and the expandability of unmanned ship cluster path planning are remarkably improved, meanwhile, the communication overhead is greatly reduced, and efficient, safe and intelligent autonomous collaboration in a complex environment is achieved.

Description

Technical Field

[0001] This invention belongs to the field of unmanned vehicle technology, specifically relating to a distributed collaborative path planning method and system for unmanned surface vessel swarms. Background Technology

[0002] With the increasing demands of humankind for marine resource exploration, environmental monitoring, maritime security patrols, and search and rescue missions, the operational capabilities of a single unmanned surface vessel (USV) have gradually revealed their limitations when facing large-scale, highly dynamic mission scenarios. Therefore, swarm systems comprised of multiple USVs, with their significant advantages in operational efficiency, system robustness, and mission coverage, have become a highly anticipated research hotspot in the field of marine technology.

[0003] In the complex process of unmanned surface vessel (USV) swarms collaboratively executing missions, path planning is one of the core key technologies, and the quality of the planning results directly affects the success, efficiency, and overall safety of the mission.

[0004] However, existing technologies still face many insurmountable technical bottlenecks and inherent theoretical flaws when attempting to solve the problem of collaborative path planning for large-scale unmanned surface vessel swarms in complex, variable, and uncertain real-world marine environments. The limitations of existing technologies are mainly reflected in the following aspects: First, traditional path planning algorithms are severely lacking in adaptability. Classic algorithms such as A* algorithm, Dijkstra's algorithm, and dynamic window method can plan feasible paths in static scenarios where environmental information is completely known. However, when faced with complex real marine environments full of uncertainties, dynamic obstacles, and multiple constraints such as energy consumption, time, and safety, their adaptability and real-time performance are seriously inadequate. These methods are usually classified as global path planning or local path planning.

[0005] Global methods, such as the A* algorithm, require frequent global replanning in dynamic environments, which incurs huge computational overhead and makes it difficult to meet the requirements of high real-time tasks. On the other hand, local methods, such as the dynamic window method, although fast, are prone to getting trapped in local optima due to the lack of a global perspective, and cannot guarantee the global optimality of the final path.

[0006] More importantly, when it is necessary to optimize multiple conflicting objectives simultaneously, such as the shortest time, the lowest energy consumption, and the highest security, these traditional methods often use a fixed, pre-defined weighted summation method to construct the cost function. This static weight allocation strategy cannot be adaptively adjusted according to real-time changes in the task context and environment, resulting in suboptimal paths generated in dynamically changing task scenarios, which are difficult to meet the refined requirements of complex tasks.

[0007] Secondly, the artificial potential field method, as a widely studied local obstacle avoidance method, has attracted much attention due to its simple model and low computational cost. However, its inherent theoretical defects are fatal in practical applications.

[0008] The most prominent problem is that in environments with complex obstacles or multiple obstacle avoidance targets, the artificial potential field method is prone to falling into the "local optimum trap," that is, the unmanned surface vessel stops at the point where the resultant force of gravity and repulsion is zero, or oscillates back and forth in front of the obstacle, and can never reach the target point smoothly.

[0009] Furthermore, traditional artificial potential field methods require relatively complete prior knowledge of environmental obstacles to construct an effective potential field, making it difficult to cope with unknown or time-varying obstacles that may appear in the marine environment at any time, such as other ships and floating ice, resulting in extremely poor environmental adaptability.

[0010] Furthermore, the application of deep reinforcement learning methods faces bottlenecks. While deep reinforcement learning provides powerful learning capabilities and decision optimization potential for unmanned surface vessel (USV) swarm path planning, its application in existing technologies is still in its early stages and faces three major bottlenecks.

[0011] First, in multi-agent cooperative scenarios, especially when multiple mutually constrained objectives need to be met, such as collision avoidance, formation, and energy conservation, the design of the reward function becomes extremely complex. An unreasonable reward function design can easily lead to policy conflicts between different objectives, causing continuous policy oscillations and making it difficult for the algorithm to converge stably to the optimal solution.

[0012] Secondly, deep reinforcement learning models are data-driven, and their performance is highly dependent on massive amounts of interactive training data. However, large-scale field training in real ocean environments is extremely costly and risky; therefore, most existing research trains models in simulation environments. However, simulation environments often cannot fully reproduce the complex dynamic characteristics of the real ocean, such as ocean currents, waves, and sensor noise. This leads to a significant decrease in the generalization ability and robustness of models that perform well in simulations, or even complete failure, when deployed in actual situations, posing serious safety risks.

[0013] Third, most current collaborative methods based on deep reinforcement learning adopt a "centralized training, distributed execution" framework. The centralized decision-making under this framework places extremely high demands on the communication bandwidth, latency, and computing power of the central node. As the cluster size increases, the joint state space and action space grow exponentially, the so-called "curse of dimensionality," making centralized methods unsuitable for large-scale cluster applications. While the distributed decision-making framework alleviates the pressure on the central node to some extent, it still requires the design of an efficient collaborative communication mechanism to share states and coordinate actions. Existing methods have failed to effectively balance communication overhead and collaborative efficiency, often leading to collaborative failures due to communication overload or untimely information.

[0014] Finally, swarm coordination and communication mechanisms have inherent flaws. Most existing studies on swarm coordination control are based on the assumption of ideal communication, systematically ignoring real-world factors such as communication delays, data packet loss, and time-varying network topology in the actual marine communication environment. These factors can severely weaken the control performance of the swarm and even lead to the instability of the entire closed-loop system.

[0015] Furthermore, many collaborative frameworks employ a fixed leader-follower model. In this model, the planning process of the entire cluster can stall once the leader is blocked on its path or stops upon reaching the goal, especially in complex environments with numerous obstacles. Existing communication mechanisms also lack effective assessment of information value, resulting in a flood of redundant, low-value, or even useless information in communication channels. This not only consumes valuable bandwidth resources but may also interfere with decision-making at critical moments.

[0016] In summary, there is a fundamental contradiction in existing technologies: distributed methods that are computationally simple and easy to implement, such as artificial potential field methods, lack sufficient security and effectiveness in complex dynamic environments; while theoretically more intelligent methods, such as centralized deep reinforcement learning or global optimizers, lack scalability and robustness in real-world environments due to their huge computational and communication overhead.

[0017] There is an urgent need for a new technological solution that can break down this binary opposition and achieve intelligent, safe, scalable, and cost-effective unmanned surface vessel (USV) swarm collaborative path planning. Summary of the Invention

[0018] The purpose of this invention is to overcome the aforementioned defects and fundamental contradictions in the existing technology and provide a distributed collaborative path planning method for unmanned surface vessel (USV) swarms. This method aims to achieve a highly scalable, highly adaptive, and highly robust distributed collaborative framework without relying on a centralized decision-making node.

[0019] Specifically, the technical problem this invention aims to solve is how to enable unmanned surface vessel (USV) swarms to optimize multiple conflict objectives, including navigation safety, mission efficiency, energy consumption, and compliance with international maritime collision avoidance rules, in real time and dynamically with extremely low communication overhead in complex, dynamic, and uncertain marine environments, thereby generating a safe, efficient, and economical optimal collaborative path.

[0020] To achieve the above objectives, the present invention provides the following technical solution: A distributed collaborative path planning method for unmanned surface vessel (USV) swarms based on hierarchical predictive negotiation is proposed. The core of this method lies in equipping each USV agent in the swarm with a fully distributed decision-making framework consisting of three core technology layers. This framework abandons the dependence on a single deterministic path and instead adopts probabilistic prediction and risk assessment; it abandons centralized control or simple reactive collision avoidance and instead adopts a distributed negotiation mechanism based on economic principles; and it abandons blind information broadcasting and instead adopts a gated communication strategy based on information value.

[0021] The aforementioned hierarchical predictive negotiation framework specifically includes the following three collaborative technical layers: The first layer: Probabilistic trajectory prediction and spatiotemporal risk assessment layer based on Transformer.

[0022] The second layer: a multi-objective negotiation layer based on fuzzy adaptive "change cost".

[0023] The third layer: consensus and traceable decision-making layer based on information value gating.

[0024] Specifically, the present invention provides a distributed cooperative path planning method for unmanned surface vessel swarms, comprising the following steps: First, each unmanned surface vessel (USV) agent acquires its own historical state sequence data and that of other USV agents within its communication range. Based on this data, it uses a trajectory prediction model to make a probabilistic prediction of its own spatial position in the future prediction time domain, generating probability distribution information that can characterize the uncertainty of its future position.

[0025] Subsequently, based on this probability distribution information, the unmanned surface vessel intelligent agent constructs a spatiotemporal risk map that can quantify the risk of collisions in the future.

[0026] Next, by sharing or receiving spatiotemporal risk maps among unmanned surface vessel agents, the system can detect whether there are spatiotemporal conflict areas where the probability of joint conflict exceeds a preset conflict triggering threshold.

[0027] When such a spatiotemporal conflict zone is detected, the unmanned surface vessel agents involved in the conflict will enter a negotiation and decision-making state. At this time, each unmanned surface vessel agent participating in the negotiation will calculate and declare a comprehensive change cost for deviating from the current planned path to resolve the conflict.

[0028] Finally, all participating unmanned surface vessel agents determine the final cooperative avoidance strategy based on the comprehensive change costs declared by each other, following a pre-set negotiation protocol, and execute the strategy to complete the cooperative path planning.

[0029] Preferably, the trajectory prediction model is a deep learning model based on the Transformer architecture. The advantage of this model is its ability to efficiently process temporal historical state sequence data and output a set of hybrid density network parameters to describe future locations, thereby accurately capturing multiple possibilities for the trajectory.

[0030] More preferably, the spatiotemporal risk map is a three-dimensional data structure, whose three dimensions represent two-dimensional spatial coordinates and time, respectively; in this data structure, the value stored in each voxel represents the specific probability value of the unmanned surface vessel agent appearing in the corresponding geographic grid at the corresponding future time step.

[0031] Furthermore, the comprehensive change cost is referred to as... Its calculation formula is designed as a multi-objective cost function. The specific mathematical expression is: In the formula, , , , These represent the weighting coefficients for each cost item. This represents the additional energy consumption resulting from implementing the avoidance strategy compared to the original planned path. This represents the overall task time delay caused by implementing the avoidance strategy. This is a key cost item, quantifying the degree to which the proposed avoidance strategy conforms to international maritime collision avoidance rules. Violations will result in extremely high costs for this item. This is a penalty item used to measure the potential negative impact of path changes on the current task priority.

[0032] Furthermore, to achieve a high degree of adaptability, the weighting coefficients in the above formula... , , , It is not fixed.

[0033] These systems are adaptively adjusted in real time by a built-in fuzzy logic controller. This controller receives and processes two types of input information: one is real-time external environment information, which includes at least the obstacle density around the unmanned surface vessel (USV) and the current sea state; the other is the USV's own state information, which includes at least the USV's current remaining energy percentage and its distance from the final target. In this way, the USV can intelligently weigh the importance of various objectives in different situations.

[0034] As another preferred embodiment of the present invention, an information value assessment step is set before the unmanned surface vessel enters the negotiation and decision-making state in order to optimize communication resources.

[0035] Specifically, before deciding whether to initiate a negotiation, the unmanned surface vessel (USV) agent will first calculate the expected benefits of the communication negotiation, mainly the degree of reduction in collision risk, and also assess the communication costs required, including energy and time.

[0036] Using these two factors, the system can construct an information value assessment metric, which we call... Only when this A positive evaluation value indicates that the expected benefits significantly outweigh the communication costs, and only then will the system actually trigger the subsequent negotiation and decision-making process, thereby avoiding a large amount of unnecessary communication from the source.

[0037] As another preferred embodiment of the present invention, the preset negotiation protocol is designed to be very simple and efficient: among all the unmanned surface vessels participating in the negotiation, the one that declares the lowest overall change cost... The unmanned surface vessel (USV) agent that receives the highest value will be identified as the primary avoidance party and will bear the main responsibility for avoidance. This mechanism, based on economic bidding, can resolve conflicts quickly and effectively.

[0038] As another preferred embodiment of the present invention, in order to enhance the interpretability and post-analysis capability of the system, the method further includes a decision log recording step.

[0039] Throughout the entire negotiation and decision-making process, all key information, such as the spatiotemporal risk map snapshot triggered by the negotiation, the assessment results of the information value, the bids in each round of the negotiation process, the detailed input and output data of the fuzzy logic controller for weight adjustment, and the finally adopted collaborative avoidance strategy, will be structured and completely recorded in the local non-volatile storage device of each unmanned surface vessel agent.

[0040] This invention also provides a distributed cooperative path planning system for unmanned surface vessel (USV) swarms. This system is typically deployed within an USV agent and includes: a state prediction module, a risk assessment module, a negotiation and decision-making module, and a strategy execution module. The functions of these modules correspond to the various steps in the aforementioned method, working collaboratively to achieve the technical solution of this invention.

[0041] Compared with the prior art, the present invention has the following advantages: This invention not only solves the inherent "curse of dimensionality" and single point of failure problems of traditional centralized planning methods, but also overcomes the drawbacks of simple distributed methods that are prone to getting trapped in local optima and lack security guarantees. Instead, it deeply and organically integrates multiple technologies such as probabilistic trajectory prediction, multi-objective optimization, economic game theory, and value communication through an original "hierarchical predictive negotiation" framework. This significantly improves the decision-making intelligence, navigation safety, and system robustness of unmanned surface vessel (USV) swarms in complex, dynamic, and uncertain marine environments. In particular, its innovative distributed negotiation mechanism based on "change cost" and information value-gated communication strategy provide an effective and feasible new paradigm for solving the two core problems of decision-making conflict and communication overload that are common in large-scale swarm collaboration. This enables the entire swarm system to spontaneously emerge with near-optimal global collaborative paths that meet international maritime collision avoidance rules with extremely low communication costs and computational overhead, without any intervention from any central node. This endows large-scale USV swarms with unprecedented high scalability, strong adaptability, and mission execution efficiency. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0043] Figure 1 This is a schematic diagram of the overall process of a distributed collaborative path planning method and system for unmanned surface vessel swarms according to the present invention; Figure 2 This is a schematic diagram of a distributed collaborative path planning method for unmanned surface vessel (USV) swarms and a hierarchical control framework for a single USV agent in the system, as described in this invention. Figure 3 This is a schematic diagram of the state machine of the "change cost" negotiation protocol in the distributed collaborative path planning method and system for unmanned surface vessel swarms of the present invention. Figure 4 This is a schematic diagram of a distributed collaborative path planning method for unmanned surface vessels (USVs) swarms and a system in which the spatiotemporal probability risk maps of two USVs overlap, triggering a negotiation mechanism. Figure 5This is a schematic diagram of the input-output relationship of the fuzzy logic adaptive weight adjustment module in the distributed collaborative path planning method for unmanned surface vessel swarms and the system of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] In implementing this invention, it is first necessary to perform mathematical and computational modeling of the unmanned surface vessel swarm system and its operating environment.

[0046] Please see Figure 2 As shown, the unmanned surface vessel (USV) intelligent agent is defined as each USV in the cluster being abstracted as an independent intelligent agent with perception, decision-making and action capabilities; each intelligent agent 100 is equipped with an onboard computing unit 110, which is the core of executing the hierarchical predictive negotiation method described in this invention.

[0047] In addition, the intelligent agent is equipped with a standard sensor suite 120, including but not limited to a global positioning system for obtaining its own position, an automatic identification system for receiving information about surrounding vessels, and radar and sonar for detecting static and dynamic obstacles; the communication module 130 is responsible for exchanging data with nearby unmanned surface vessel intelligent agents, and the propulsion and control system 140 is responsible for executing navigation commands output by the computing unit 110.

[0048] To perform route planning and risk assessment, digital modeling of the mission area is required. This invention preferably employs a grid method for environmental modeling, specifically dividing the target mission area into a series of uniformly sized two-dimensional or three-dimensional grid cells.

[0049] Each grid cell stores two types of information: static information, including the grid's geographic coordinates, water depth, whether it is land or a known static obstacle, such as an island or drilling platform, and whether it is a no-navigation zone. This information can be pre-loaded from electronic nautical charts; and dynamic information, including real-time or predicted environmental data for the grid, such as ocean current speed and direction, wind field information, etc. This data can be obtained in real time from weather forecasting systems or through sensors.

[0050] This invention adopts a fully distributed network topology, which means that there is no central control station in the system. Each unmanned surface vessel (USV) agent only needs to communicate point-to-point with its neighboring nodes within its communication range. This architecture has good scalability and robustness. The communication protocol is designed to support dynamic topology changes. That is, when a new USV joins the cluster, a member leaves, or communication is interrupted due to distance, the network topology can be adaptively reconstructed without affecting the operation of the entire system.

[0051] Please see Figure 2 As shown, the first layer of the method of this invention is the probabilistic trajectory prediction and spatiotemporal risk assessment layer 111. The purpose of this layer is to deduce uncertain future risks from the deterministic current state, providing a basis for subsequent negotiation and decision-making. This process breaks through the limitations of traditional path planning methods that rely on a single, deterministic trajectory, fundamentally improving the system's ability to cope with uncertainty.

[0052] Traditional methods for predicting future trajectories typically output a single path, which fails to reflect the diversity of decision-making in the real world. For example, at a waterway intersection, a ship may go straight, turn left, or turn right. This invention employs a more advanced model to capture this multimodal behavior.

[0053] Each unmanned surface vessel (USV) agent 100 deploys a trajectory prediction model based on the Transformer architecture. The Transformer model is well-suited for processing time-series trajectory data due to its powerful sequential data processing capabilities and ability to capture long-distance dependencies. The input to the model is the USV's own historical state sequence over a period of time, including position, speed, heading, acceleration, etc., as well as state information received from neighboring USVs through the communication module 130.

[0054] One of the key innovations of this invention is that the output layer of the Transformer model does not directly regress future coordinates, but is combined with a hybrid density network or a similar generative model; the output of the hybrid density network is not a single predicted value, but a set of parameters describing the probability distribution, typically the mean, standard deviation and mixing coefficients of multiple Gaussian components.

[0055] Therefore, the model ultimately outputs the probability density function of the unmanned surface vessel's position in the future predicted time domain T. This probabilistic output can effectively represent and quantify future uncertainties and capture multiple possibilities of navigation behavior, i.e., multimodal characteristics. For example, at a certain decision point, the model may predict two probability peaks, corresponding to two high-probability future trajectories, left turn and right turn, respectively.

[0056] After obtaining the probability distribution of future locations, the next step is to transform it into a structured representation that can be used for decision-making, namely a spatiotemporal risk map.

[0057] First, the prediction time domain T is discretized into Each time step is denoted as . At each time step Using the probability density function obtained in the first step, we can calculate the location of the unmanned surface vessel in any grid cell of the environmental model. probability .

[0058] Then, these probability values ​​are integrated to construct a three-dimensional spatiotemporal risk map, namely two spatial dimensions plus one temporal dimension. The value of each voxel in this map represents the future... At that moment, the unmanned surface vessel occupied the grid. This could potentially lead to a collision. The simplest risk value can be directly equal to the probability value. In more complex implementations, the risks can be weighted by factors such as the size, speed, and maneuverability of the unmanned surface vessel. For example, the faster the speed, the larger the potential impact range, and the higher the risk value should be.

[0059] Please see Figure 4 As shown, conflict detection is no longer a simple geometric problem of whether two line segments intersect, but a probabilistic problem of whether two probabilistic bodies overlap; when unmanned surface vessel A (the vessel itself) receives its own spatiotemporal risk map broadcast by unmanned surface vessel B (the other vessel) via communication... Then, it will compare the risk map with its own risk map. Superimpose them.

[0060] Define any spacetime point joint conflict probability This is the product of two risk values, assuming their movements are independent: The system will traverse the entire spatiotemporal risk map and detect any joint conflict probability at any spatiotemporal point. The conflict trigger threshold was exceeded. ,Right now If a collision occurs, the system determines that there is an unacceptable risk of collision. At this point, the first layer completes its work and transmits the conflict information, including the conflict area, time, and involved parties, to the second layer to initiate a negotiation resolution mechanism.

[0061] Please see Figure 2As shown, the second layer of the method of the present invention is the adaptive multi-objective negotiation layer 112, which is the core innovation of the present invention. It transforms the complex path conflict problem into a distributed, multi-party negotiation process based on economic principles. Its fundamental idea is that the allocation of avoidance responsibility should not be a simple geometric rule, but a dynamic optimization process that comprehensively considers task objectives, its own state and environmental constraints.

[0062] To achieve efficient and fair distributed conflict resolution, this invention innovatively employs a method based on "changing costs," namely... The core logic of the competitive negotiation agreement is that, among the conflicting parties, the one who pays less "cost" to avoid a collision by changing their original plan should bear more responsibility for avoiding conflict.

[0063] The execution of this protocol follows a defined finite state machine; please refer to [link / reference]. Figure 3 As shown, the state machine defines the various stages of the negotiation process, including: idle state 301, negotiated state 302, proposal sent state 303, proposal received state 304, consensus reached state 305, and negotiation failed state 306. The transitions between states are triggered by specific events, such as receiving a negotiation request or timeout, which ensures the orderliness and controllability of the negotiation process.

[0064] To achieve low-overhead communication in bandwidth-constrained marine environments, the negotiation protocol employs a highly structured and lightweight message format; each message contains standardized fields such as: message type (request, proposal, accept, reject, etc.); initiator ID; receiver ID; a unique negotiation session ID; and proposal content, such as... Quotations, suggested spacetime tunnels, etc.; and digital signatures used to verify identity and prevent tampering.

[0065] The calculation is the cornerstone of the negotiation process. When an unmanned surface vessel agent needs to evaluate the cost of deviating from its current optimal predicted path, i.e. the highest probability path predicted in the first layer, to avoid the conflict zone, it calculates a comprehensive multi-objective cost function. This function transforms the abstract "cost" into a concrete and comparable value.

[0066] The mathematical expression for this cost function is: in, The energy consumption increment represents the additional energy consumed in executing the avoidance path compared to the original path. This value is calculated based on a precise unmanned surface vessel (USV) dynamics and energy consumption model. This model needs to consider the hull's hydrodynamic characteristics, propulsion system efficiency, and the influence of environmental factors such as currents and wind fields. For example, sailing against the current or frequent acceleration and deceleration can lead to energy consumption increases. Significant increase.

[0067] Mission time delays represent the additional travel time incurred by executing avoidance paths. This cost is significant for missions with strict time windows, such as arriving at a rendezvous point on time.

[0068] Safety and regulatory compliance costs are a key penalty item. If a proposed avoidance path violates the International Maritime Collision Avoidance Regulations (IMDG), such as turning to port in a head-on collision situation, or causing the safe distance from other vessels or obstacles to be less than the preset minimum safe distance, this item will be assigned a very large positive value, thus effectively rejecting it in negotiations.

[0069] Task priority penalty: Different tasks have different priorities. For high-priority tasks, such as emergency search and rescue, any deviation from the planned action may lead to serious consequences, and therefore a fixed, high cost penalty will be imposed.

[0070] Through calculation Each unmanned surface vessel can quantify the "pain level" of its own avoidance maneuver, and in the negotiation process, The party with the lower value naturally becomes the more suitable party to avoid collisions; this mechanism cleverly transforms the complex collision avoidance decision-making problem into a clear, cost-minimizing economic game problem.

[0071] A major drawback of existing multi-objective optimization techniques is the weighting coefficients, such as those in the above formula. , , , These weights are usually statically preset and cannot adapt to dynamically changing environments and task requirements. This invention introduces a fuzzy logic controller to realize real-time adaptive adjustment of these weights, thereby giving the unmanned surface vessel agent human-like context-based decision-making capabilities.

[0072] Please see Figure 5 As shown, the working principle of this fuzzy logic controller is as follows: The controller receives multiple real-time input variables describing the current "situation" and fuzzifies them. These input variables include: the density of nearby obstacles, calculated from sensor data and categorized into fuzzy sets such as "low", "medium", and "high"; the distance to the final target, categorized into "far", "medium", and "near"; the percentage of remaining energy, categorized into "sufficient", "moderate", and "low"; and the current sea state level, assessed based on data such as wind speed and wave height, categorized into "good", "moderate", and "severe".

[0073] Pre - define a set of "IF - THEN" rules based on expert experience, which link situations with decision preferences. For example: Rule 1, IF the obstacle density is "high" AND the distance to the target is "near" THEN the safety weight is "very high" AND the energy consumption weight is "low", which means that in crowded inshore waters, safety is the top priority and some energy consumption can be sacrificed. Rule 2, IF the remaining energy is "low" AND the sea state level is "severe" THEN the energy consumption weight.

[0074] is "very high" AND the time weight is "medium", which means that when the power is insufficient and the environment is harsh, the primary goal is to save energy to ensure the ability to return, and a certain time delay can be accepted.

[0075] The controller performs fuzzy inference based on the input fuzzy values and the rule base to obtain each output weight 、 、 、 of the fuzzy results. Finally, through a defuzzification process, such as the centroid method, an accurate and normalized weight value is calculated for the calculation at the current moment.

[0076] This mechanism makes the calculation no longer rigid, but dynamic and intelligent. The unmanned boat can automatically adjust the focus of its decision - making according to its own and the real - time situation of the environment, so as to make more reasonable and optimal collaborative decisions in various complex situations.

[0077] Please refer to Figure 2 shown. The third layer of the method of the present invention is the information value gating, consensus and traceability layer 113, which mainly solves two problems: how to avoid unnecessary communication to achieve large - scale scalability, and how to ensure the reliability and transparency of decisions.

[0078] In a multi - agent system, unrestricted communication is one of the main reasons for the inability of the system to scale. To solve this problem, the present invention introduces information value, that is , which is evaluated as the "gatekeeper" of communication. Its core idea is that communication has a cost, and only when the benefit brought by communication is greater than the cost, the communication behavior is reasonable.

[0079] Before the unmanned boat agent decides whether to initiate a negotiation, it will first perform an evaluation.

[0080] The calculation of can be formalized as: in: This represents the expected utility after communication. It represents the expected benefit of communicating with another vessel and successfully reaching an avoidance agreement. This benefit is mainly reflected by the probability of successfully avoiding a collision. The lower the probability of a collision, the higher the utility.

[0081] This represents the expected utility of not communicating, which means that if communication is abandoned and the current best predicted path is continued, the utility will be significantly reduced due to the known risk of collision.

[0082] This represents the communication cost, which includes the energy, computing resources, and time required to initiate and conduct a complete negotiation process.

[0083] Only when calculated In other words, the expected benefits of communication, primarily improved security, must cover the costs before the unmanned surface vessel (USV) agent will actually initiate a negotiation request through communication module 130. In many low-risk encounter scenarios, although there may be slight overlap in spatiotemporal risk maps, the potential collision probability is extremely low. The calculation result may be negative, thus the agent actively suppresses communication behavior. Gating mechanisms can filter out massive amounts of low-value interaction requests at the source, ensuring that valuable communication and computing resources are used to resolve truly critical, high-risk conflicts. This greatly reduces the communication load of the entire cluster without sacrificing security, making large-scale deployment possible.

[0084] Once the two conflicting parties reach an agreement through a second-level negotiation mechanism, for example, if unmanned surface vessel A agrees to turn right while unmanned surface vessel B maintains its course, this final decision needs to be reliably implemented and the potentially affected local neighbors notified.

[0085] To this end, this invention employs a lightweight distributed consensus protocol, such as a simplified variant of Paxos or Raft, or a simple acknowledgment-response mechanism. The unmanned surface vessels (USVs) that reach an agreement will broadcast their final collaborative solution—their respective committed, adjusted spatiotemporal corridors—to their one-hop neighbors. Upon receiving this information, the neighboring nodes will update their understanding of the future intentions of the two USVs, thus incorporating this information as a new constraint in their own path planning, thereby avoiding secondary conflicts caused by information inconsistencies.

[0086] To enhance the system's transparency, credibility, and post-analysis capabilities, this invention features a decision log mechanism. Throughout the entire "prediction-evaluation-negotiation-decision" closed loop, all key intermediate data and final results are structured and recorded in the local non-volatile storage of each unmanned surface vessel.

[0087] The log records, but is not limited to: a snapshot of the spatiotemporal risk map that triggered the negotiation, including conflict areas and probability values; and information prior to initiating the negotiation. Calculation results; each round of offers during the negotiation process, i.e., the offers made by each party that include... Value proposal; fuzzy logic weight adjustment module in each The detailed inputs during the calculation include the environment, state, and output, namely the weights of each objective; the final agreement reached and the adopted path adjustment scheme.

[0088] This decision log is of dual importance. First, it provides detailed data support for performance evaluation and algorithm optimization after the task is completed. Engineers can review each negotiation process, analyze the rationality of the decisions, and make adjustments accordingly. The model and fuzzy rule base are iteratively improved; secondly, and more importantly, the log provides the foundation for achieving explainable artificial intelligence; when the system makes a critical decision, such as a sharp turn, the operator can clearly trace the complete logical chain of the decision by consulting the log: from risk prediction to value judgment, to cost trade-offs and the final negotiated result; this traceability greatly enhances human understanding and trust in the behavior of complex autonomous systems, and is a core element for promoting the acceptance and application of unmanned systems in critical missions.

[0089] To illustrate the beneficial effects of the present invention more specifically, a complex multi-ship encounter scenario that conforms to international maritime collision avoidance rules will be used as an example below.

[0090] Suppose that unmanned surface vessel A is sailing and its course is simultaneously at risk of collision with two other vessels: unmanned surface vessel B is approaching from starboard ahead, and unmanned surface vessel C is approaching from starboard. According to the International Regulations for Preventing Collisions at Sea (ICMLS), unmanned surface vessel A should turn to starboard to avoid unmanned surface vessel B. Simultaneously, as the vessel giving way, it should also take action, typically slowing down or turning to starboard, to pass behind unmanned surface vessel C. This is a typical scenario that places high demands on decision-making algorithms.

[0091] The performance of existing technologies in this scenario: For the artificial potential field method, in this scenario, both unmanned surface vessel B and unmanned surface vessel C will generate strong repulsive forces on unmanned surface vessel A. These two repulsive forces are likely to form a region in front of unmanned surface vessel A with zero net force or an unclear direction, i.e., a potential field "trap". As a result, unmanned surface vessel A is likely to hesitate in place, oscillate back and forth, or even stop completely, unable to make any effective avoidance actions, ultimately leading to a collision.

[0092] For traditional deep reinforcement learning methods, if the model has not been specifically trained in a large number of similar and diverse scenarios, its behavior will be unpredictable. It may oscillate between the two objectives of "avoiding B" and "avoiding C" due to improper design of the reward function, and may even make dangerous actions that violate international maritime collision avoidance rules in order to pursue the shortest path, such as trying to pass between two ships or turning to the left.

[0093] The specific implementation process of the method of the present invention in actual use is as follows: In the prediction and risk identification process, the first layer of unmanned surface vessel A, namely the prediction layer, continues to operate. Its Transformer-MDN model not only predicts its most likely straight path, but also predicts the probabilistic trajectories of B and C based on information obtained from automatic identification systems. The spatiotemporal risk map generated by the system clearly shows that within a certain time window in the future, the risk area of ​​unmanned surface vessel A will overlap with the risk areas of unmanned surface vessels B and C with a high probability.

[0094] conduct Gating and negotiation initiation, the third layer of unmanned surface vessel A, namely The gating layer evaluates the value of communication negotiation with UAVs B and C separately. Due to the high predicted collision probability, the expected utility of not communicating is extremely low; therefore, both... The calculation results were all much greater than zero; the system immediately initiated negotiation requests to both unmanned surface vessel B and unmanned surface vessel C.

[0095] The second layer of the unmanned surface vessel A, namely the negotiation layer, begins its multi-objective negotiation process.

[0096] Negotiations were conducted with the encountering unmanned vessel B. Both parties were aware of the encounter rules in the International Regulations for Preventing Collisions at Sea (ICMLS); if either party proposed a turn to port as a avoidance maneuver, its... In calculation The cost of each item would be astronomical, so such options were ruled out in the first round of bidding; both parties calculated the cost of turning right at different angles. Assuming both sides have similar task priorities and states, they are likely to reach a fair agreement with the lowest cost of "each turning slightly to the right".

[0097] Negotiations were conducted with the intersecting unmanned surface vessel (USV) C. According to the International Regulations for Preventing Collisions at Sea (ICMLS), USV A was the give-way vessel, and USV C was the vessel proceeding in the straight course. Therefore, any changes in course or speed calculated by USV C were within its jurisdiction. All will contain huge The penalty clause would be very expensive; however, as the yielding vessel, unmanned surface vessel A's evasive action was compliant, therefore its calculated right turn or deceleration plan was reasonable. Zero, The value is relatively much lower.

[0098] During the fuzzy logic weight adjustment process, the fuzzy logic layer of unmanned surface vessel A detects multiple nearby threatening vessels in the environment, indicating a "high" obstacle density, and automatically adjusts the safety weights accordingly. and rule compliance weight Adjust to the highest level to ensure all calculated... Safety and regulations are always given top priority.

[0099] Through two parallel negotiations, UAV A will almost inevitably reach an agreement with UAV B to "turn right together" and with UAV C to "take unilateral evasive action". UAV A's decision-making system will integrate these two agreements and plan a single, smooth right-turn path that meets all constraints, which can safely avoid B and C in sequence.

[0100] The decision log records the entire decision-making process, including the risk map, Values, with B and C in each round The pricing, details of the fuzzy logic weighting adjustments, and the final trajectory plan were all fully recorded in the decision log of the unmanned surface vessel A, providing complete transparency for subsequent analysis.

[0101] As can be seen from this embodiment, the method of the present invention can decompose the complex, multi-constraint collision avoidance problem into a series of clear, distributed negotiation processes based on quantified costs and values, thereby finding the globally optimal solution while ensuring safety and compliance, which is difficult for existing technologies to achieve.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A distributed collaborative path planning method for unmanned surface vessel swarms, characterized in that, Includes the following steps: Step S1: Each unmanned surface vessel (USV) acquires its own historical state sequence data and that of other USVs within its communication range. Based on the historical state sequence data, it uses a trajectory prediction model to make a probabilistic prediction of its own spatial position in the future prediction time domain, generating probability distribution information that characterizes the uncertainty of its future position. Step S2: Based on the probability distribution information generated in step S1, the unmanned surface vessel intelligent agent constructs a spatiotemporal risk map that quantifies the collision risk in the future. Step S3: The unmanned surface vessel (USV) agent detects whether there are spatiotemporal conflict areas where the probability of joint conflict exceeds a preset conflict triggering threshold by sharing or receiving spatiotemporal risk maps from other USV agents. Step S4: When the spatiotemporal conflict area is detected, the unmanned surface vessel agents involved in the conflict enter a negotiation decision state. Each unmanned surface vessel agent participating in the negotiation calculates and announces a comprehensive change cost for deviating from the current planned path to resolve the conflict. In steps S5 and S6, the unmanned surface vessel (USV) agent determines a cooperative avoidance strategy based on the comprehensive change cost declared by all participating USV agents, following a preset negotiation protocol, and executes the cooperative avoidance strategy to complete the cooperative path planning.

2. The distributed collaborative path planning method for unmanned surface vessel swarms according to claim 1, characterized in that, In step S1, the trajectory prediction model is a deep learning model based on the Transformer architecture. This model can process temporal historical state sequence data and output hybrid density network parameters describing future locations.

3. The distributed cooperative path planning method for unmanned surface vessel swarms according to claim 1, characterized in that, In step S2, the spatiotemporal risk map is a three-dimensional data structure with dimensions of spatial two-dimensional coordinates and time. The value of each voxel in this data structure represents the probability that the unmanned surface vessel agent appears in the corresponding geographic grid at the corresponding time step.

4. The distributed collaborative path planning method for unmanned surface vessel swarms according to claim 1, characterized in that, The calculation of the comprehensive change cost is based on a multi-objective cost function that incorporates at least two of the following cost components: the energy consumption increment resulting from implementing the avoidance strategy compared to the original path planning; the mission time delay resulting from implementing the avoidance strategy; and the compliance cost of the avoidance strategy with international maritime collision avoidance rules. And the penalty cost caused by changes in path priority to tasks.

5. The distributed collaborative path planning method for unmanned surface vessel swarms according to claim 4, characterized in that, The weights of each cost component in the multi-objective cost function are not fixed values, but are adaptively adjusted by a fuzzy logic controller based on the current real-time environmental information and the unmanned surface vessel's own state information. The real-time environmental information includes at least the density of nearby obstacles and the current sea state level. The unmanned surface vessel's own state information includes at least the unmanned surface vessel's current remaining energy and its distance from the final target.

6. The distributed collaborative path planning method for unmanned surface vessel swarms according to claim 1, characterized in that, Before step S4, there is an information value assessment step: before initiating negotiation, the unmanned surface vessel agent first calculates the expected benefits and communication costs that may be brought about by this communication negotiation, and constructs an information value assessment quantity. Only when the information value assessment quantity is positive will the negotiation decision state be triggered.

7. The distributed cooperative path planning method for unmanned surface vessel swarms according to claim 1, characterized in that, The preset negotiation agreement determines the allocation of avoidance responsibility based on the comprehensive change cost value declared by each participating party. A preferred agreement is to allocate the main avoidance responsibility to the unmanned surface vessel agent with the lowest declared comprehensive change cost value.

8. A distributed cooperative path planning method for unmanned surface vessel swarms according to any one of claims 1-7, characterized in that, The method also includes a decision log recording step: throughout the negotiation decision process, key information such as the spatiotemporal risk map snapshot that triggers the negotiation, the information value assessment results, the bids in each round of negotiation, the inputs and outputs of the fuzzy logic weight adjustment, and the finally reached collaborative avoidance strategy are structured and recorded in the local non-volatile storage of each unmanned surface vessel agent.

9. A distributed cooperative path planning system for unmanned surface vessel (USV) swarms, applied to USV intelligent agents, characterized in that, include: The state prediction module is used to acquire historical state sequence data of itself and other unmanned surface vessels within the communication range, and based on the historical state sequence data, to make a probabilistic prediction of its spatial position in the future prediction time domain using a trajectory prediction model, thereby generating probability distribution information that characterizes the uncertainty of the future position. The risk assessment module is used to construct a spatiotemporal risk map that quantifies the collision risk in the future spatiotemporal space based on the probability distribution information, and to detect whether there are spatiotemporal conflict areas where the probability of joint conflict exceeds a preset conflict triggering threshold by sharing or receiving spatiotemporal risk maps from other unmanned surface vessels. The negotiation and decision-making module is used to initiate negotiation with relevant unmanned surface vessel agents when the spatiotemporal conflict area is detected, and to calculate and announce a comprehensive change cost for deviating from the current planned path to resolve the conflict; The strategy execution module is used to determine and execute a cooperative avoidance strategy in accordance with a preset negotiation protocol based on the comprehensive change cost declared by all participating unmanned surface vessel agents.

10. The unmanned surface vessel swarm distributed cooperative path planning system according to claim 9, characterized in that, The negotiation decision-making module further includes a fuzzy logic control unit, which is used to adaptively adjust the weight coefficients of various costs used in calculating the comprehensive change cost based on the current real-time environmental information and the unmanned surface vessel's own state information.

Images