CN-121994208-A - Intelligent shipping collision risk prediction and avoidance method based on large model reasoning

CN121994208ACN 121994208 ACN121994208 ACN 121994208ACN-121994208-A

Abstract

The invention discloses an intelligent shipping collision risk prediction and avoidance method based on large model reasoning, which relates to the technical field of ship navigation and comprises the steps of acquiring multi-mode data, carrying out weighted fusion, alignment and feature enhancement on image sequences of marine radar points, automatic identification system tracks and optical cameras in the multi-mode data to obtain enhanced features, and acquiring future moments The method comprises the steps of obtaining additional cost of underwater obstacle based on sonar echo, generating a candidate track set based on enhanced features, obstacle occupation, obstacle speed and additional cost, calculating comprehensive scores based on the candidate track set, selecting high-score tracks, obtaining nominal control according to the high-score tracks, and correcting and calibrating the nominal control to obtain final control. The method ensures the integrity and consistency of environmental characterization by self-adaptive alignment and enhanced fusion of the multi-mode data of vision, radar, AIS and sonar, and uniformly models the risks of underwater and water surfaces to avoid avoiding blind areas.

Inventors

WANG ZONGYUE
DU JING
XIAO LONGYUAN
CAO MENGYUN
WU DEFENG
WU JUN

Assignees

集美大学
厦门快商通科技股份有限公司
罗普特科技集团股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260312

Claims (10)

1. An intelligent shipping collision risk prediction and avoidance method based on large model reasoning is characterized by comprising the following steps: s1, acquiring multi-mode data in ship shipping and performing time scale alignment, wherein the multi-mode data comprises an image sequence of an optical camera, maritime radar points, an automatic recognition system track and sonar echoes; S2, projecting the maritime radar trace, the automatic recognition system trace and the image sequence of the optical camera to a overlook grid for weighted fusion to obtain fusion features; S3, carrying out feature enhancement on the spatial relationship of the alignment features to obtain enhanced features; S4, constructing a continuous time state, and decoding the continuous time state into the future time Obstacle occupancy and obstacle velocity of (2); s5, calculating the nearest approaching time and the nearest approaching distance of the ship body and the obstacle according to the occupation of the obstacle and the speed of the obstacle; S6, generating an underwater target set based on sonar echo, and converting the influence of the underwater obstacle on the water surface track into additional cost of the underwater obstacle according to the underwater target set; s7, analyzing navigation rule text and channel boundary to obtain space rule cost, and generating candidate track set based on the occupation of obstacle, speed of obstacle, space rule cost, additional cost of underwater obstacle and enhanced feature; S8, calculating track interaction cost according to the relative azimuth and speed of the candidate track set and the peripheral targets, calculating risk items based on the nearest time, nearest distance, candidate track set and space rule cost map, weighting and summing the risk items, track interaction cost, underwater obstacle additional cost and space rule cost to obtain comprehensive scores, and selecting the former one Dividing the altitude into tracks; s9, based on the front Constructing a prompt vector by using the altitude flight path, the alignment feature, the enhancement feature, the obstacle occupation and the obstacle speed, inputting the prompt vector into a large model to obtain nominal control of a driving ship dynamics model; S10, updating new environment beliefs according to real-time online observation values and checking, and if the verification is satisfied, driving a ship dynamics model to execute by using nominal control; and S11, solving the minimum modified safety projection control according to the feasible set, the correction control and the current motion state of the ship body to obtain final control, and driving the ship dynamics model to execute by using the final control.
2. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the alignment features are expressed as: ; Wherein, the Representing an alignment feature; representing non-negative sampling weight, adopting normalization constraint; And Representing the learned spatial offset; Representing fusion features Increasing the value after the space offset; representing the number of sample offsets set for each location; Representing coordinates.
3. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the spatial relationship of the alignment features is feature enhanced to obtain enhanced features, which specifically comprises the following steps: Computing alignment features Significance score of (2) ; Selecting the most significant score Coordinate sets corresponding to the positions are collected to obtain a center set; constructing inter-center relationship embeddings, expressed as: ; Wherein, the Representing the embedding of relationships between centers; Represent the first Points and the first The euclidean distance of the individual points, Representing a leachable map; Represent center set A plurality of points; Represent center set A plurality of points; Obtaining enhanced features based on inter-center relational embedding 。
4. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the continuous time state is represented as: ; ; Wherein, the Representing a continuous time state; Representing a parameterized vector field; representing a continuous drive input that is to be applied, Represent the first At the time of the secondary observation, Indicating the time status before the update, Representing the updated time status; Represent the first An observation abstract at the secondary observation moment; Representing an update operator; Said decoding from continuous-time state at future time The obstacle occupation and obstacle speed of (a) are specifically: ; ; Wherein, the Representing a linear readout matrix; representing a Sigmoid logical compression function; indicating future time of day Occupation of (C) A value; indicating future time of day Speed of (2) Values.
5. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the track interaction cost is expressed as: ; ; Wherein, the Representing the track interaction cost; Representing sector weights; Representing an angle domain influence function; Represent the first The center angle of the individual sectors is set, Represent the first The relative azimuth angle of the individual targets, Representing the weight as a function of distance and relative speed; Representing a periodic kernel function.
6. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the risk term is expressed as: ; Wherein, the Representing a risk item; representing meeting targets Is the closest approach distance to (a); representing meeting targets Is the most recent approach time to (1); The weight is represented by a weight that, Representing a threshold value; the composite score is expressed as: ; Wherein, the Representing the composite score; 、 And Representing the weight; Representing a space rule cost graph; representing the track interaction cost.
7. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the performing control calibration obtains correction control, performing set calibration based on correction control, candidate track set and historical success sample library to obtain a feasible set, and specifically comprising: For nominal control Performing correction control to obtain correction control ; According to correction control Performing introspection score and feasible set calibration, and expressing as: ; ; Wherein, the A difference between the score representing the consistency of the track and the observation and the threshold; a score representing the consistency of the track with the observation; Representing a track; representing an new observation; Representing a threshold value; representing a feasible set; represents a division point and is used for dividing the division point, Indicating the allowable error rate.
8. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the solving the minimum-change safety projection control is specifically implemented by using an optimization target to solve the minimum-change safety projection control, and the optimization target is expressed as: ,s.t. ; Wherein, the Representing the square of the modulus; represents the optimal variables to be used in the process of optimizing, S.t. represents a constraint condition; Representing a security function; representing the current motion state of the ship body; Representation Li Daoshu; Representing the relaxation coefficient.
9. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, wherein the additional cost of underwater obstacle is expressed as: ; Wherein, the Representing candidate tracks And the first A safe distance function of the underwater targets; A monotonic penalty function representing the underwater constraint; representing the weights.
10. The intelligent shipping collision risk prediction and avoidance method based on large model reasoning of claim 1, further comprising performing uncertainty control after S11, specifically comprising: Calculating the comprehensive uncertainty, expressed as: ; Wherein, the Representing the degree of uncertainty of the occupancy prediction, The degree of uncertainty in the plan is indicated, Representing a synthesis operator; if the uncertainty exceeds a preset uncertainty threshold, curtailing the feasible set And tighten the weight settings in the composite score calculation.

Description

Intelligent shipping collision risk prediction and avoidance method based on large model reasoning Technical Field The invention relates to the technical field of ship navigation, in particular to an intelligent shipping collision risk prediction and avoidance method based on large model reasoning. Background With the continued expansion of the global shipping industry, the environment in which ships are operated offshore is increasingly complex. The density of the offshore traffic flow is increased, the types of ships are various, severe weather and complex sea conditions frequently occur, and the traditional navigation mode depending on artificial experience is gradually difficult to adapt. The existing ship navigation is mainly based on manual operation and rule constraint, and although the requirements can be met under the conventional conditions, risks are very easy to occur under the conditions of insufficient visibility, information delay or multi-target interaction and the like. In recent years, intelligent shipping is proposed as a solution, and the core idea is to use a multi-mode sensor to realize environmental perception, and complete path planning and avoidance control through an intelligent algorithm, so as to improve autonomous navigation and safety level of a ship. However, intelligent shipping is still under exploration from the state of the art. Currently, most of the mainstream research focuses on single sensor or local function improvement, such as target detection using vision, range measurement using radar, ship identification using AIS, or underwater detection using sonar. While these approaches are effective in local scenarios, it is difficult to create a global understanding of the environment due to the lack of unified modeling and cross-modal alignment of data. The data of different modes have differences in time, space and noise characteristics, and direct fusion often causes error accumulation, so that downstream risk prediction and avoidance decision is affected. Meanwhile, a large model in the artificial intelligence field shows strong reasoning and generalization capability in recent years, and remarkable results are achieved in image recognition, natural language processing and three-dimensional scene modeling. The large model can be inferred and corrected in complex tasks through prompt construction and a self-learning mechanism, and possibility is provided for autonomous decision making in a dynamic environment. However, in the field of aeronautics, the application of such capabilities is still relatively limited. Most of the existing researches only stay in single-point application of a small-scale model, and multi-mode reasoning and introspection advantages of a large model are not fully exerted. Particularly in collision risk prediction and avoidance links of ship navigation, most of existing methods are based on static rules or fixed models for calculation, and lack dynamic adaptability and long-term optimization mechanisms. In addition, detection of the underwater environment avoids fracturing from the water surface for a long time. Submerged reefs, sunken ships or floating obstacles form a great threat to navigation, but current multi-reliance sonar or independent underwater detection methods cannot be modeled uniformly with a water surface collision avoidance strategy. Such fracturing results in circumvention decisions often covering only two-dimensional planes, ignoring potential risks in three-dimensional space. In addition, the sensor performance is reduced in severe weather, the model robustness is insufficient, and the intelligent shipping still has obvious defects in actual deployment. Therefore, there is a need for an intelligent shipping method that can integrate multi-modal awareness, large model reasoning and unified risk avoidance mechanisms to ensure that a ship realizes stable, safe and efficient autonomous navigation in a complex and diverse environment. The prior art has a plurality of limitations in the field of navigation autonomous navigation and avoidance, so that the system is difficult to realize stable and safe operation in a complex environment. First, there is a significant fracture in the sensor information. The vision sensor has high resolution, but the effect is obviously reduced under the condition of low illumination or strong fog, the radar has the advantages of long-distance detection, but has a blind area in short-distance obstacle recognition, the AIS provides flight path and identity information, the updating delay is larger, the quick-change environment cannot be reflected in real time, and the sonar can detect underwater obstacle and cannot be uniformly used with water surface perception. Such fracturing makes it difficult for the vessel to develop a continuous, comprehensive environmental understanding, directly affecting the accuracy of risk prediction. Second, the predictive and evasive functions are generally separated. Most m