Methods and systems for ship navigation assistance in typical waterways based on UAV escort

Abstract

This invention discloses a method and system for assisted navigation of ships in typical waterways based on UAV escort, belonging to the field of intelligent shipping auxiliary control technology. It constructs a multi-dimensional risk perception model for the waterway area, extracts historical accident, hydrological, and obstacle information to generate a risk density matrix; forms high-risk clusters based on risk spectrum dynamic clustering and assigns adjustable intervention weights; schedules UAVs with self-localization and differential recognition capabilities for escort flight, and acquires track deviation, velocity vector, and attitude information through multi-modal perception and low-latency links; calculates intervention coefficients using a risk decision engine and generates heading, speed, and turning point control parameters; and finally achieves local path dynamic optimization and feedback closed-loop control. This invention improves the heading accuracy and hazard avoidance capabilities of ships in complex typical waterways and is suitable for intelligent shipping applications in high-density navigation environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent shipping auxiliary control technology, specifically to a method and system for ship navigation assistance in typical waterways based on unmanned aerial vehicle (UAV) escort. Background Technology

[0002] Currently, ships navigating in typical waterways face complex and ever-changing navigation environments, such as narrow channels, turning areas, bridge areas, intersecting navigation areas, and shoals. Strong current disturbances, high ship density, and limited visibility often lead to accidents such as collisions, groundings, and misnavigation. Traditional auxiliary navigation systems mostly rely on shore-based radar, AIS systems, or shipborne sensors. However, these systems often experience response delays under environmental obstructions or sudden hydrological and meteorological conditions, making it impossible to dynamically identify and intervene in specific high-risk areas.

[0003] While existing technologies have attempted to use drones for maritime surveillance and monitoring of ship arrivals and departures, these efforts remain limited to information collection and lack real-time integration with ship control systems. In particular, they cannot adjust ship trajectories in real time based on the spatiotemporal changes in navigation risk distribution. Furthermore, existing systems often employ fixed-parameter path planning algorithms, neglecting the "dynamic aggregation" of risk factors in typical waterways, such as temporary increases in underwater obstacles, sudden changes in current velocity, and shifts in obstructions.

[0004] Therefore, there is an urgent need to provide an auxiliary navigation method that integrates multi-source risk perception, dynamic UAV escort, intelligent classification of waterway areas, and path control. This method can accurately identify risk areas in typical waterways, model risk density, and calculate intervention priorities. Furthermore, it can achieve more proactive and efficient intelligent navigation control by using aerial UAVs to provide bidirectional coupling feedback to the ship's navigation system. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for assisting ship navigation in typical waterways based on unmanned aerial vehicle (UAV) escort, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for assisted navigation of ships in typical waterways based on unmanned aerial vehicle (UAV) escort, comprising:

[0007] S100. Construct a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution map and vessel type feature set in typical waterways, to generate an initial risk density matrix H.

[0008] S200. Based on the initial risk density matrix H, the waterway is dynamically partitioned to identify high-risk dynamic clusters R with time-varying characteristics, and each region is assigned a specific risk level. Assign adjustable navigation intervention priority weights ;

[0009] S300: At least one UAV with self-localization and differential recognition capabilities is dispatched to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time.

[0010] S400, by drones Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision engine. The data is transmitted back to the ship's navigation module in real time to adjust its control commands.

[0011] S500, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power.

[0012] S600, during the passage of ships After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

[0013] Preferably, the construction of a multi-dimensional risk perception model for the waterway area includes:

[0014] An accident triggering map is constructed based on historical accident data, ship trajectory records, and traffic types of typical waterways. Each node represents a waterway segment, and the edge weights are composed of the accident frequency and ship type weights within a time window, generating structured graph data with time-series labels.

[0015] The graph structure is spliced ​​with the multi-source hydrological change parameters of the waterway and the spatiotemporal distribution map of obstacles using tensor-level processing, and a local attention mechanism is introduced to identify risk-related paths in high-accident-probability areas in order to form a preliminary regional risk thermal index matrix.

[0016] The risk heat index matrix is ​​embedded and learned by a multi-channel convolutional graph neural network, and an initial risk density matrix H is output, which integrates time series and spatial correlation weights, to characterize the risk distribution probability of each typical flight segment within the prediction period.

[0017] Preferably, dynamic partitioning of waterways based on the risk density matrix H includes:

[0018] A sliding window time-domain spectral analysis was performed on the risk density matrix H to extract the main fluctuation factor of the risk spectrum for each flight segment, and the time-varying sensitivity index was calculated based on the rate of change of risk amplitude. It is used to identify potentially high-risk segments with frequency-varying characteristics;

[0019] A multi-objective density clustering algorithm based on fuzzy entropy weights is used to divide the frequently changing high-sensitivity region into a dynamic clustering region R, where the cluster centers are automatically updated according to the entropy mutation points of the risk spectrum;

[0020] For each cluster area Based on the frequency response intensity of the covered flight segment and the adjacent risk diffusion coefficient, a dynamically adjustable intervention priority weight is generated. This serves as an input parameter for optimizing subsequent intervention strategies.

[0021] Preferably, scheduling at least one UAV with self-localization and differential recognition capabilities for escort flight includes:

[0022] Self-localization of accompanying UAVs is achieved using a fusion-based heterogeneous sensing mechanism, including joint computation of visual SLAM and multi-source differential GNSS, to generate a continuous set of localization vectors. ;

[0023] The target ship's multi-dimensional trajectory characteristics, including heading change rate, rudder angle response, and drift speed, are used to perform real-time differential matching with the UAV flight trajectory to construct the trajectory deviation δ and calibrate the flight path in real time.

[0024] A low-latency wireless communication link based on a coded feedforward compression algorithm is established between the UAV and the target ship to synchronously transmit the trajectory deviation δ, velocity vector V, and pitch-roll-yaw three-axis attitude information θ.

[0025] Preferably, the process of collecting multimodal data and calculating intervention coefficients by drones in the dynamic aggregation area R includes:

[0026] A drone was deployed over the flight path, and image frames, wind disturbance fields, and radar reflection sequences were jointly acquired using infrared vision, millimeter-wave radar, and micro-meteorological sensors to construct a time-series perception dataset. ;

[0027] Dataset The input is fed into a multimodal cross-attention module, and the correlation feature channels between heterogeneous signals are extracted based on a risk label-guided learning mechanism to achieve a high-dimensional risk state vector through graph-flow-radar ternary fusion. ;

[0028] By the integrated risk decision engine Perform real-time reasoning and combine current ship behavior response parameters with regional risk levels. Output the current intervention coefficient The data is transmitted back to the ship's navigation and control system for dynamic adjustment of control commands τ, including course fine-tuning, speed limiting, and obstacle avoidance switching strategies.

[0029] Preferably, based on the intervention coefficient Intervention priority weight The coupling relationship between them, determining control parameters and optimizing the path includes:

[0030] Constructing a weighted system based on regional risk impact With real-time intervention coefficient The coupled dynamic function M(ψ, a, P) is used as a control factor to guide the generation of the parameter solution space of the heading correction angle Δψ, the deceleration threshold amin, and the optimal turning point position Popt.

[0031] The parameter solution space is input into a nonlinear multivariable function regression model for state fitting, and combined with the ship's current dynamic state and inertial response, a set of local correction control commands is output. ;

[0032] Through a dual-channel joint adjustment mechanism of rudder control and propulsion, control commands are transmitted... It is synchronously loaded into the ship's navigation system to achieve continuous optimization of the local path for the current voyage, specifically including smoothing of the course trajectory, limiting control of power output, and minimum turning radius constraint.

[0033] Preferably, during the passage of ships through dynamic clustering areas Subsequently, the drone uploads the perception data and updates the risk model, including:

[0034] The images, wind flow, radar data, and corresponding control response information collected during the escort process are combined into a high-dimensional sensing response sequence. By drone The channel intelligent update module uploads data from the region to the shore-based control center.

[0035] Using the distributed entropy difference calculation method to The perceived data is compared with the risk feature vector of the corresponding region in the original risk density matrix H to determine the risk model error and trigger the local map reconstruction mechanism.

[0036] Based on the reinforcement learning policy evolution module, and using historical policy feedback returns as the evaluation metric, the parameter set of the intervention policy for the next cycle is evaluated. Perform weight fine-tuning and path updates to generate an optimized set of navigation strategies.

[0037] This invention also provides a typical navigational aid system for ships based on unmanned aerial vehicle (UAV) escort, comprising:

[0038] The matrix construction module constructs a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution maps, and a set of vessel type characteristics in typical waterways, to generate an initial risk density matrix H.

[0039] The risk area identification module dynamically partitions the waterway based on the initial risk density matrix H, identifies high-risk dynamic clusters R with time-varying characteristics, and assigns a risk area classification to each region. Assign adjustable navigation intervention priority weights ;

[0040] The data exchange module schedules at least one UAV with self-localization and differential recognition capabilities to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time.

[0041] Control and adjustment module, by drone Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision engine. The data is transmitted back to the ship's navigation module in real time to adjust its control commands.

[0042] Path optimization module, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power.

[0043] Strategy correction module, during ship passage After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0045] 1. This invention constructs a multi-dimensional risk perception model for waterway areas and combines it with a dynamic partitioning mechanism driven by spectrum and fuzzy entropy to achieve accurate identification and real-time modeling of high-risk clusters in typical waterways. Based on this, the proposed UAV escort control system possesses heterogeneous perception, self-localization, and differential recognition capabilities. It can form a low-latency data link with the vessel and acquire track deviation, velocity vector, and attitude information in real time, effectively improving the foresight of risk response and the accuracy of control intervention. Through a coupled optimization mechanism of intervention coefficients and risk weights, multi-variable dynamic adjustments to heading, deceleration, and turning points are achieved. Supplemented by a rudder-propulsion dual-channel linkage strategy, this enables the vessel to maintain higher path stability and maneuverability in complex environments.

[0046] 2. Compared with existing assisted navigation systems based on static rules or single-modal perception, this invention not only possesses full-link closed-loop control capabilities but also introduces a reinforcement learning strategy evolution module, supporting cross-cycle updates and adaptive optimization of risk models and control strategies. Experimental results show that this invention outperforms traditional methods in key indicators such as course correction accuracy, hazard avoidance success rate, response delay, and path deviation, significantly improving the intelligent navigation capabilities and overall passage efficiency of ships in typical waterways. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0048] Figure 1 This is a mind map of the method of the present invention.

[0049] Figure 2 This is a mind map of the system modules of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0051] Example 1, please refer to Figure 1 As shown in this embodiment, the ship navigation assistance method based on UAV escort in typical waterways includes:

[0052] S100. Construct a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution map and vessel type feature set in typical waterways, to generate an initial risk density matrix H.

[0053] S200. Based on the initial risk density matrix H, the waterway is dynamically partitioned to identify high-risk dynamic clusters R with time-varying characteristics, and each region is assigned a specific risk level. Assign adjustable navigation intervention priority weights ;

[0054] S300: At least one UAV with self-localization and differential recognition capabilities is dispatched to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time.

[0055] S400, by drones Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision engine. The data is transmitted back to the ship's navigation module in real time to adjust its control commands.

[0056] S500, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power.

[0057] S600, during the passage of ships After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

[0058] In this invention, to achieve intelligent assisted navigation control for ships in typical waterways, the primary task is to construct a dynamically updatable and quantitatively analyzable multidimensional risk perception model for waterway areas. This model serves as the foundational data source for subsequent risk identification, area zoning, UAV intervention path generation, and control strategy optimization. The model described in this embodiment integrates structured historical data, spatiotemporal perception information, and deep learning network inference mechanisms, aiming to output a matrix structure representing the risk density of different sections of a typical waterway, namely, the risk density matrix H.

[0059] First, based on historical accident datasets, vessel trajectory records, and vessel type traffic statistics in typical waterways, a segment-level accident triggering map is constructed. In this map, each node represents a standardized waterway segment unit, and the nodes are connected by edges to form a complete waterway map structure.

[0060] The weight of an edge is determined by two main factors:

[0061] Accident frequency factor: Counts the number of accidents that occur between each pair of adjacent flight segments within a specified time window (e.g., the past 12 months);

[0062] Ship type weighting coefficient: Taking into account the proportion of different ship types passing through this section, weights are set according to tonnage and structural complexity, so that the side weight of heavy-load or poorly maneuverable ship types that cause accidents is higher.

[0063] The final structure is a type of directed weighted graph with temporal attributes, called the accident triggering graph G = (V,E, T), where V represents the set of nodes (i.e., flight segments), E represents the set of edges (with weights), and T represents the additional time series label, reflecting the state changes of the graph in different time periods.

[0064] To improve the model's adaptability to hydrological and environmental changes, it is necessary to fuse the aforementioned maps with multi-source environmental data. This embodiment introduces a tensor stitching method to combine the following three types of data with the graph structure:

[0065] Multi-source hydrological parameters of the waterway include water depth, flow velocity, flow direction, and turbidity.

[0066] Spatiotemporal distribution map of obstacles: generated by linking high-resolution remote sensing images with the AIS automatic recognition system, recording both variable and fixed obstacles;

[0067] High-frequency meteorological disturbance factors: such as instantaneous wind speed, wind direction changes, visibility levels, etc.

[0068] The aforementioned data is uniformly embedded into a tensor space corresponding to each graph node, forming a multi-channel tensor structure. To extract high-risk relevance paths, a local attention mechanism is introduced. An attention score function is trained on each frame of the graph structure, and the weights are dynamically adjusted based on the differences in risk impact factors to identify potential "risk path segments" with strong accident inducing factors.

[0069] The above processing results are encoded to generate a preliminary regional risk thermal index matrix. Each row of the matrix corresponds to a flight segment node, and each column corresponds to a perception factor dimension. The matrix values ​​represent the aggregate score of risk indicators.

[0070] The above index matrix The input is fed into a pre-designed multi-channel convolutional graph neural network (MCGNN) for structured learning. This network has the following key characteristics:

[0071] Supports multiple input channels: simultaneously processing graph topology (from G) and tensor quantization environment metrics (from G). );

[0072] Fusion of time series: The state of each node forms a time sliding window in the first T time steps to extract temporal fluctuation features;

[0073] Convolutional diffusion: Performs feature diffusion operations on the graph structure to achieve spatial upstream and downstream transmission of risk indicators and expansion of the perception region.

[0074] The output of MCGNN is the predicted risk density for each node at the current moment, with a value ranging from [0, 1], representing the probability that the flight segment will trigger an abnormal navigation state in the next cycle (e.g., within 10 minutes). The outputs of all flight segments constitute the final risk density matrix H∈ , where n is the total number of modeled segments in the waterway, and R is the set of real numbers.

[0075] By constructing the aforementioned multidimensional risk perception model, this invention achieves a quantifiable, time-varying, and controllable representation of risk distribution in waterway areas. Compared with traditional rule-based or statistical regression-based assessment methods, this model has the following significant advantages:

[0076] High-dimensional fusion: Integrates multi-source, heterogeneous, and time-series data, possessing stronger environmental adaptability; Regional correlation capture: Utilizes graph neural network mechanisms to uncover hidden risk propagation paths between navigation segments; Dynamic evolution: Risk heatmaps can evolve over time, continuously correcting perception errors and prediction biases; Data-driven self-learning capability: The network structure can be continuously trained and updated, suitable for generalization and promotion in complex waters.

[0077] In summary, the construction of a multi-dimensional risk perception model for the waterway area lays the foundational capability support for the entire set of assisted navigation methods based on UAV escort. Subsequent dynamic zoning, strategy optimization, path control, and feedback optimization can all be based on this model to achieve improved accuracy and stable output.

[0078] This method differs from traditional static segmentation based on waterway geometry or administrative divisions. Instead, it incorporates temporal spectrum analysis and fuzzy density clustering mechanisms to construct a dynamic, multi-objective-driven system for classifying high-risk waterway areas. This system possesses the ability to identify the dynamic evolution trends and coupling diffusion paths of regional risks, providing a strategic basis for subsequent UAV intervention and ship navigation control.

[0079] First, the risk density matrix H ∈ output during the construction phase. (Where n is the number of flight segments) Perform a sliding window-based time-domain spectral analysis. Let the sliding window size be T and the window step size be Δt. At each time step, perform a Fast Fourier Transform (FFT) on the risk value sequence of each flight segment to extract the main fluctuation factor in the frequency domain, i.e., the frequency... The corresponding energy peak.

[0080] Next, the gradient of the risk amplitude over time is calculated, i.e.: ;in Let represent the risk amplitude of the i-th flight segment at time t, and α represent its rate of change of amplitude. Based on this, a time-varying sensitivity index is introduced: Where std represents the standard deviation and var represents the variance of frequency variation. This is the weighting factor. The indicators are used to identify potentially highly sensitive areas characterized by high-frequency fluctuations and unstable risk distribution. These areas often occur at points of sudden changes in ship currents, confluences, or frequent environmental disturbances, and are candidate areas for the formation of dynamic high-risk clusters.

[0081] Based on the aforementioned high-sensitivity region sequences, an improved Fuzzy-Entropy Weighted Multi-objective Density Clustering (FEW-MDC) algorithm is employed to achieve spatiotemporal partitioning of the flight segment set. This clustering algorithm mainly includes the following core elements:

[0082] Fuzzy entropy weighting function: Calculate the entropy value of the risk spectrum for each flight segment. The entropy mutation point is used to dynamically define the cluster center, so that the boundary of the cluster region R adapts to the change over time.

[0083] Density objective function: maximize intra-cluster risk similarity, minimize inter-cluster boundary diffusion, and support joint clustering of multi-dimensional attributes of flight segments (including risk magnitude, frequency change rate, spatial adjacency, etc.).

[0084] Evolutionary path modeling mechanism: Record the drift trajectory of each cluster center over time to construct the evolutionary path map of the "risk cluster" for predicting future regional expansion trends.

[0085] The clustering output results in multiple dynamic risk clusters. Each This represents a set of flight segments with strong correlation in risk density and similar disturbance characteristics within a certain time period. The number, shape, and distribution of clusters change dynamically with the input matrix H and window parameters.

[0086] After identifying dynamic clustering areas Subsequently, intervention priority weights with adjustability and strategic value need to be generated for it. This serves as the basis for downstream UAV trajectory allocation and path intervention parameters.

[0087] Weight The generation is mainly based on two dimensions:

[0088] Frequency response strength This refers to the weighted average of the frequency energy peak values ​​of each flight segment within the cluster area, reflecting the degree of fluctuation risk in that area.

[0089] Adjacent risk diffusion coefficient : Measures the trend of risk in the current cluster area spreading to adjacent areas. It is calculated as the average diffusion gradient between the edge node of the current area and the adjacent low-risk flight segment.

[0090] Finally, the following mapping function is used to generate... : ;in These are adjustable policy coefficients used to switch between security-first and efficiency-first policies. The generated weights... In the main control module, the intervention coefficient will be used. Coupling is used to generate control parameters such as heading correction angle Δψ, deceleration threshold amin, and turning point position Popt.

[0091] This dynamic partitioning method has the following significant technical effects and advantages:

[0092] High time-varying dynamism: Through the sliding spectrum and cluster center evolution mechanism, it can accurately identify clustered areas with rapidly changing risks and drifting locations; High identification granularity: It supports sub-segment level zoning and is suitable for local risk analysis such as narrow waterways and tributary channels; Strong strategy-driven: The output intervention weights are continuously adjustable and support the coordinated execution of multiple control strategies; Strong evolutionary modeling capability: Through the risk cluster evolution path prediction mechanism, it has the ability to provide early warning of potential high-risk clustering trends.

[0093] In summary, this dynamic partitioning scheme not only differs structurally from the traditional fixed segmentation strategy, but also adopts a frequency domain-clustering-graph evolution fusion approach in its technical path. This significantly enhances the risk modeling accuracy and response capability of the system in complex waterway environments, providing high-quality data support for subsequent UAV escort control systems.

[0094] To achieve precise and efficient escort and auxiliary control of ships in typical waterways, this invention further designs an unmanned aerial vehicle (UAV) escort control system that integrates multi-source sensing, differential recognition, and low-latency communication capabilities. The core of this system is to schedule at least one UAV with high-precision self-localization and path differential recognition capabilities to establish a closed-loop information feedback mechanism with the target ship. This enables synchronous sensing and sharing of trajectory deviation δ, velocity vector V, and three-axis attitude angle θ, supporting the generation of subsequent control intervention decisions.

[0095] Traditional drones rely heavily on a single GPS signal for positioning, which is easily affected by factors such as building obstruction, water reflection, and weather disturbances in the flight path environment, leading to position errors and attitude drift. To address these issues, this invention proposes a fusion-based heterogeneous sensing mechanism to enhance the self-localization capability of drones, specifically including:

[0096] Visual SLAM system: A simultaneous localization and mapping (SLAM) system is deployed on the UAV's onboard computing platform to acquire feature points around the ship through continuous image frames, thereby achieving relative environmental positioning.

[0097] Multi-source differential GNSS solution: A dual-frequency differential GPS, BeiDou and GLONASS joint solution module is introduced, and correction data from shore-based reference stations are received simultaneously to form a real-time differential positioning model, which effectively suppresses satellite signal drift;

[0098] Inertial navigation system assistance: Using a miniature IMU module to collect linear acceleration and angular velocity, it provides short-term inertial attitude estimation when GNSS temporarily loses lock.

[0099] After the three are fused, a fused positioning solution unit is constructed using a Kalman filter to output a set of accurate three-dimensional positioning vectors for the UAV in continuous time: ; This represents the X coordinate (usually the spatial position in the east-west direction) of the UAV in the airway reference coordinate system at the m-th time point. This represents the Y-coordinate (usually the spatial position in the north-south direction) corresponding to the m-th time point. This represents the Z-coordinate (the altitude or elevation of the drone) corresponding to the m-th time point. The timestamp corresponding to this set of spatial coordinates represents the time of the UAV's m-th sampling, in seconds or milliseconds; M represents the total number of positioning data points recorded throughout the entire escort cycle. This positioning set provides a high-precision data source for subsequent dynamic calibration of the flight path and target tracking.

[0100] To ensure that UAVs can stably track and observe the navigation status of target ships, this invention proposes a trajectory differential matching algorithm based on multi-dimensional behavioral features. This algorithm constructs a trajectory deviation index δ by comparing the current navigation characteristics of the target ship with the dynamic offset of the UAV trajectory, which is used to correct the flight path in real time.

[0101] The main steps include:

[0102] Target behavior feature extraction: Extracting heading rate of change (dψ / dt), rudder angle response (γ), and drift speed from ship AIS data and inertial navigation system. Key dynamic features, such as these, constitute the target state vector Sship;

[0103] Flight trajectory comparison and analysis: combined with the current positioning group of the UAV Calculate its relative position error, velocity vector angle, and heading offset in the ship's reference frame;

[0104] Track deviation calculation: Construct a nonlinear difference model, compare the ship's behavioral response with the UAV's relative motion trajectory, and generate the track deviation. ;in For feature weight parameters, These are the velocity vectors of the UAV and the ship, respectively, and ψ is the heading angle. Standard positioning group for drones.

[0105] The generated deviation index The control module can input data in real time, enabling the UAV to maintain a stable position and adjust its viewing angle in front of and behind the ship, thus ensuring the quality of the sensing data and the stability of the observation.

[0106] To enable data interaction between UAVs and ships, this invention constructs a wireless communication link based on a coded feedforward compression algorithm. Its core objective is to improve the real-time performance and stability of trajectory and attitude data. This communication mechanism includes:

[0107] Compression coding mechanism: Feedforward prediction compression technology is used for three-dimensional attitude information and velocity vector V to transmit only key transition points and prediction errors, thereby reducing the amount of redundant data;

[0108] Frequency domain modulation optimization: Use low-power, high-frequency narrowband communication protocols (such as LoRa or custom WiFi-AC channels) to transmit encoded δ, V, and θ key data packets;

[0109] Feedback confirmation mechanism: A link layer confirmation mechanism is designed to ensure stable data packet transmission and allow for rapid retransmission, maintaining low latency and high stability in complex maritime environments.

[0110] Once this link is established, corresponding receiving and decoding modules are deployed on the ship navigation system side, which can use the received data in real time for control strategy calculation and navigation adjustment, effectively forming a data closed loop of "air-ship collaboration".

[0111] Through the above multi-module collaborative design, the UAV escort system of this invention has the following outstanding technical advantages in complex waterway environments: high positioning accuracy and strong adaptability: the fusion perception mechanism significantly enhances the UAV's positioning capability under conditions of obstruction and electromagnetic interference; accurate differential recognition and fast control response: multi-dimensional track feature comparison improves the UAV's tracking stability of ship motion state; high communication efficiency and strong link stability: based on feedforward compression and feedback mechanisms, it ensures real-time sharing of key control data and meets the reliability requirements under dynamic navigation conditions.

[0112] In summary, the UAV scheduling and control technology described in this embodiment not only realizes "track-level coordinated observation" but also constructs a closed-loop information feedback control framework, significantly improving the UAV's active perception capability and track following accuracy in intelligent assisted navigation systems.

[0113] To achieve real-time perception and intelligent intervention control of dynamic high-risk flight segments, this invention, based on the aforementioned identification of dynamic clustering areas (R), further deploys unmanned aerial vehicle (UAV) units with heterogeneous perception and decision-making capabilities to perform multimodal environmental data acquisition and intervention coefficient calculation tasks. This process not only expands the ship's perception dimensions of the surrounding high-risk environment but also enables closed-loop parameter feedback from the aerial unit to the ship's control system, thereby constructing a ship-air collaborative risk response mechanism.

[0114] High-risk clusters in typical waterways In this invention, at least one escort drone is deployed in the airspace above the target vessel to collect multimodal environmental information based on its onboard sensor array, specifically including:

[0115] Infrared vision subsystem: It acquires image frame sequences through infrared cameras to identify events such as heat sources, floating objects on the water surface, and navigation mark obstruction under low visibility conditions;

[0116] Millimeter-wave radar subsystem: Scans multiple layers of obstacles at near, medium, and far distances in a dynamic clustering area, forms a radar reflection intensity sequence, and identifies dynamic objects, changes in the relative positions of ships, and clustering patterns of suspicious obstacles;

[0117] Micro-meteorological sensing subsystem: Real-time measurement of local wind speed, wind direction disturbance gradient, humidity changes, etc., used to identify disturbances caused by wind shear, high-speed air masses, and sudden weather anomalies.

[0118] The aforementioned sensing results were collected using a unified timestamp synchronization mechanism and stored as a multimodal sensing dataset in a continuous time domain. The data structure is as follows: ;in Infrared image frame, For millimeter-wave radar reflection data, This represents the vector of micro-meteorological disturbance.

[0119] To extract feature channels with practical intervention value from multimodal data, this invention designs a cross-attention-driven risk state modeling mechanism. Specifically, it includes:

[0120] Guide tag generation: generated by the shore-based control center based on historical flight segment accident records and the current area. Weights are used to generate guided risk labels for supervised learning during training.

[0121] Multimodal cross-attention module design: Implement cross-attention mechanism for different modal feature subspaces, enhance contextual features between adjacent frames and modalities, and construct a strongly coupled attention channel between heterogeneous signals;

[0122] Feature vector construction: Feature vectors such as image edge changes, radar scatter density, and wind flow mutation index are fused into a set of time-continuous high-dimensional risk state vectors. Its dimensions and structure remain consistent with the input of the intervention decision model.

[0123] The cross-attention mechanism effectively solves problems such as redundant interference between modal data and differences in perceptual response during the learning process, significantly improving the risk state vector. Its expressive power and predictive accuracy.

[0124] To achieve coordinated response with the ship's control system, this invention introduces an integrated risk decision engine, based on the aforementioned risk state vector. By combining ship behavior response parameters, the current intervention coefficient f can be adjusted. k Real-time calculation:

[0125] Input parameter fusion: Engine input includes the current risk state vector. Regional risk level And the target vessel's heading angle change rate, speed stability index and control history;

[0126] Inference network architecture: A discriminative network combining a lightweight multilayer perceptron and a gating mechanism is used to perform feature cross-validation and pattern recognition on the input data;

[0127] Output intervention coefficient : Output ∈ [0, 1] represents the current regional intervention intensity recommendation value, which is used as the input to the control module.

[0128] This intervention coefficient will be fed back to the ship's navigation control system, and the following control optimization behavior will be achieved by dynamically adjusting the control command τ:

[0129] Heading fine-tuning Δψ: Adjust the current rudder angle according to the direction of yaw risk;

[0130] Speed ​​limit amin: Limits the current maximum propulsion speed to prevent acceleration through high-risk areas;

[0131] Obstacle avoidance mode switching: Based on the obstacle density or sudden object recognition results, activate active obstacle avoidance strategy or lane change control mechanism.

[0132] This implementation method deploys unmanned aerial vehicle (UAV) units with multimodal perception and decision-making capabilities to achieve accurate identification of the local environment and risk status modeling within the dynamic aggregation area R. This significantly improves the response speed and processing accuracy of the ship control system to sudden risks. Specifically, this is manifested in: expanded perception dimensions: surpassing traditional radar and AIS positioning, it adds comprehensive perception capabilities of infrared, radar, and meteorological signals; strong real-time intervention and decision-making: the link delay from risk status to intervention coefficient to control feedback is less than 200ms, enabling continuous dynamic heading control; and tight coupling with the ship system: the output results are directly connected to the ship's main control system, possessing engineering feasibility and reliability.

[0133] To effectively address the volatile situation in high-risk areas of typical waterways and achieve dynamic adjustment and continuous control optimization of vessel paths, this invention obtains intervention coefficients. Regional intervention priority weight Building upon this foundation, a further intelligent control parameter reasoning and path optimization system based on coupling relationships is constructed. This system, through mathematical modeling and control feedback mechanisms, jointly considers the ship's current motion state, risk gradient distribution, and segment inertia effects to dynamically solve and load the heading correction angle Δψ, deceleration threshold amin, and optimal turning point position Popt.

[0134] Before optimizing control, intervention coefficients need to be established. Intervention priority weight A mathematical coupling model is used to quantify the coupling strength between the ship's risk environment and the desired control response. This invention designs a regional coupling dynamic optimization function M(ψ, a, P), as follows: ;in: Weighting of the current cluster area's risk impact; The adjustment function for the intervention coefficient is, for example... s is the response sensitivity factor; ψ represents the risk gradient sensitivity index for the current flight segment, i.e., the local gradient value of the risk density H in its spatial distribution, indicating an increasing risk trend; ψ represents the candidate heading angle, a represents the deceleration magnitude, and P represents the set of possible turning points. This coupling function is used to generate a feasible solution space for a set of multidimensional control parameters. Each solution corresponds to a set of candidate adjustment strategies for ships to cope with the risks of the current voyage segment.

[0135] To address the aforementioned solution space for control parameters, this invention further designs a nonlinear multivariable function regression model to screen for the optimal solution in the control strategy solution space and transform it into executable control instructions. The model primarily considers the following factors:

[0136] Current power parameters of the ship, including propulsion power rudder speed , rudder angle ;

[0137] The ship's inertial response model includes the center of mass shift, moment of inertia J, and current disturbance coefficient.

[0138] The wind disturbance prediction input is used to dynamically adjust the weights of the objective function in the regression model.

[0139] The regression objective function is defined as: ;in: This represents the deviation between the desired path and the current path. To control the response hysteresis loss caused by changes in control parameters to the ship's inertial system; To control the estimated latency between loading and execution;

[0140] The regression model outputs a set of robust control commands τ′ = {Δψ*, amin*, Popt*}, which are adapted to the optimal path correction objective under the current complex environment.

[0141] To ensure the effective execution of the τ′ command in the ship system, this invention employs a dual-channel joint adjustment mechanism of rudder control and propulsion, specifically including the following execution methods:

[0142] Heading trajectory smoothing: The differential control controller performs smooth interpolation on Δψ* to make the heading change process continuous and stable, avoiding oscillations caused by sudden changes;

[0143] Power output limiting control: Set the upper limit of the thruster power based on the deceleration threshold amin*, and make adaptive adjustments considering the current ship speed and forward distance;

[0144] Minimum turning radius constraint mechanism: Evaluate whether the turning radius near the selected turning point Popt* is less than the ship's safe minimum value rmin. If it is less than rmin, execute the turning point offset compensation strategy.

[0145] The execution process is periodically invoked by the ship's main navigation system controller, in conjunction with data transmitted back from the UAV. The frequency is dynamically adjusted to maintain a consistent level (e.g., once every 0.5 seconds) to ensure that local path optimization remains continuous and real-time.

[0146] This implementation constructs a strongly coupled mathematical model of risk parameters and control variables. Through adjustable function solution space mapping and nonlinear regression strategies, it effectively solves the following problems: unstable control response: traditional path planning methods lag in response to sudden risks, while this solution improves predictability through weight-intervention coupling; incoordination in regulation: linking rudder control and propulsion control logic prevents command conflicts between path correction and speed control; local optimum trap problem: through multi-solution space modeling and dynamic regression path selection mechanism, it effectively avoids local suboptimal solutions.

[0147] To ensure the risk density matrix H has adaptive updating capabilities during ship navigation and to support the continuous evolution and optimized control of intervention strategies, this invention designs a risk model update mechanism centered on perception response data-driven analysis, entropy difference judgment, and iterative strategy reconstruction. This mechanism relies on the perception and control feedback data from the accompanying UAV within the dynamic aggregation zone R, and promptly uploads this data to the shore-based system for intelligent map reconstruction and strategy evolution updates after the ship passes through this area.

[0148] Ships pass through dynamic gathering area During this period, the accompanying UAV has collected multiple consecutive image frame sequences, wind and air disturbance data, radar reflection maps, and corresponding control response information, including heading adjustment angles, power output parameters, and intervention execution results.

[0149] The aforementioned multimodal data, after being synchronized with timestamps, are combined to form a unified high-dimensional sensing response sequence. Its structure is as follows: ;in: For image frame sequences (infrared or visible light); This is millimeter-wave radar reflection data; This represents the wind disturbance vector (wind speed, wind direction). To execute control commands; This corresponds to the intervention coefficient. When the ship sails out... After reaching the area, the drone automatically transmits the above information via an encrypted communication module. The data is uploaded to the intelligent channel update module of the shore-based control center for subsequent model update operations.

[0150] To determine the accuracy and timeliness of the model for this region in the original risk density matrix H, this invention introduces a distributed entropy difference calculation method to compare the differences between the newly uploaded data and the current risk feature vector. The specific steps are as follows:

[0151] Risk feature extraction: from Key indicators such as image frequency domain features, radar point cloud density, and disturbance peak frequency are extracted to construct a newly generated regional risk vector. ;

[0152] Original feature retrieval: Extracting corresponding features from the risk density matrix H Historical risk vector ;

[0153] Entropy difference calculation: The model bias is defined by comparing the KL divergence and the Shannon entropy difference. If the entropy difference If the threshold is exceeded, it indicates that the current regional risk model significantly deviates from the true state, and the system will automatically trigger a local map reconstruction mechanism. This mechanism reconstructs the graph node status of the region based on the latest data and adjusts the edge weights with adjacent flight segments, updating them as follows: ; where η is the corrected learning rate and ΔF is the vector difference.

[0154] After the map reconstruction is completed, in order to achieve more effective control intervention and path selection in the next cycle, this invention introduces a policy evolution module based on reinforcement learning to process the intervention policy parameter set. Perform weight fine-tuning and path updates.

[0155] Its core logic is:

[0156] Strategy benefit assessment: utilizing The path deviation correction effect, speed stability, and change in the final risk index corresponding to τt are used to construct the strategy scoring function R(τt);

[0157] Reinforcement learning update rules: The update algorithm based on policy gradient method (such as REINFORCE or PPO) is adopted to adjust the intervention parameters according to the policy reward and optimize the policy probability distribution π(τ|s), where s is the current state;

[0158] Strategy set output: The system outputs a set of optimized intervention parameters: ;in This represents the reinforcement learning policy function. This parameter set will be used as the initial control input in the next cycle segment control task, effectively improving the control response's foresight and environmental adaptability.

[0159] Through the risk model adaptive update mechanism described in this embodiment, the present invention achieves significant technical improvements in the following aspects: data-driven closed-loop modeling: real-time sensing response data is used for feedback verification and map reconstruction, breaking the limitations of static risk presets; high error identification accuracy: the risk model deviation is accurately judged through entropy difference criteria, improving local modeling accuracy; continuous evolution of strategy optimization: the reinforcement learning-driven strategy update mechanism ensures that the control strategy is adjusted synchronously with environmental changes; cross-cycle memory mechanism: the sensing response is stored for a long time for subsequent model training, and the system has continuous learning capabilities.

[0160] Example 2, to verify the effectiveness of the "UAV-based ship-assisted navigation method in typical waterways" described in this invention in a real navigation environment, constructs a digital simulation platform based on a typical inland waterway, uses a ship-UAV cooperative control model, and sets up a comparison group for performance evaluation. The simulation platform is built on ROS+Gazebo to simulate a synchronous navigation environment between ships and UAVs, with a system sampling period of 100ms.

[0161] Experimental setup:

[0162] Test segment: A typical curve segment (1.2km long, including 3 high-risk dynamic clustering areas) was selected. );

[0163] Simulated ship type: 3,000-ton container ship, maximum speed 18 km / h;

[0164] Drone configuration: Equipped with an infrared camera, millimeter-wave radar, GNSS module, and micro-weather sensor;

[0165] Experimental group: Deploy the method of this invention to achieve a closed loop of risk perception + control optimization + path adjustment;

[0166] Comparison group: using traditional trajectory tracking algorithm (PID control + fixed path planning);

[0167] Test metric definition:

[0168] Heading correction accuracy (°): The average angular deviation between the actual heading and the target path;

[0169] Control response latency (ms): The average response time from risk identification to the execution of control actions;

[0170] Risk avoidance success rate (%): The percentage of simulated risk events in which collisions or groundings are avoided;

[0171] Average path deviation (m): The average distance of the ship's center trajectory from the target centerline.

[0172] Experimental results:

[0173]

[0174] Technical Effects: As can be seen from the above results, in the three risk areas with strong water flow disturbances and obstacle accumulation, the method of this invention effectively improves the trajectory control accuracy and response speed of ships by constructing a risk density matrix H, dynamically identifying accumulation areas, and generating control parameters based on intervention coupling. Specifically, the significantly improved heading correction accuracy is mainly due to the UAV's coordinated identification of real-time attitude and risk; the reduced response latency is attributed to the... and The coupled predictive control strategy and the improved risk avoidance success rate demonstrate that the invention has stronger operational stability and safety margin in highly dynamic environments; the significant reduction in path deviation verifies the actual effectiveness of the path optimization algorithm.

[0175] Therefore, this embodiment fully demonstrates the beneficial effects of the present invention in intelligent assisted navigation systems, and has significant engineering application value and promotion prospects.

[0176] Example 3, please refer to Figure 2 As shown in this embodiment, the ship navigation assistance system based on UAV escort includes:

[0177] The matrix construction module constructs a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution maps, and a set of vessel type characteristics in typical waterways, to generate an initial risk density matrix H.

[0178] The risk area identification module dynamically partitions the waterway based on the initial risk density matrix H, identifies high-risk dynamic clusters R with time-varying characteristics, and assigns a risk area classification to each region. Assign adjustable navigation intervention priority weights ;

[0179] The data exchange module schedules at least one UAV with self-localization and differential recognition capabilities to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time.

[0180] Control and adjustment module, by drone Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision engine. The data is transmitted back to the ship's navigation module in real time to adjust its control commands.

[0181] Path optimization module, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power.

[0182] Strategy correction module, during ship passage After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

[0183] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assisted navigation of ships in typical waterways based on unmanned aerial vehicle (UAV) escort, characterized in that: include: S100. Construct a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution map and vessel type feature set in typical waterways, to generate an initial risk density matrix H. S200. Based on the initial risk density matrix H, the waterway is dynamically partitioned to identify high-risk dynamic clusters R with time-varying characteristics, and each region is assigned a specific risk level. Assign adjustable navigation intervention priority weights ; S300: At least one UAV with self-localization and differential recognition capabilities is dispatched to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time. S400, by drones Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision-making calculation unit. The data is transmitted back to the ship's navigation module in real time to adjust its control commands. S500, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power. S600, during the passage of ships After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

2. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: The construction of a multi-dimensional risk perception model for the waterway area includes: An accident triggering map is constructed based on historical accident data, ship trajectory records, and traffic types of typical waterways. Each node represents a waterway segment, and the edge weights are composed of the accident frequency and ship type weights within a time window, generating structured graph data with time-series labels. Specifically, this includes: in this map, each node represents a standardized waterway segment unit, and the nodes are connected by edges to form a complete waterway map structure; The weight of an edge is determined by two main factors: Accident frequency factor: Counts the number of accidents that occur between each pair of adjacent flight segments within a specified time window; Ship type weighting factor: taking into account the proportion of different ship types passing through this section of the waterway; The final structure is a type of directed weighted graph with temporal attributes, called the incident triggering graph G = (V, E, T), where V represents the set of nodes, E represents the set of edges, and T represents the additional time series label. The graph structure is spliced ​​with the multi-source hydrological change parameters of the waterway and the spatiotemporal distribution map of obstacles using tensor-level processing, and a local attention mechanism is introduced to identify risk-related paths in high-accident-probability areas in order to form a preliminary regional risk thermal index matrix. The risk heat index matrix is ​​embedded and learned by a multi-channel convolutional graph neural network, and an initial risk density matrix H is output, which integrates time series and spatial correlation weights, to characterize the risk distribution probability of each typical flight segment within the prediction period. The embedding learning supports multiple input channels: that is, it simultaneously processes the accident triggering map and the regional risk heat index matrix.

3. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: Dynamic zoning of waterways based on the risk density matrix H includes: A sliding window time-domain spectral analysis was performed on the risk density matrix H to extract the main fluctuation factor of the risk spectrum for each flight segment, and the time-varying sensitivity index was calculated based on the rate of change of risk amplitude. It is used to identify potentially high-risk segments with frequency-varying characteristics; A multi-objective density clustering algorithm based on fuzzy entropy weights is used to divide the frequently changing high-sensitivity region into a dynamic clustering region R, where the cluster centers are automatically updated according to the entropy mutation points of the risk spectrum; The multi-objective density clustering algorithm specifically includes: a fuzzy entropy weight function, which calculates the entropy value of the risk spectrum for each flight segment, and entropy mutation points are used to dynamically define the cluster center so that the boundary of the clustering region R adapts to changes over time; and a density objective function, which maximizes the risk similarity within the cluster and minimizes the boundary diffusion between clusters, supporting joint clustering of multi-dimensional attributes of flight segments. For each cluster area Based on the frequency response intensity of the covered flight segment and the adjacent risk diffusion coefficient, a dynamically adjustable intervention priority weight is generated. This serves as an input parameter for optimizing subsequent intervention strategies.

4. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: Dispatching at least one drone with self-localization and differential recognition capabilities for escort flight includes: Self-localization of accompanying UAVs is achieved using a fusion-based heterogeneous sensing mechanism, including joint computation of visual SLAM and multi-source differential GNSS, to generate a continuous set of localization vectors. ; The target ship's multi-dimensional trajectory characteristics, including heading change rate, rudder angle response, and drift speed, are used to perform real-time differential matching with the UAV flight trajectory to construct the trajectory deviation δ and calibrate the flight path in real time. A low-latency wireless communication link based on a data compression algorithm with a joint structure of forward prediction and entropy coding is established between the UAV and the target ship to synchronously transmit the trajectory deviation δ, velocity vector V, and pitch-roll-yaw three-axis attitude information θ. The data compression algorithm based on the joint structure of forward prediction and entropy coding includes: using feedforward prediction compression technology for three-dimensional attitude information and velocity vector V, and using low-power, high-frequency narrowband communication protocol to transmit the encoded data packets of δ, V, and θ.

5. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: The process of collecting multimodal data and calculating intervention coefficients in a dynamic aggregation area R using drones includes: A drone was deployed over the flight path, and image frames, wind disturbance fields, and radar reflection sequences were jointly acquired using infrared vision, millimeter-wave radar, and micro-meteorological sensors to construct a time-series perception dataset. ; Dataset The input is fed into a multimodal cross-attention module, and the correlation feature channels between heterogeneous signals are extracted based on a risk label-guided learning mechanism to achieve a high-dimensional risk state vector through graph-flow-radar ternary fusion. ; By the integrated risk decision calculation unit Perform real-time reasoning and combine current ship behavior response parameters with regional risk levels. Output the current intervention coefficient The data is transmitted back to the ship's navigation and control system for dynamic adjustment of control commands τ, including course fine-tuning, speed limiting, and obstacle avoidance switching strategies.

6. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: Based on intervention coefficient Intervention priority weight The coupling relationship between them, determining control parameters and optimizing the path includes: Constructing a weighted system based on regional risk impact With real-time intervention coefficient The control parameter calculation model M(ψ, a, P) uses risk gradient sensitivity as a control factor to guide the generation of the parameter solution space for the heading correction angle Δψ, the deceleration threshold amin, and the optimal turning point position Popt. Specifically, it includes: ;in: Weighting of the current cluster area's risk impact; Let be the adjustment function for the intervention coefficient, and s be the response sensitivity factor; ψ represents the risk gradient sensitivity index for the current flight segment, i.e., the local gradient value of the risk density H in the spatial distribution, indicating the trend of increasing risk; ψ represents the candidate heading angle, a represents the deceleration level, and P represents the set of possible turning points. The parameter solution space is input into a nonlinear multivariable function regression model for state fitting, and combined with the ship's current dynamic state and inertial response, a set of local correction control commands is output. ; The nonlinear multivariate function regression model considers the ship's current dynamic state parameters, the ship's inertial response model, and the predicted input of wind and current disturbances. The regression objective function is defined as: ;in: This represents the deviation between the desired path and the current path. To control the response hysteresis loss caused by changes in control parameters to the ship's inertial system; To control the estimated latency between loading and execution; The regression model outputs a set of robust control commands τ′ = {Δψ*, amin*, Popt*}; Through a dual-channel joint adjustment mechanism of rudder control and propulsion, control commands are transmitted... It is synchronously loaded into the ship's navigation system to achieve continuous optimization of the local path for the current voyage, specifically including smoothing of the course trajectory, limiting control of power output, and minimum turning radius constraint.

7. The method for assisted navigation of ships in typical waterways based on UAV escort according to claim 1, characterized in that: Ships pass through dynamic gathering area Subsequently, the drone uploads the perception data and updates the risk model, including: The images, wind flow, radar data, and corresponding control response information collected during the escort process are combined into a high-dimensional sensing response sequence. By drone The channel intelligent update module uploads data from the region to the shore-based control center. Using the distributed entropy difference calculation method to The perceived data is compared with the risk feature vector of the corresponding region in the original risk density matrix H to determine the risk model error and trigger the local map reconstruction mechanism. Based on the reinforcement learning policy evolution module, and using historical policy feedback returns as the evaluation metric, the parameter set of the intervention policy for the next cycle is evaluated. Perform weight fine-tuning and path updates to generate an optimized set of navigation strategies.

8. A ship navigation assistance system based on UAV escort, used to implement the ship navigation assistance method based on UAV escort as described in any one of claims 1-7, characterized in that: include: The matrix construction module constructs a multi-dimensional risk perception model for the waterway area, including historical accident data, hydrological change parameters, obstacle distribution maps, and a set of vessel type characteristics in typical waterways, to generate an initial risk density matrix H. The risk area identification module dynamically partitions the waterway based on the initial risk density matrix H, identifies high-risk dynamic clusters R with time-varying characteristics, and assigns a risk area classification to each region. Assign adjustable navigation intervention priority weights ; The data exchange module schedules at least one UAV with self-localization and differential recognition capabilities to accompany the target ship and establish a low-latency wireless communication link to exchange track deviation δ, velocity vector V and attitude information θ in real time. Control and adjustment module, by drone Data including images, wind flow, and radar reflections are collected within the region, and the current intervention coefficient is calculated using an integrated risk decision-making calculation unit. The data is transmitted back to the ship's navigation module in real time to adjust its control commands. Path optimization module, based on and The coupling relationship between them determines the heading correction angle Δψ, deceleration threshold amin and optimal turning point position Popt for the current flight segment, and achieves local path optimization through joint adjustment of rudder control and power. Strategy correction module, during ship passage After the area is identified, the drone uploads its sensing data to the shore-based control center to update the corresponding risk distribution map in the waterway area and make corrections. And parameters for the next cycle of intervention strategies.

Images