Bridge collision warning method based on three-dimensional point cloud data for constructing a digital twin of a waterway

Abstract

The application discloses a bridge anti-collision early warning method based on construction of a channel digital twin based on three-dimensional point cloud data, and belongs to the field of channel traffic safety monitoring and bridge protection. In order to solve the problems that the three-dimensional position and posture of a ship are difficult to be acquired in real time under the condition of complex water flow and wind waves in a bridge area, risk assessment is inaccurate and prone to false positives and false negatives, the application loads a bridge and a channel model on a digital twin platform and establishes a structural safety boundary. The application obtains point clouds by using a bridge area laser radar and extracts a ship outer cuboid to form a state sequence of position, heading, speed, roll angle and pitch angle and a scene context feature sequence. A multi-scale attention timing network is used to predict a short-time position, and a physical constraint correction and adaptive fusion are carried out in combination with a ship maneuvering motion mathematical model. The minimum distance between the predicted outer cuboid and the safety boundary is calculated, and a hierarchical warning is output in combination with a clearance margin, so that the technical effects of improving the warning accuracy and robustness are achieved.

Description

Technical Field

[0001] This invention relates to the field of waterway traffic safety monitoring, and in particular to a bridge collision avoidance early warning method based on constructing a digital twin of a waterway using three-dimensional point cloud data. Background Technology

[0002] Bridges spanning waterways typically involve dense traffic, limited navigation space, and significant environmental disturbances such as currents and waves. Vessels yawing, drifting laterally, or changing attitude can easily lead to collisions with bridge piers or beams. To mitigate the risk of bridge collisions, existing technologies mainly include the following: First, passive protection through physical barriers such as anti-collision piers, piles, and rings; second, monitoring vessels using video surveillance, microwave radar, and manual lookout, and providing early warnings or interventions when approaching the bridge; third, acquiring vessel position and heading information based on Automatic Identification System (AIS) data, and using trajectory extrapolation or data-driven models for short-term trajectory prediction; and fourth, some solutions incorporate 3D modeling or digital platforms to visualize bridge and waterway elements for management and risk assessment.

[0003] However, the aforementioned existing technologies still have the following shortcomings:

[0004] First, the monitoring accuracy and environmental adaptability are insufficient. Video is affected by lighting and obstruction, and radar is easily interfered with by clutter. When relying solely on the Automatic Identification System (AIS), it is difficult to cover targets that are not activated or have abnormal signals. Furthermore, it is difficult to obtain real-time information on the ship's three-dimensional position and attitude information such as roll and pitch angles, resulting in insufficient assessment of airspace risk and attitude-related risks.

[0005] Second, the reliability of early warning criteria and prediction methods is insufficient. Some schemes only use distance thresholds or simple extrapolation, which are prone to false alarms or missed alarms under complex water flow and wind conditions. Moreover, the prediction results may not conform to the laws of ship maneuvering and movement, affecting the credibility of the early warning.

[0006] Third, the risk assessment of the bridge area is not sufficiently linked to the scene elements, making it difficult to comprehensively consider the interaction between the safety boundary of the bridge structure, the geometric constraints of the waterway and the envelope of the ship shape under a unified spatial coordinate system, thus making it difficult to meet the real-time and refined collision avoidance early warning needs of the bridge area.

[0007] Therefore, a bridge collision warning method is needed to address the shortcomings of the existing technologies. Summary of the Invention

[0008] One objective of this invention is to propose a bridge collision avoidance and early warning method based on a digital twin of a waterway constructed from 3D point cloud data. Addressing the shortcomings of existing technologies in monitoring accuracy and environmental adaptability under complex water flow and wave conditions in bridge areas, the difficulty in obtaining real-time ship 3D position and attitude information such as roll and pitch angles, and the tendency for false alarms and missed alarms due to predictions based solely on distance thresholds or simple data-driven methods, and the potential for prediction results not conforming to ship dynamics, the following technical solution is proposed: A 3D model of the bridge and waterway is loaded onto a digital twin platform, and a safety boundary for the bridge structure is established; a lidar system is deployed to collect point clouds, and the ship point clouds are segmented in a unified coordinate system; a ship outer cuboid is fitted, and a seven-tuple sequence containing position, heading, speed, roll angle, and pitch angle, as well as a scene context feature sequence containing relative position and clearance margin, are extracted; historical sequences are input into a multi-scale attention temporal prediction network and scene context features are injected to obtain predicted seven-tuples; physical feasibility constraints are corrected by combining the ship maneuvering motion mathematical model, and adaptive fusion is performed using fusion coefficients generated by a neural network; the minimum distance between the predicted outer cuboid and the safety boundary is calculated in the digital twin scene, and a graded early warning is output based on the clearance margin, with online updates performed as needed based on prediction errors. This invention has the technical effects of improving the accuracy of bridge collision risk assessment and the reliability of early warning, reducing false alarms and missed alarms, and enhancing robustness in complex environments.

[0009] This invention provides a bridge collision avoidance and early warning method based on constructing a digital twin of a waterway using 3D point cloud data, comprising:

[0010] S1. Load the three-dimensional models of the bridge and waterway onto the digital twin platform, establish the safety boundary of the bridge structure, and obtain the digital twin scene model;

[0011] S2. Deploy lidar at the bridge to collect and monitor the three-dimensional point cloud of the water area and convert it to the coordinate system of the digital twin scene model. Extract the ship point cloud from the three-dimensional point cloud and fit it to obtain the ship's outer cuboid.

[0012] S3. The ship's state seven-tuple is determined by the ship's outer cuboid and the ship's point cloud, and a seven-tuple sequence is formed. The seven-tuple includes position, heading, speed, roll angle and pitch angle. The scene context feature vector is determined by the digital twin scene model, the bridge structure safety boundary and the ship's outer cuboid, and a scene context feature sequence is formed. The scene context feature vector includes relative position information and clearance margin information.

[0013] S4. Input the historical fragments of the seven-tuple sequence and the scene context feature sequence into the temporal prediction neural network, and output the predicted seven-tuple within the preset prediction time domain. Input the predicted seven-tuple, environmental parameters, and ship parameters into the ship maneuvering motion mathematical model MMG to obtain the MMG output seven-tuple. Under the physical feasibility constraint, the predicted seven-tuple is corrected to obtain the corrected seven-tuple. The correction includes fusing the predicted seven-tuple and the MMG output seven-tuple according to the fusion coefficient, which is output by the neural network.

[0014] S5. Generate the predicted ship outer cuboid from the modified seven-tuple and load it into the digital twin scene model. Calculate the minimum distance between it and the safety boundary of the bridge structure. Combine the roll angle, pitch angle and clearance margin information to determine the collision risk assessment result.

[0015] S6. Output graded early warning information based on the collision risk assessment results.

[0016] Optionally, S1 includes:

[0017] High-precision 3D model data of the bridge and waterway are acquired and imported into the digital twin platform. Based on the high-precision 3D model data, a unified coordinate system consistent with the actual spatial position of the bridge and waterway is established to obtain the digital twin scene model.

[0018] The structural geometric information of the bridge piers and beams is extracted from the digital twin scene model, and the bridge structural safety boundary is generated based on the structural geometric information according to the preset safety distance and preset clearance threshold.

[0019] Optionally, S2 includes:

[0020] The lidar is deployed at a predetermined installation location on the bridge so that the detection field of the lidar covers the predetermined monitoring water area;

[0021] A high-precision positioning device is used to determine the spatial position and attitude of the lidar, and a coordinate transformation relationship between the lidar coordinate system and the digital twin scene model coordinate system is established based on the spatial position and attitude.

[0022] Based on the three-dimensional point cloud data of the preset monitoring water area collected by the lidar, the three-dimensional point cloud data is transformed into the coordinate system of the digital twin scene model using the coordinate transformation relationship to form a point cloud dataset;

[0023] Outlier removal and noise suppression are performed on the point cloud dataset to obtain a preprocessed point cloud;

[0024] Point cloud segmentation is performed on the preprocessed point cloud to obtain ship point cloud; minimum bounding box fitting is performed on the ship point cloud to obtain ship bounding box cuboid;

[0025] The point cloud quality index is calculated based on the ship point cloud and the ship's outer cuboid. The point cloud quality index includes the number of ship point cloud points, the ship point cloud density, and the minimum bounding box fitting error.

[0026] Optionally, S3 includes:

[0027] The three components of the ship's three-dimensional position parameters are determined based on the ship's outer cuboid, which are the coordinate components of the center point of the ship's outer cuboid in the coordinate system of the digital twin scene model.

[0028] The heading parameter is determined based on the major axis direction of the ship's outer cuboid, where the major axis direction is the principal direction of the ship's outer cuboid on the horizontal plane.

[0029] The velocity parameters are determined based on the coordinates of the center point of the ship's outer cuboid at adjacent time points and the time interval between adjacent time points;

[0030] The roll angle parameter and the pitch angle parameter are determined based on the ship point cloud. The roll angle parameter and the pitch angle parameter are determined by the tilt relationship of the attitude principal axis relative to the vertical direction obtained by principal component analysis of the ship point cloud. Target association is performed on the ship's outer cuboid at different times to form a ship trajectory. A target identifier is assigned to each ship. The target association is based on matching one or more of the following: the distance between the center points of the outer cuboids at adjacent times, the consistency of their velocities, and the similarity of the outer cuboid dimensions.

[0031] The ship status seven-tuples corresponding to multiple consecutive moments of the same target identifier are arranged into a seven-tuple sequence in chronological order; the scene context feature vector is calculated based on the digital twin scene model, the bridge structure safety boundary, and the ship outer cuboid; the relative position information includes the relative displacement information of the center point of the ship outer cuboid relative to the bridge structure safety boundary; and the clearance margin information includes the height margin between the highest point of the ship outer cuboid and the lowest clearance position of the corresponding bridge structure in the high-precision three-dimensional model.

[0032] The scene context feature vectors from multiple consecutive time points are arranged in chronological order to form a scene context feature sequence.

[0033] Optionally, S4 includes:

[0034] The seven-tuple sequence is time-aligned and normalized to obtain temporal input features, and the scene context feature sequence is encoded to obtain context input features.

[0035] The temporal input features are input into the multi-scale attention feature fusion temporal prediction neural network to obtain temporal latent features, and the temporal latent features are fused with the context input features through the cross attention mechanism, thereby outputting the prediction seven-tuple corresponding to the prediction step size Δt within the preset prediction time domain T.

[0036] The multi-scale attention feature fusion temporal prediction neural network adopts the pyramid multi-scale attention structure of the Pyraformer network. The environmental parameters include at least one of the following: water flow velocity, water flow direction, wind speed, wind direction, wave height, and water level. These environmental parameters are obtained by hydrological sensors deployed in the bridge area and / or by meteorological sensors deployed in the bridge area and / or by accessing the waterway management system database and / or by manual setting. The ship parameters include at least one of the following: ship length, ship width, draft, displacement, rudder area, propeller parameters, and MMG model hydrodynamic coefficients. These ship parameters are obtained from the ship static information database and / or from the static messages of the Automatic Identification System (AIS) and / or by matching a preset ship type parameter table and / or by manual input. The predicted 7-tuple, the environmental parameters, and the ship parameters are input into the ship maneuvering motion mathematical model MMG to obtain the MMG output 7-tuple. The physical feasibility constraints are constructed based on the model residual information between the predicted 7-tuple and the MMG output 7-tuple.

[0037] The environmental parameters, the point cloud quality index, and the model residual information are input into the fusion coefficient generation neural network to output the fusion coefficients. The determination of whether the predicted seven-tuple satisfies the physical feasibility constraints includes at least one of the following: the residual norm between the predicted seven-tuple and the MMG output seven-tuple is greater than a preset residual threshold; one or more of the longitudinal acceleration, lateral acceleration, bow angular velocity, roll angular velocity, and pitch angular velocity calculated by the predicted seven-tuple exceed a preset upper limit; the feasible region constraint consisting of one or more of the rudder angle range, steering rate range, and velocity change range given by the MMG model is violated. When the determination is not satisfied, the predicted seven-tuple and the MMG output seven-tuple are fused based on the fusion coefficients to obtain the corrected seven-tuple.

[0038] When the condition is met, the predicted 7-tuple is determined as the corrected 7-tuple, or the predicted 7-tuple and the MMG output 7-tuple are fused based on the fusion coefficient to obtain the corrected 7-tuple.

[0039] Furthermore, the context input features further include channel geometric constraint features, which include: the lateral offset of the center point of the ship's outer cuboid from the channel centerline; and the angle between the heading of the ship's outer cuboid and the tangential direction of the channel centerline.

[0040] Furthermore, the fusion coefficients are a sequence of fusion coefficients that change with the prediction step, and fusion coefficients are generated for different components in the seven-tuple respectively; wherein, corresponding fusion coefficients are generated for the position component, heading component, velocity component, roll component and pitch component respectively, so that the correction intensity for different components is independent of each other.

[0041] Optionally, S5 includes:

[0042] Using the corrected seven-tuple as input, the size parameters of the ship's outer cuboid corresponding to the ship are called, and the size parameters are transformed based on the three-dimensional position parameters, heading parameters, roll angle parameters and pitch angle parameters in the corrected seven-tuple to generate the predicted ship's outer cuboid corresponding to the prediction step size Δt within the preset prediction time domain T.

[0043] The predicted ship outer cuboid is loaded into the digital twin scene model, and both the predicted ship outer cuboid and the bridge structure safety boundary are characterized as convex polyhedra.

[0044] The minimum distance between the predicted ship's outer cuboid and the safety boundary of the bridge structure is calculated using a minimum distance calculation algorithm between convex polyhedra;

[0045] The minimum distance is combined with the roll angle parameter, the pitch angle parameter, and the clearance information to obtain the collision risk assessment result. The combined determination includes determining the lateral collision risk based on the minimum distance and determining the vertical collision risk based on the clearance information.

[0046] Optionally, S6 includes:

[0047] Based on the collision risk assessment results, graded early warning information is output;

[0048] Obtain the actual seven-tuple at the time corresponding to the corrected seven-tuple, and calculate the difference between the actual seven-tuple and the corrected seven-tuple to obtain the prediction error;

[0049] When the prediction error exceeds a preset error threshold, the prediction error, together with the temporal input features and contextual input features corresponding to the prediction error, are used to construct an online learning sample, and an online learning strategy is adopted to update the network parameters of the multi-scale attention feature fusion temporal prediction neural network.

[0050] Furthermore, the online learning strategy includes a concept drift detection step, which includes: determining a drift indicator based on the mean and variance changes of the prediction error within a preset time window; and triggering the update of the network parameters of the multi-scale attention feature fusion temporal prediction neural network only when the drift indicator exceeds a preset drift threshold, and performing weighted training on the online learning samples according to time decay weights.

[0051] The beneficial effects of this invention are:

[0052] 1. By extracting the ship's outer cuboid from the 3D point cloud of the bridge area lidar, and further obtaining the state seven-tuple of position, heading, speed, roll angle and pitch angle, the ship's 3D pose can be perceived in real time. Compared with the solution that only relies on video, radar or ship automatic identification system, the monitoring accuracy and environmental adaptability are improved, and the airspace-related risks can be effectively assessed.

[0053] 2. The scene context feature sequence is injected into a multi-scale attention temporal prediction network for short-term pose prediction, and a ship maneuvering motion mathematical model is introduced for physical feasibility constraint correction. At the same time, the fusion coefficient generation network adaptively fuses the data-driven prediction and the physical model output, making the prediction results more consistent with the laws of ship dynamics, reducing false alarms and false negatives caused by relying solely on distance thresholds or pure data-driven prediction, and improving the reliability of early warning.

[0054] 3. In the digital twin scenario, the minimum distance between the predicted ship outer body and the safety boundary of the bridge structure is calculated, and the lateral and vertical joint judgment is made in combination with the roll angle, pitch angle and clearance margin to form a graded early warning output. The online update can be triggered based on the prediction error, which improves the real-time performance, stability and robustness of collision avoidance early warning under complex working conditions in the bridge area. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 A flowchart illustrating a bridge collision avoidance and early warning method based on 3D point cloud data for constructing a digital twin of a waterway;

[0057] Figure 2 This is a flowchart of step S4 of the present invention, which describes short-term pose prediction and physical constraint correction. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figure 1 A bridge collision avoidance and early warning method based on 3D point cloud data to construct a digital twin of the waterway includes:

[0060] S1. Load the three-dimensional models of the bridge and waterway onto the digital twin platform, establish the safety boundary of the bridge structure, and obtain the digital twin scene model;

[0061] S2. Deploy lidar at the bridge to collect and monitor the three-dimensional point cloud of the water area and convert it to the coordinate system of the digital twin scene model. Extract the ship point cloud from the three-dimensional point cloud and fit it to obtain the ship's outer cuboid.

[0062] S3. The ship's state seven-tuple is determined by the ship's outer cuboid and the ship's point cloud, and a seven-tuple sequence is formed. The seven-tuple includes position, heading, speed, roll angle and pitch angle. The scene context feature vector is determined by the digital twin scene model, the bridge structure safety boundary and the ship's outer cuboid, and a scene context feature sequence is formed. The scene context feature vector includes relative position information and clearance margin information.

[0063] S4. Input the historical fragments of the seven-tuple sequence and the scene context feature sequence into the temporal prediction neural network, and output the predicted seven-tuple within the preset prediction time domain. Input the predicted seven-tuple, environmental parameters, and ship parameters into the ship maneuvering motion mathematical model MMG to obtain the MMG output seven-tuple. Under the physical feasibility constraint, the predicted seven-tuple is corrected to obtain the corrected seven-tuple. The correction includes fusing the predicted seven-tuple and the MMG output seven-tuple according to the fusion coefficient, which is output by the neural network.

[0064] S5. Generate the predicted ship outer cuboid from the modified seven-tuple and load it into the digital twin scene model. Calculate the minimum distance between it and the safety boundary of the bridge structure. Combine the roll angle, pitch angle and clearance margin information to determine the collision risk assessment result.

[0065] S6. Output graded early warning information based on the collision risk assessment results.

[0066] In this specific embodiment, S1 includes:

[0067] High-precision three-dimensional model data of bridges and waterways are acquired and imported into a digital twin platform. The high-precision three-dimensional model data includes triangular mesh models of bridge piers and beams, as well as vector elements of the center line and shoreline of the waterway. The length unit of all geometric data is uniformly set to meters and stored using the same engineering coordinate reference.

[0068] A unified coordinate system consistent with the actual spatial location of the bridge and waterway is established in the digital twin platform. This unified coordinate system is a right-handed coordinate system with the projection point of the bridge's navigation opening center as the origin. The axis points tangentially downstream along the centerline of the waterway. The axis points towards the left bank. The axis points vertically upwards and is opposite to the direction of gravity.

[0069] To ensure strict alignment between the imported model and the actual site location, at least three non-collinear coordinate control points were established on both sides of the bridge, and their measured coordinates in a unified coordinate system were obtained using a high-precision positioning device. Simultaneously, the model coordinates of corresponding control points were extracted from the high-precision 3D model data. Based on rigid body registration relationships, the model coordinates were transformed to a unified coordinate system, whereby the rigid body registration relationships satisfy… ,in This represents a column vector of three-dimensional coordinates of control points in a unified coordinate system. This represents the column vector of three-dimensional coordinates of the corresponding control points in the coordinate system of the high-precision three-dimensional model data. This indicates a rotation from the model coordinate system to the unified coordinate system. Orthogonal rotation matrix and satisfying This represents a three-dimensional translation column vector that translates from the model coordinate system to the unified coordinate system. express identity matrix Indicates the transpose operation;

[0070] The and The solution is obtained by minimizing the sum of squared registration residuals for all control point pairs. The process involves first calculating the centroids of the two sets of coordinates for each control point and then removing the mean. Next, a covariance matrix is ​​constructed for the corresponding point sets after the mean removal, and singular value decomposition is used to obtain the least squares result. Then, the calculation is performed using the centroid relation. and the and The model loading parameters are solidified into the digital twin platform to generate a digital twin scene model that is consistent with the actual site.

[0071] The structural geometric information of the bridge piers and beams is extracted from the digital twin scene model. The structural geometric information of the bridge piers consists of the outer contour polygon of the bridge pier triangular mesh after projection onto the horizontal plane and its elevation range. The structural geometric information of the beams consists of the set of lower edge facets in the area above the navigation hole in the beam triangular mesh and its corresponding minimum clearance elevation field.

[0072] Based on the structural geometry information, a bridge structural safety boundary is generated according to a preset safety distance and a preset clearance threshold, wherein the preset safety distance is set as follows: It is used to perform planar equidistant expansion of the outer contour polygon of the bridge pier to form the lateral safety boundary of the bridge pier. The expanded polygon is along... The shaft is at the design minimum navigable water level With the design maximum navigable water level The space between them is stretched to form a closed lateral boundary volume for subsequent spatial distance calculations, and the preset clearance threshold is set to... And it is used to shift the minimum clearance elevation of the beam as a whole downward. This forms the vertical safety boundary surface of the beam, thereby combining the "lateral safety boundary body of the pier" and the "vertical safety boundary surface of the beam" to constitute the safety boundary of the bridge structure and storing it in a triangular mesh on the digital twin platform.

[0073] In this specific embodiment, S2 includes:

[0074] A lidar unit is deployed at a pre-designated installation location on the bridge. This pre-designated location is the fixed support mounting point on the outer side of the cap beam of the upstream pier, with the lidar's emission window facing the center of the navigation channel. The installation height of the lidar is set to a uniform coordinate system. And ensure that its detection field of view covers the preset monitoring water area, which is defined in a unified coordinate system as follows: The spatial region, in which The minimum navigable water level determined in S1;

[0075] A high-precision positioning device is used to determine the spatial position and attitude of the lidar. This device employs a total station to measure the three-dimensional coordinates of the lidar's optical reference point in the coordinate system of the digital twin scene model, and an electronic inclinometer to measure the lidar's roll and pitch angles around its own coordinate axes. Simultaneously, the azimuth angle is determined by aligning the total station with the markings on the lidar's housing. This yields the rotation matrix and translation vector of the lidar coordinate system relative to the digital twin scene model coordinate system, and establishes coordinate transformation relationships accordingly.

[0076] ;

[0077] in This represents a column vector of three-dimensional coordinates of a point in the coordinate system of the digital twin scene model. This represents a column vector of three-dimensional coordinates of the same point in the lidar coordinate system. This represents the rotation from the lidar coordinate system to the digital twin scene model coordinate system. Orthogonal rotation matrix and satisfying This represents the three-dimensional translation column vector from the lidar coordinate system to the digital twin scene model coordinate system. express identity matrix Indicates the transpose operation;

[0078] The lidar acquires three-dimensional point clouds of a preset monitored water area at a frame rate of 10Hz and adds a timestamp to each frame of point cloud. The acquired raw point cloud is transformed to the coordinate system of the digital twin scene model through the coordinate transformation relationship to form a point cloud dataset. The point cloud dataset is stored in the order of timestamps when it is entered into the database, and each frame contains the three-dimensional coordinates of the points and the echo intensity.

[0079] Outlier removal and noise suppression were performed on the point cloud dataset to obtain a preprocessed point cloud. Outlier removal was performed using statistical outlier filtering with the number of nearest neighbor points set to 30 and the standard deviation multiplier threshold set to 2.0. Noise suppression was performed using voxel grid downsampling with the voxel side length set to 0.10m. Points not in the preset monitoring water area were directly cropped and removed.

[0080] Point cloud segmentation is performed on the preprocessed point cloud to obtain the ship point cloud. Point cloud segmentation includes static structure removal based on the digital twin scene model and target extraction based on Euclidean clustering. Static structure removal is performed by calculating the nearest distance from the point to the triangular mesh surface of the bridge pier, beam and shoreline in the digital twin scene model and deleting points belonging to static structures with a distance threshold of 0.20m. Target extraction performs Euclidean clustering on the remaining points, setting the cluster search radius to 0.80m, the minimum number of cluster points to 300, and the maximum number of cluster points to 200,000. The centroid of the clustering result is located within the navigation channel range and the centroid elevation meets the following conditions. The clustering was determined to be a ship point cloud, among which... To design the highest navigable water level;

[0081] Minimum bounding box fitting is performed on the ship point cloud to obtain the ship's outer cuboid. The minimum bounding box fitting adopts the directed bounding box construction method based on principal component analysis and uses the eigenvectors of the three-dimensional covariance matrix of the ship point cloud as the three orthogonal axis directions of the cuboid. Then, the minimum and maximum values ​​of the projection of the ship point cloud in the three axis directions are used as the positions of the six faces of the cuboid to obtain the center point, length, width and height dimensions and spatial orientation of the cuboid.

[0082] Point cloud quality indices are calculated based on ship point clouds and ship-enclosed cuboids. These indices include the number of ship point clouds. Ship point cloud density and minimum bounding box fitting error ,in Defined as the total number of points within the ship point cloud. Defined as The ratio of the point cloud volume to the volume of the ship's outer cuboid is used to characterize the point cloud's occupancy density within the target volume. Defined as the root mean square value of the minimum distance from each point in the ship's point cloud to the six faces of the ship's outer cuboid, it characterizes the consistency of the outer fitting and... and Write the point cloud quality record along with the corresponding timestamp.

[0083] In this specific embodiment, S3 includes:

[0084] Based on the ship point cloud and the ship's outer cuboid, the ship's state seven-tuple is calculated frame by frame in the coordinate system of the digital twin scene model and a seven-tuple sequence is formed. At the same time, the scene context feature vector is calculated based on the digital twin scene model, the bridge structure safety boundary and the ship's outer cuboid and a scene context feature sequence is formed.

[0085] For any timestamp The frame data uses the coordinates of the center point of the ship's outer cuboid as the three components of the position parameter and denotes them as follows: ,in This indicates the center point in the coordinate system of the digital twin scene model. Coordinate components along the axial direction, This indicates the center point in the coordinate system of the digital twin scene model. Coordinate components along the axial direction, This indicates the center point in the coordinate system of the digital twin scene model. Coordinate components along the axial direction;

[0086] The major axis direction of the ship's outer cuboid, in the horizontal plane, is used as the directional reference for the heading. This major axis direction is given by the first principal axis obtained during minimum bounding box fitting and is used to define the course reference. The projection of a plane determines the heading angle. The heading angle Using the coordinate system of digital twin scene model The positive direction of the axis is the zero-degree reference and is determined according to the right-hand rule. In a plane, counterclockwise is the positive direction;

[0087] The velocity parameters are calculated using the coordinate difference of the center point of the outer cuboid at adjacent time points and denoted as... Its calculation satisfies:

[0088] ;

[0089] in Indicates timestamp as The velocity scalar at time t. Indicates by The three-dimensional position column vector of the center point, Indicates timestamp as The column vector of the three-dimensional position of the center point at time. Describing the Euclidean norm, This represents the time interval between two adjacent frames, which is determined to be 0.1s by the lidar frame rate.

[0090] The roll angle parameter is denoted as The pitch angle parameter is denoted as Both are determined by the principal component analysis results of the ship point cloud. Specifically, the three-dimensional covariance matrix of the ship point cloud is calculated in the coordinate system of the digital twin scene model, and its eigenvectors are solved. The eigenvectors corresponding to the smallest eigenvalue and... The eigenvector with the smallest included-axis angle is determined as the principal axis of the ship's vertical attitude and denoted as . ,Will exist Projection of a plane and The angle between the axes is defined as the roll angle. ,Will exist Projection of a plane and The angle between the axes is defined as the pitch angle. When the right-hand rule determines the direction of the sign, causing the ship to list to port... When the bow is tilted upwards and the ship is upright It is positive;

[0091] The above position, heading, speed, roll angle, and pitch angle are combined into a seven-element group representing the ship's state. A seven-tuple sequence is formed by concatenating elements under the same target identifier in chronological order. The target identifier is obtained through target association. In each frame, target association performs trajectory matching on all ship outer cuboids and uses the Hungarian algorithm to solve for the minimum total cost matching. The cost is constructed from the center point distance, speed consistency, and outer cuboid size similarity, and matching is only allowed when the threshold constraints are met. The threshold constraints are set as follows: the center point distance does not exceed 15m, and the speed difference does not exceed [a certain value]. The relative differences in the length, width, and height of the outer cuboid do not exceed [a certain value]. Unmatched outer cuboids generate new target identifiers, and the output sequence terminates after 10 consecutive frames of unmatched target identifiers.

[0092] The scene context feature vector is calculated synchronously in each frame and includes relative position information and clearance margin information. The relative position information is obtained by calculating the relative displacement component from the center point of the ship's outer cuboid to the nearest point of the bridge structure's safety boundary. The nearest point search is achieved on the triangular mesh of the bridge structure's safety boundary by establishing a hierarchical structure of bounding boxes for triangular faces and calculating the shortest distance from a point to a triangular face. The clearance margin information is obtained by calculating the elevation of the highest point of the ship from the eight vertices of its outer cuboid and denoting it as [the highest point elevation]. Simultaneously, a rasterized query is performed on the triangular mesh at the lower edge of the beam in the digital twin scene model to obtain the lowest beam bottom elevation within the corresponding area projected onto the horizontal plane by the ship's outer cuboid, and this elevation is recorded as... The rasterization query uses a regular grid with a resolution of 0.20m and generates a "lowest beam bottom elevation field" offline after loading the model in S1 to ensure real-time performance. Finally, the net clearance margin is used as the basis for the query. The margin characterizing the vertical collision risk of the beam will be... As scene context feature vectors, they are concatenated in chronological order to form a scene context feature sequence, which is then bound to the corresponding target identifier for output.

[0093] In this specific embodiment, S4 includes:

[0094] The seven-tuple sequence corresponding to the same target identifier and the scene context feature sequence are time-aligned and normalized to obtain the network input. The time alignment uses the LiDAR timestamp and the sequence is resampled to a fixed step size. The length of the historical segment is set to The corresponding historical time domain is 3 seconds, and the prediction time domain is set to... And the corresponding prediction steps are The missing steps are filled by linear interpolation of adjacent valid steps and the most recent valid target identifier association result is used as the interpolation constraint. The normalization adopts the mean vector and standard deviation vector calculated on the training set during the offline training stage and is solidified into the running parameters of the digital twin platform. The seven-tuple sequence is normalized by components to form the temporal input feature, and the scene context feature sequence is normalized by components to form the context input feature.

[0095] The context input features consist of relative position information ( ), Net airspace margin information Together with the channel geometric constraints, the system comprises the lateral offset and heading deviation angle from the center point of the ship's outer cuboid to the channel centerline. The channel centerline, represented by the channel vector elements imported from S1, is a polyline with mileage and is parameterized by arc length before operation to support nearest point lookup. The lateral offset is calculated by connecting the center point to the nearest point on the centerline in the coordinate system of the digital twin scene model. The signed projection along the axis is obtained, and the direction pointing to the left bank is defined as positive. The heading deviation angle is calculated by the heading. The signed angle between the tangential direction at the nearest point to the centerline is obtained and the result is mapped to... To eliminate angular jumps;

[0096] The temporal input features and contextual input features are fed into a multi-scale attention feature fusion temporal prediction neural network to output a predicted seven-tuple. This multi-scale attention feature fusion temporal prediction neural network employs a pyramidal multi-scale attention structure of the Pyraformer network and fixes the network configuration to the input embedding dimension. The system employs a 3-layer pyramid structure with a downsampling factor of 2 per layer, 4 multi-head attention heads, a 256-dimensional hidden layer in the feedforward network with ReLU activation and a dropout rate of 0.1. Temporal input features are linearly mapped to form temporal latent features, while contextual input features are mapped to the same dimension via a two-layer fully connected encoder to become contextual latent features. At each pyramid layer, a cross-attention mechanism is used to fuse features in a manner where "temporal latent features are the query, and contextual latent features are the key and value," thus outputting the future... The predicted seven-tuple sequence is generated step by step and the intermediate residual information encoding for fusion coefficient generation is output synchronously.

[0097] Using the current 7-tuple as the initial state, and inputting environmental and ship parameters into the ship maneuvering motion mathematical model MMG, the MMG output 7-tuple sequence is obtained. The environmental parameters are obtained from the bridge area hydrological and meteorological sensors according to... The data is collected and fixed into six parameters: water flow velocity, water flow direction, wind speed, wind direction, wave height, and water level. The ship parameters are obtained from the ship static information database by querying the unique ship identifier associated with the target identifier and are fixed as follows: length, beam, draft, displacement, rudder area, propeller diameter, and a set of hydrodynamic coefficients from the MMG model. The MMG model uses a three-degree-of-freedom planar equation to calculate the pitch, sway, and heading motions of the ship's maneuvering motion, and obtains the future position and heading through numerical integration. The numerical integration method is fixed as a fourth-order Runge-Kutta with a step size of [missing value]. The velocity is obtained by combining the sway velocity and the transverse velocity, and the position... The component uses the current time. The elevation datum is obtained by superimposing the water level changes within the prediction time domain to ensure consistency with the digital twin scene model. The roll and pitch angles are calculated using a first-order inertial model. Based on the wave height drive and damped decay, the time constant is fixed at 2s to characterize the short-time attitude response of the bridge area;

[0098] Physical feasibility constraints are constructed and determined based on the predicted 7-tuple sequence and the MMG output 7-tuple sequence. The determination rule is fixed to simultaneously check the residual norm threshold, the upper limits of longitudinal and lateral acceleration obtained from the difference of the predicted 7-tuple, and the upper limits of bow angular velocity, roll angular velocity, and pitch angular velocity obtained from the difference of the predicted 7-tuple. The residual norm threshold is set as follows: The upper limit of longitudinal acceleration is set to The upper limit of lateral acceleration is set to The upper limit of the bow angular velocity is set to The maximum roll rate is set to The upper limit of the pitch angular velocity is set to And any violation of a rule will be determined as a failure to meet the physical feasibility constraint;

[0099] Environmental parameters, point cloud quality indicators, and model residual information are input into a fusion coefficient generation neural network to output a fusion coefficient sequence. The fusion coefficient generation neural network is fixed as a two-layer gated recurrent unit (GRU) and outputs seven components of the fusion coefficient at each prediction step, corresponding to different components of the seven-tuple, and changing with each prediction step. The fusion coefficients are compressed using a sigmoid function to... And recorded as When it is determined that the physical feasibility constraint is not met, The overall limit is set to no more than 0.3 to enhance the strength of the physical model correction, and this limit is maintained when the physical feasibility constraint is satisfied. Generate the raw output of the neural network for the fusion coefficients;

[0100] Based on the fusion coefficient, the predicted 7-tuple and the MMG output 7-tuple are adaptively fused to obtain the corrected 7-tuple, which is then output to S5. The fusion calculation satisfies the following:

[0101] ;

[0102] in Indicates the first The corrected seven-tuple column vector corresponding to each prediction step is then sorted according to... arrangement, Indicates the first Each prediction step corresponds to a prediction 7-tuple column vector, and the meaning of each component is... Consistent, Indicates the first Each prediction step corresponds to a 7-tuple column vector of MMG output, and the meaning of each component is the same as... Consistent Indicates the first The column vector of fusion coefficients corresponding to each prediction step and Representing the position Components, position Components, position weight, heading Components, speed Components, roll angle Components and pitch angle Component fusion weights, Indicates and A column vector of all 1s of the same dimension This represents the Hadamard product operation, which multiplies components. This represents the index in the resampled sequence at the current moment. Represents the prediction step index and satisfies .

[0103] In this specific embodiment, S5 includes:

[0104] Using the corrected 7-tuple sequence as input, a preset prediction time domain is generated. Each prediction step predicts the ship's outer cuboid shape and completes the collision risk assessment in the digital twin scenario model;

[0105] The current target identifier calls the dimensions of the ship's outer cuboid corresponding to the ship and fixes them as follows: ,in Indicates the length of the rectangular prism surrounding the ship and takes The median length of the bounding box obtained by minimum bounding box fitting over the most recent 10 frames. This represents the width of the ship's outer cuboid, taken as the median of the most recent 10 frames. The height of the ship's outer cuboid is represented by the median of the most recent 10 frames, thereby suppressing size jitter caused by point cloud occlusion in a single frame;

[0106] For the The correction septum corresponding to each prediction step Generate the coordinates of the eight vertices of the predicted ship's outer cuboid, where These represent the coordinate components of the center point of the outer cuboid in the coordinate system of the digital twin scene model. Indicates the heading angle. Indicates the roll angle. The pitch angle is represented by a local coordinate system of the ship's hull, with the center point of the outer cuboid as the origin. Within this local coordinate system, the relative coordinates of the eight vertices are constructed. Then, following a fixed rotation sequence, the local coordinates are sequentially rotated around the coordinate system of the digital twin scene model. Axis rotation , local area after rotation Axis rotation , local area after rotation Axis rotation And complete the translation ( The coordinates of the eight vertices of the prediction step are obtained, thus forming the first step. Each prediction step predicts the ship's outer cuboid shape, denoted as a convex polyhedron. and will The data is loaded into the digital twin scene model using a triangular mesh for spatial computation.

[0107] Subsequently, the safety boundary of the bridge structure was also characterized as a set of convex polyhedra. ,in represents the number of convex polyhedrons, Indicates the first A convex polyhedron, the lateral safety boundary body of the pier, after being generated in S1, performs triangulation on its horizontal cross-sectional outer contour polygon and divides each triangle along... Axis in Stretching within a range to form a triangular prism to ensure each As a convex polyhedron, the vertical safety boundary surface of the beam, after being generated in S1, is discretized into several quadrilateral patches using a regular grid with a resolution of 0.20m, and each patch is given a fixed thickness along both sides of the normal direction. The material is stretched to form a thin box to ensure it can be used in the distance calculation of the convex polyhedron.

[0108] Calculate for each prediction step and The minimum distance between them is denoted as Its definition is:

[0109] ;

[0110] in Indicates the first The minimum distance in each prediction step. This indicates a predicted ship encased in a rectangular convex polyhedron. A column vector of the three-dimensional coordinates of any point on the surface. The first term representing the safety boundary of a bridge structure A convex polyhedron A column vector of the three-dimensional coordinates of any point on the surface. The minimum distance calculation uses the GJK convex polyhedron distance algorithm and constructs a Minkowski difference volume using kernel-supported mapping to iteratively search for the nearest simple form. The upper limit of iteration is set to 30 times and the distance convergence threshold is set to 0.01m. and Output directly when they intersect ;

[0111] Simultaneously, the vertical clearance margin is calculated in each prediction step and denoted as... Its calculation uses The maximum elevation of the eight vertices is used as the elevation of the highest point of the ship, and the lowest beam bottom elevation field is called in. The minimum beam bottom elevation is taken within the grid cell covered by the horizontal projection, and the difference between the two is obtained. and They are used together for joint determination, among which Used for side collision risk assessment, with a fixed threshold. , Used for vertical collision risk assessment, with a fixed threshold. and with preset air clearance threshold Maintain consistency;

[0112] In the joint decision, if any prediction step satisfies or If any prediction step satisfies the following conditions, the collision risk assessment result is determined to be high risk. or Then it is determined to be of medium risk and meets the following conditions in the same prediction step. or The risk level was raised from medium to high, among which The pitch angle is expressed in radians. The pitch angle is expressed in radians, and if all prediction steps satisfy the following conditions: and If the collision is classified as low risk, it is classified as moderate risk. The collision risk assessment result is then compared with the minimum threshold that would trigger the result. and minimum The information is output together to provide S6 with tiered early warning information.

[0113] In this specific embodiment, S6 includes:

[0114] Based on the collision risk assessment results, a tiered early warning system is generated and released. The tiered early warning system is divided into four levels, and the fixed mapping relationship is that low risk corresponds to... General risk response Medium risk response High-risk response ,in Indicates the warning level code;

[0115] when The system sends a warning to the waterway management system and highlights the target identifier and navigation hole number on the digital twin platform interface. Simultaneously, the navigation channel management system and the on-site audio-visual terminal in the bridge area send a warning-level alert and display the predicted risk step index, corresponding minimum distance, and minimum clearance margin on the interface. The system simultaneously sends emergency warnings to the waterway management system, the on-site audio-visual terminal in the bridge area, and the emergency linkage interface, and locks the display color and flashing frequency of the target identifier in the digital twin scene model;

[0116] Simultaneously, a "prediction-measurement" alignment cache is established to obtain the actual 7-tuple at the time corresponding to the corrected 7-tuple and to calculate the prediction error. Specifically, when outputting the corrected 7-tuple sequence in S4, the prediction error is calculated by... The corresponding corrected 7-tuple is written to the cache and its target time is marked as [time value]. When S3 is at time Output the actual 7-tuple The corrected 7-tuple is read from the cache at the same time. ),in The actual position components are defined in the coordinate system of the digital twin scene model. This is the actual heading angle. For actual speed, This is the actual roll angle. This is the actual pitch angle. These are the corrected output components corresponding to the same moment;

[0117] Prediction error is denoted as The position error is obtained by normalizing and summing the component errors according to deterministic rules. The position error is taken as the Euclidean distance of the three-dimensional position difference and expressed as follows: Normalization, heading error is taken as Mapped to The absolute value after and with Normalization, speed error is taken and with Normalization, roll angle error is taken as and with Normalization, pitch angle error is taken as and with Normalization, the sum of the five normalization errors yields And set the error threshold to Used to determine whether a single prediction is inaccurate;

[0118] The online learning strategy includes a concept drift detection step and triggers a network update only when drift exists. Concept drift detection determines the drift indicator based on the change in the mean and variance of the prediction error within a preset time window, with the preset time window length being fixed. Furthermore, corresponding to 30 seconds, the reference statistic is calculated and fixed within the first stable window after the system is put into operation. and The drift indicator is denoted as And calculate according to the following formula:

[0119] ;

[0120] in Indicates the time index as The drift indicator at that time Indicates the most recent Prediction error The calculated mean Indicates the most recent Prediction error The calculated standard deviation, This represents the mean of the prediction error of the reference window. This represents the standard deviation of the reference window prediction error. This represents the absolute value operation. Indicates the length of the time window;

[0121] Drift threshold set to If and only if both conditions are met and The temporal input features corresponding to that moment, along with the context input features and the actual seven-tuple, are used to construct an online learning sample. The input of the online learning sample is the input from S4 used to generate the target moment. The predicted historical segment temporal input features and contextual input features, with the supervised label of the online learning sample being the target time. The actual seven-tuple;

[0122] Online learning sample input capacity is The sample queue is weighted and trained according to time decay weights, with the time decay coefficient fixed at 1. Furthermore, the newer the sample, the greater its weight.

[0123] When an update is triggered, online fine-tuning is performed on the multi-scale attention feature fusion temporal prediction neural network, updating only the network parameters without changing the MMG parameters of the ship maneuvering motion mathematical model. The online fine-tuning uses the Adam optimizer with a fixed learning rate. The batch size is fixed at 32, the gradient update steps are fixed at 100, the loss function adopts the weighted SmoothL1 loss of each component of the seven-tuple, and the weights of the position component, heading component, velocity component, roll component, and pitch component are set to 2, 1, 1, and 1 respectively. After the update is completed, the new network parameters are written into the running instance and the subsequent prediction and correction results are output according to S4. The graded early warning information is also output according to this step.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0125] This invention employs a holistic process: "digital twin scene modeling and safety boundary construction, 3D point cloud perception using LiDAR in the bridge area, fitting of the ship's outer cuboid and extraction of the seven-tuple pose sequence, short-term pose prediction and physical constraint correction, and joint assessment of minimum distance and clearance." This process unifies the calculation of bridge structure geometry, waterway spatial constraints, and the ship's 3D motion state into a single coordinate system. This not only enables real-time acquisition of the ship's 3D position and attitude information such as roll and pitch angles, but also allows for precise determination of the spatial distance and clearance margin between the ship's outer cuboid and the bridge structure's safety boundary within the prediction time domain. This improves the accuracy of collision risk assessment and the reliability of graded early warning under complex water flow and wave conditions in bridge areas, while reducing false alarms and missed alarms caused by single distance thresholds or single-source data.

[0126] To address the technical challenges of "strong environmental disturbances, significant attitude and clearance risks, and susceptibility to distortion in purely data-driven predictions," this invention makes targeted improvements to the algorithm structure: First, it constructs temporal features from the contextual features of the digital twin scene and injects them into the prediction network through a cross-attention mechanism. This allows the model to explicitly utilize constraint information such as bridge safety boundaries, relative positions, and clearance margins during the prediction process, enhancing its ability to perceive spatial constraints in the bridge area. Second, it introduces a mathematical model of ship maneuvering motion to construct physical feasibility constraints and corrects the neural network prediction results through residual and feasible region determination, making the predicted trajectory more consistent with the laws of ship maneuvering motion. Third, it sets up a fusion coefficient generation network and adaptively determines the fusion strength by combining environmental parameters, point cloud quality indicators, and model residual information, achieving robust fusion for different operating conditions and different prediction step sizes, thereby further improving the accuracy and robustness of early warnings.

Claims

1. A bridge collision avoidance and early warning method based on constructing a digital twin of a waterway using 3D point cloud data, characterized in that, include: S1. Load the three-dimensional models of the bridge and waterway onto the digital twin platform, establish the safety boundary of the bridge structure, and obtain the digital twin scene model; S2. Deploy lidar at the bridge to collect and monitor the three-dimensional point cloud of the water area and convert it to the coordinate system of the digital twin scene model. Extract the ship point cloud from the three-dimensional point cloud and fit it to obtain the ship's outer cuboid. S3. The ship's state seven-tuple is determined by the ship's outer cuboid and the ship's point cloud, and a seven-tuple sequence is formed. The seven-tuple includes position, heading, speed, roll angle and pitch angle. The scene context feature vector is determined by the digital twin scene model, the bridge structure safety boundary and the ship's outer cuboid, and a scene context feature sequence is formed. The scene context feature vector includes relative position information and clearance margin information. S4. Input the historical fragments of the seven-tuple sequence and the scene context feature sequence into the temporal prediction neural network, and output the predicted seven-tuple within the preset prediction time domain. Input the predicted seven-tuple, environmental parameters, and ship parameters into the ship maneuvering motion mathematical model MMG to obtain the MMG output seven-tuple. Under the physical feasibility constraint, the predicted seven-tuple is corrected to obtain the corrected seven-tuple. The correction includes fusing the predicted seven-tuple and the MMG output seven-tuple according to the fusion coefficient, which is output by the neural network. S5. Generate the predicted ship outer cuboid from the modified seven-tuple and load it into the digital twin scene model. Calculate the minimum distance between it and the safety boundary of the bridge structure. Combine the roll angle, pitch angle and clearance margin information to determine the collision risk assessment result. S6. Output graded early warning information based on the collision risk assessment results.

2. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway, as described in claim 1, is characterized in that... S1 includes: High-precision 3D model data of the bridge and waterway are acquired and imported into the digital twin platform. Based on the high-precision 3D model data, a unified coordinate system consistent with the actual spatial position of the bridge and waterway is established to obtain the digital twin scene model. The structural geometric information of the bridge piers and beams is extracted from the digital twin scene model, and the bridge structural safety boundary is generated based on the structural geometric information according to the preset safety distance and preset clearance threshold.

3. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway as described in claim 1, characterized in that, S2 include: The lidar is deployed at a predetermined installation location on the bridge so that the detection field of the lidar covers the predetermined monitoring water area; A high-precision positioning device is used to determine the spatial position and attitude of the lidar, and a coordinate transformation relationship between the lidar coordinate system and the digital twin scene model coordinate system is established based on the spatial position and attitude. Based on the three-dimensional point cloud data of the preset monitoring water area collected by the lidar, the three-dimensional point cloud data is transformed into the coordinate system of the digital twin scene model using the coordinate transformation relationship to form a point cloud dataset; Outlier removal and noise suppression are performed on the point cloud dataset to obtain a preprocessed point cloud; Point cloud segmentation is performed on the preprocessed point cloud to obtain ship point cloud; minimum bounding box fitting is performed on the ship point cloud to obtain ship bounding box cuboid; The point cloud quality index is calculated based on the ship point cloud and the ship's outer cuboid. The point cloud quality index includes the number of ship point cloud points, the ship point cloud density, and the minimum bounding box fitting error.

4. The bridge collision avoidance and early warning method based on constructing a digital twin of a waterway using three-dimensional point cloud data according to claim 1, characterized in that, S3 includes: The three components of the ship's three-dimensional position parameters are determined based on the ship's outer cuboid, which are the coordinate components of the center point of the ship's outer cuboid in the coordinate system of the digital twin scene model. The heading parameter is determined based on the major axis direction of the ship's outer cuboid, where the major axis direction is the principal direction of the ship's outer cuboid on the horizontal plane. The velocity parameters are determined based on the coordinates of the center point of the ship's outer cuboid at adjacent time points and the time interval between adjacent time points; The roll angle parameter and the pitch angle parameter are determined based on the ship point cloud. The roll angle parameter and the pitch angle parameter are determined by the tilt relationship of the attitude principal axis relative to the vertical direction obtained by principal component analysis of the ship point cloud. Target association is performed on the ship's outer cuboid at different times to form a ship trajectory. A target identifier is assigned to each ship. The target association is based on matching one or more of the following: the distance between the center points of the outer cuboids at adjacent times, the consistency of their velocities, and the similarity of the outer cuboid dimensions. The ship status seven-tuples corresponding to multiple consecutive moments of the same target identifier are arranged into a seven-tuple sequence in chronological order; the scene context feature vector is calculated based on the digital twin scene model, the bridge structure safety boundary, and the ship outer cuboid; the relative position information includes the relative displacement information of the center point of the ship outer cuboid relative to the bridge structure safety boundary; and the clearance margin information includes the height margin between the highest point of the ship outer cuboid and the lowest clearance position of the corresponding bridge structure in the high-precision three-dimensional model. The scene context feature vectors from multiple consecutive time points are arranged in chronological order to form a scene context feature sequence.

5. The bridge collision avoidance and early warning method based on constructing a digital twin of a waterway using three-dimensional point cloud data according to claim 1, characterized in that, S4 includes: The seven-tuple sequence is time-aligned and normalized to obtain temporal input features, and the scene context feature sequence is encoded to obtain context input features. The temporal input features are input into the multi-scale attention feature fusion temporal prediction neural network to obtain temporal latent features, and the temporal latent features are fused with the context input features through the cross attention mechanism, thereby outputting the prediction seven-tuple corresponding to the prediction step size Δt within the preset prediction time domain T. The multi-scale attention feature fusion temporal prediction neural network adopts the pyramid multi-scale attention structure of the Pyraformer network. The environmental parameters include at least one of the following: water flow velocity, water flow direction, wind speed, wind direction, wave height, and water level. These environmental parameters are obtained by hydrological sensors deployed in the bridge area and / or by meteorological sensors deployed in the bridge area and / or by accessing the waterway management system database and / or by manual setting. The ship parameters include at least one of the following: ship length, ship width, draft, displacement, rudder area, propeller parameters, and MMG model hydrodynamic coefficients. These ship parameters are obtained from the ship static information database and / or from the static messages of the Automatic Identification System (AIS) and / or by matching a preset ship type parameter table and / or by manual input. The predicted 7-tuple, the environmental parameters, and the ship parameters are input into the ship maneuvering motion mathematical model MMG to obtain the MMG output 7-tuple. The physical feasibility constraints are constructed based on the model residual information between the predicted 7-tuple and the MMG output 7-tuple. The environmental parameters, the point cloud quality index, and the model residual information are input into the fusion coefficient generation neural network to output the fusion coefficients. The determination of whether the predicted seven-tuple satisfies the physical feasibility constraints includes at least one of the following: the residual norm between the predicted seven-tuple and the MMG output seven-tuple is greater than a preset residual threshold; one or more of the longitudinal acceleration, lateral acceleration, bow angular velocity, roll angular velocity, and pitch angular velocity calculated by the predicted seven-tuple exceed a preset upper limit; the feasible region constraint consisting of one or more of the rudder angle range, steering rate range, and velocity change range given by the MMG model is violated. When the determination is not satisfied, the predicted seven-tuple and the MMG output seven-tuple are fused based on the fusion coefficients to obtain the corrected seven-tuple. When the condition is met, the predicted 7-tuple is determined as the corrected 7-tuple, or the predicted 7-tuple and the MMG output 7-tuple are fused based on the fusion coefficient to obtain the corrected 7-tuple.

6. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway according to claim 1, characterized in that, S5 include: Using the corrected seven-tuple as input, the size parameters of the ship's outer cuboid corresponding to the ship are called, and the size parameters are transformed based on the three-dimensional position parameters, heading parameters, roll angle parameters and pitch angle parameters in the corrected seven-tuple to generate the predicted ship's outer cuboid corresponding to the prediction step size Δt within the preset prediction time domain T. The predicted ship outer cuboid is loaded into the digital twin scene model, and both the predicted ship outer cuboid and the bridge structure safety boundary are characterized as convex polyhedra. The minimum distance between the predicted ship's outer cuboid and the safety boundary of the bridge structure is calculated using a minimum distance calculation algorithm between convex polyhedra; The minimum distance is combined with the roll angle parameter, the pitch angle parameter, and the clearance information to obtain the collision risk assessment result. The combined determination includes determining the lateral collision risk based on the minimum distance and determining the vertical collision risk based on the clearance information.

7. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway according to claim 1, characterized in that, S6 include: Based on the collision risk assessment results, graded early warning information is output; Obtain the actual seven-tuple at the time corresponding to the corrected seven-tuple, and calculate the difference between the actual seven-tuple and the corrected seven-tuple to obtain the prediction error; When the prediction error exceeds a preset error threshold, the prediction error, together with the temporal input features and contextual input features corresponding to the prediction error, are used to construct an online learning sample, and an online learning strategy is adopted to update the network parameters of the multi-scale attention feature fusion temporal prediction neural network.

8. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway, as described in claim 5, is characterized in that... The context input features further include channel geometric constraint features, which include: the lateral offset of the center point of the ship's outer cuboid from the channel centerline; and the angle between the heading of the ship's outer cuboid and the tangential direction of the channel centerline.

9. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway, as described in claim 5, is characterized in that... The fusion coefficients are a sequence of fusion coefficients that change with the prediction step, and fusion coefficients are generated for different components in the seven-tuple respectively; wherein, corresponding fusion coefficients are generated for the position component, heading component, velocity component, roll component and pitch component respectively, so that the correction intensity for different components is independent of each other.

10. The bridge collision avoidance and early warning method based on three-dimensional point cloud data to construct a digital twin of a waterway, as described in claim 7, is characterized in that... The online learning strategy includes a concept drift detection step, which includes: determining a drift indicator based on the mean and variance changes of the prediction error within a preset time window; and triggering the update of the network parameters of the multi-scale attention feature fusion temporal prediction neural network only when the drift indicator exceeds a preset drift threshold, and performing weighted training on the online learning samples according to time decay weights.

Images