A method for updating dynamic data visualization of submarine pipeline by integrating beidou positioning

Abstract

A method for dynamic data visualization and updating of subsea pipelines integrating BeiDou positioning is disclosed, relating to the field of marine engineering digitization and 3D visualization technology. This method constructs a marine digital twin data architecture consisting of a perception layer, a data layer, a fusion layer, a model layer, and an application layer. The perception layer integrates BeiDou positioning terminals, pressure sensors, temperature sensors, and sonar detection equipment deployed along the subsea pipeline to collect pipeline position coordinate sequences, environmental parameters, and seabed topography data in real time. The data layer standardizes, cleans, and encrypts multi-source heterogeneous data for storage. The fusion layer employs an adaptive Kalman filter algorithm to tightly couple BeiDou positioning data and inertial navigation data to obtain centimeter-level accuracy in the dynamic spatial position of the pipeline. By constructing a real-time fusion architecture of high-precision BeiDou positioning data and multi-source sensor data, dynamic visualization updates and secure interaction of the subsea pipeline's spatial position, operational status, and structural morphology are achieved.

Description

Technical Field

[0001] This invention relates to the field of marine engineering digitization and 3D visualization technology, and to a method for dynamic data visualization and updating of subsea pipelines that integrates BeiDou positioning. Background Technology

[0002] In the current marine sector, driven by the dual demands of energy security and marine ecological protection, the importance of real-time sensing technology for the structural integrity and operational status of subsea pipelines continues to rise. As marine engineering increasingly extends into deep-sea areas, pipeline networks are expanding, and the complexity of operating environments is increasing year by year, leading to higher safety risks for subsea pipelines. Existing subsea pipeline management models suffer from data update delays; static 3D scenes constructed from offline imported surveying data cannot reflect in real-time the deformation, displacement, and structural response of pipelines under dynamic loads in complex environments. Furthermore, conventional positioning methods lack accuracy, traditional electronic charts and geographic information systems have low integration levels, and there is a lack of high-precision positioning methods to support dynamic calibration of pipeline spatial positions, resulting in discrepancies between 3D models and actual geographic coordinates. Data security mechanisms are also lacking; sensitive pipeline data involving geological structures and energy transmission lacks effective encryption and desensitization processing during transmission and visualization, posing a risk of data leakage. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, the present invention aims to provide a method for dynamic visualization and updating of subsea pipeline data that integrates BeiDou positioning. This method constructs a real-time fusion architecture of BeiDou high-precision positioning data and multi-source sensor data, and combines a lightweight rendering engine based on WebGPU and an intelligent deformation prediction model to achieve dynamic visualization and safe interaction of the spatial location, operating status, and structural morphology of subsea pipelines.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for visualizing and updating dynamic data of subsea pipelines using BeiDou positioning, comprising the following steps:

[0006] Step S1: Construct a marine digital twin data architecture consisting of a perception layer, a data layer, a fusion layer, a model layer, and an application layer. The perception layer integrates BeiDou positioning terminals, pressure sensors, temperature sensors, and sonar detection equipment deployed on the subsea pipeline to collect pipeline location coordinate sequences, environmental parameters, and seabed topography data in real time. The data layer standardizes, cleans, and encrypts multi-source heterogeneous data for storage. The fusion layer uses an adaptive Kalman filter algorithm to tightly couple BeiDou positioning data and inertial navigation data to obtain the dynamic spatial position of the pipeline with centimeter-level accuracy. The model layer uses a two-stage graph neural network that integrates physical information and attention mechanisms as its core carrier. Through three core components—physically enhanced graph convolutional layers, multi-head adaptive graph attention mechanisms, and parallel spatiotemporal feature fusion modules—it provides accurate deformation prediction data support for visualization updates. The application layer relies on a lightweight 3D visualization rendering engine built on WebGPU. It receives centimeter-level high-precision real-time positioning data from the fusion layer and deformation prediction data from the model layer as its core driver, ultimately realizing the dynamic visualization and interaction of the spatial position, operating status, and structural morphology of the subsea pipeline, adapting to the visualization needs of ultra-long-distance pipeline scenarios.

[0007] Step S2: Build a 3D visualization rendering engine based on the WebGPU graphics interface. Implement parallel decoding and vertex transformation of the digital elevation model of the seabed topography and the 3D model of the pipeline through the computation shader. Use an improved ray casting algorithm to perform volume rendering of the scalar field of the ocean water body to generate a real ocean environment scene containing ocean current gradient, temperature field and pressure field.

[0008] Step S3: Construct a two-stage graph neural network that integrates physical information and attention mechanism as the pipeline deformation prediction model. It includes a physically enhanced graph convolutional layer combined with a four-head adaptive graph attention mechanism and a parallel spatiotemporal feature fusion module. The historical location sequence obtained by the Beidou positioning terminal and the environmental load data collected by the sensor are used as input features. The graph convolutional layer is used to capture the spatial correlation features between each monitoring point of the pipeline. Finally, the temporal evolution features are extracted through a gated recurrent unit to output the predicted deformation displacement and stress distribution cloud map of the pipeline at future time.

[0009] Step S4: Inject the high-precision pipeline positioning data calculated in real time in step S1 and the predicted deformation data generated in step S3 into the 3D visualization rendering engine to drive the real-time update of the geometric mesh vertices of the pipeline model and realize the visualization mapping of the dynamic deformation process; at the same time, overlay the seabed image data collected by multibeam sonar with the digital elevation model to generate realistic seabed surface texture.

[0010] Step S5: Establish a data security desensitization and encryption transmission channel based on SM4. By dynamically deflecting and remapping the geographic coordinates of classified pipelines, ensure that the visualization terminal does not display the original precise coordinate data; use the SM9 identifier cryptography algorithm to encrypt the sensor data stream in the transmission link end-to-end to achieve secure interaction of sensitive data.

[0011] Furthermore, the adaptive Kalman filter algorithm used in the fusion layer in step S1 includes a noise adaptive estimation module based on information covariance matching. This module adjusts the filter gain matrix in real time to suppress positioning noise caused by multipath effects and sea state changes, and the output stability is better than that of differential positioning results of single-point positioning.

[0012] Furthermore, in step S2, the volume rendering based on WebGPU adopts an adaptive step size control and an early ray termination strategy. The sampling step size is dynamically adjusted according to the seabed topography and pipeline spatial distribution. A fine sampling step size is used for regions of interest containing pipelines, and a coarse sampling step size is used for water areas without targets to reduce computational overhead.

[0013] Furthermore, the dual-stage graph neural network constructed in step S3, which integrates physical information and attention mechanism, includes three layers of physically enhanced graph convolutional layers, a four-head adaptive graph attention mechanism, and two layers of parallel spatiotemporal feature fusion modules. The graph convolutional layers capture the spatial correlation features between pipeline nodes by normalizing the Laplacian matrix with mechanical coupling weights. The bidirectional gated recurrent unit extracts the forward and backward correlation features of the time series of each monitoring point. Finally, the multidimensional deformation prediction results are output through the fully connected layer.

[0014] Furthermore, in step S4, the pipeline model's geometric mesh vertex update adopts a deformation constraint solution method based on position dynamics, using BeiDou positioning data as position constraints and the material mechanics constitutive model as energy constraints, and generating a smooth deformation surface that satisfies physical laws through iterative solution.

[0015] Furthermore, in step S5, the dynamic deflection processing adopts a grid scrambling algorithm based on geofencing. While maintaining the topological connectivity of the pipeline network, the same pseudo-random deflection amount is applied to the coordinate points falling into the same grid cell to achieve reversible spatial location desensitization.

[0016] Furthermore, the physically enhanced graph convolutional layer introduces material mechanics constraints during the graph convolution process; in the adjacency matrix construction stage, adjacency relationships are defined based on pipeline topology connections, and the mechanical coupling coefficient is used as a weighting factor for the adjacency matrix; the formula for calculating the mechanical coupling coefficient is:

[0017]

[0018] Where E is the elastic modulus of the pipe, I is the moment of inertia of the section, and L... ij The length of the pipe segment between adjacent monitoring nodes i and j; this coefficient characterizes the mechanical sensitivity of deformation transmission between adjacent nodes.

[0019] Furthermore, a four-head adaptive graph attention mechanism is introduced after the physically enhanced graph convolutional layer; the adaptive graph attention mechanism calculates the attention coefficients between nodes and dynamically adjusts the information aggregation weights.

[0020] The attention coefficient is calculated by taking into account both node feature similarity and physical distance.

[0021]

[0022] Where h i with h j W represents the feature vectors of node i and node j, respectively. q With W k It is a linear transformation matrix. (d ij ) is based on physical distance d ij Radial basis function encoding, where || denotes vector concatenation operation. The attention vector is trained; the attention coefficients are normalized using the softmax function to obtain the final aggregate weights; the four-head adaptive graph attention mechanism concatenates or averages the features output by multiple independent attention heads to enhance the model's ability to express different feature subspaces.

[0023] Furthermore, the parallel spatiotemporal feature fusion module includes a space-first branch and a time-first branch;

[0024] The spatial priority branch first extracts spatial features through a physically enhanced graph convolutional layer, and then inputs them into a bidirectional gated recurrent unit to extract temporal features;

[0025] The time-first branch first extracts the independent temporal features of each node through a bidirectional gated recurrent unit, and then performs spatial information interaction through a physically enhanced graph convolutional layer.

[0026] The output features of the two branches are adaptively weighted and fused through a gated fusion unit:

[0027]

[0028] Where F spatial For space-first branch output, F temporal For time-priority branch output, W z σ is a trainable weight matrix, and σ is the sigmoid activation function; the gated fusion unit dynamically adjusts the contribution weights of the two branches according to the input features.

[0029] Beneficial effects:

[0030] (1) High-precision dynamic positioning: By tightly coupling Beidou and inertial navigation to solve the problem, multipath noise is suppressed, and high-precision output of the three-dimensional coordinates of the monitoring point is achieved.

[0031] (2) Realistic real-time rendering: A 3D visualization engine is built based on the WebGPU graphics interface, which supports high frame rate smooth rendering and low latency updates for ultra-long pipeline scenes.

[0032] (3) Intelligent deformation prediction: A graph neural network that integrates physical information and attention mechanism is used to achieve high-precision short-term deformation prediction, which significantly improves the prediction effect compared with traditional methods.

[0033] (4) Physical constraint geometric reconstruction: Combining positioning data and material mechanics model, a smooth deformation surface that conforms to physical laws is generated based on position dynamics method.

[0034] (5) High-security data interaction: A data desensitization and encryption transmission channel is established based on the national cryptographic algorithm. Reversible desensitization of coordinates is achieved through grid scrambling to ensure the security of sensitive data.

[0035] (6) Strong generalization and scalability: Physically enhanced graph convolution and attention mechanism enable the model to adapt to different pipeline types, materials and environmental conditions, and support unified management of multiple types of pipelines. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of the method for visualizing and updating dynamic data of submarine pipelines that integrates BeiDou positioning, as described in this invention.

[0037] Figure 2 This is a diagram of the WebGPU-based 3D visualization rendering engine architecture in this invention.

[0038] Figure 3 This is a schematic diagram of the structure of the two-stage graph neural network that integrates physical information and attention mechanism in this invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to specific embodiments.

[0040] This embodiment provides a method for visualizing and updating dynamic data of submarine pipelines by integrating BeiDou positioning, which is specifically applied to the construction of a digital twin system for oil and gas pipelines in a certain sea area.

[0041] Implementation, for example Figure 1 As shown, a method for visualizing and updating dynamic data of submarine pipelines using BeiDou positioning includes the following steps:

[0042] Step S1: Construct a marine digital twin data architecture consisting of a perception layer, a data layer, a fusion layer, a model layer, and an application layer. The perception layer integrates BeiDou positioning terminals, pressure sensors, temperature sensors, and sonar detection equipment deployed on the subsea pipeline to collect pipeline location coordinate sequences, environmental parameters, and seabed topography data in real time. The data layer standardizes, cleans, and encrypts multi-source heterogeneous data for storage. The fusion layer uses an adaptive Kalman filter algorithm to tightly couple BeiDou positioning data and inertial navigation data to obtain the dynamic spatial position of the pipeline with centimeter-level accuracy. The model layer uses a two-stage graph neural network that integrates physical information and attention mechanisms as its core carrier. Through three core components—physically enhanced graph convolutional layers, multi-head adaptive graph attention mechanisms, and parallel spatiotemporal feature fusion modules—it provides accurate deformation prediction data support for visualization updates. The application layer relies on a lightweight 3D visualization rendering engine built on WebGPU. It receives centimeter-level high-precision real-time positioning data from the fusion layer and deformation prediction data from the model layer as its core driver, ultimately realizing the dynamic visualization and interaction of the spatial position, operating status, and structural morphology of the subsea pipeline, adapting to the visualization needs of ultra-long-distance pipeline scenarios.

[0043] Step S2: Build a 3D visualization rendering engine based on the WebGPU graphics interface. Implement parallel decoding and vertex transformation of the digital elevation model of the seabed topography and the 3D model of the pipeline through the computation shader. Use an improved ray casting algorithm to perform volume rendering of the scalar field of the ocean water body to generate a real ocean environment scene containing ocean current gradient, temperature field and pressure field.

[0044] The improved ray casting algorithm employs a volume rendering strategy that combines adaptive step size control with early ray termination: it dynamically adjusts the ray sampling step size according to the seabed topography and pipeline spatial distribution, performs fine sampling of pipeline areas and coarse sampling of blank areas; and terminates sampling in advance when the cumulative ray opacity reaches the target, reducing redundant calculations and achieving efficient and realistic volume rendering of the ocean water scalar field.

[0045] Step S3: Construct a two-stage graph neural network that integrates physical information and attention mechanism as the pipeline deformation prediction model. It includes a physically enhanced graph convolutional layer combined with a four-head adaptive graph attention mechanism and a parallel spatiotemporal feature fusion module. The historical location sequence obtained by the Beidou positioning terminal and the environmental load data collected by the sensor are used as input features. The graph convolutional layer is used to capture the spatial correlation features between each monitoring point of the pipeline. Finally, the temporal evolution features are extracted through a gated recurrent unit to output the predicted deformation displacement and stress distribution cloud map of the pipeline at future time.

[0046] Step S4: Inject the high-precision pipeline positioning data calculated in real time in step S1 and the predicted deformation data generated in step S3 into the 3D visualization rendering engine to drive the real-time update of the geometric mesh vertices of the pipeline model and realize the visualization mapping of the dynamic deformation process; at the same time, overlay the seabed image data collected by multibeam sonar with the digital elevation model to generate realistic seabed surface texture.

[0047] Step S5: Establish a data security desensitization and encryption transmission channel based on SM4. By dynamically deflecting and remapping the geographic coordinates of classified pipelines, ensure that the visualization terminal does not display the original precise coordinate data; use the SM9 identifier cryptography algorithm to encrypt the sensor data stream in the transmission link end-to-end to achieve secure interaction of sensitive data.

[0048] Specifically, the adaptive Kalman filter algorithm used in step S1 includes a noise adaptive estimation module based on information covariance matching. This module adjusts the filter gain matrix in real time to suppress positioning noise caused by multipath effects and sea state changes, and the output stability is better than that of differential positioning results of single-point positioning.

[0049] Specifically, in step S2, the volume rendering based on WebGPU adopts an adaptive step size control and an early ray termination strategy. The sampling step size is dynamically adjusted according to the seabed topography and pipeline spatial distribution. A fine sampling step size is used for regions of interest containing pipelines, and a coarse sampling step size is used for water areas without targets to reduce computational overhead.

[0050] Specifically, the dual-stage graph neural network constructed in step S3, which integrates physical information and attention mechanism, includes three layers of physically enhanced graph convolutional layers, a four-head adaptive graph attention mechanism, and two layers of parallel spatiotemporal feature fusion modules. The graph convolutional layers capture the spatial correlation features between pipeline nodes by normalizing the Laplacian matrix with mechanical coupling weights. The bidirectional gated recurrent unit extracts the forward and backward correlation features of the time series of each monitoring point. Finally, the multidimensional deformation prediction results are output through the fully connected layer.

[0051] Specifically, in step S4, the pipeline model's geometric mesh vertex update adopts a deformation constraint solution method based on position dynamics. The BeiDou positioning data is used as the position constraint term, and the material mechanics constitutive model is used as the energy constraint term. Through iterative solution, a smooth deformation surface that satisfies physical laws is generated.

[0052] Specifically, in step S5, the dynamic deflection processing adopts a grid scrambling algorithm based on geofencing. While maintaining the topological connectivity of the pipeline network, the same pseudo-random deflection amount is applied to the coordinate points falling into the same grid cell to achieve reversible spatial location desensitization.

[0053] During the deployment phase of the sensing layer, a monitoring node integrating a BeiDou-3 satellite positioning terminal and a microelectromechanical inertial measurement unit is deployed every 500 meters along the pipeline. The positioning terminal supports BeiDou-3 global short message communication and precise point positioning services. Fiber optic pressure sensors and distributed temperature sensors are deployed at key nodes of the pipeline, with a sampling frequency set to 10 Hz. An underwater autonomous vehicle carrying a multibeam echo sounder is used to collect seabed topographic data around the pipeline, with a resolution better than 0.5 meters. The raw data streams output by all sensing devices are aggregated through an underwater junction box and then transmitted back to the shore-based data center in real time via a 5G communication network.

[0054] During the data fusion processing stage, the data center receives carrier phase observations from the BeiDou positioning terminal and angular velocity and acceleration data from the inertial measurement unit, and performs tightly coupled decomposition using an adaptive Kalman filter algorithm. The core of this algorithm lies in constructing a 21-dimensional state vector, specifically defined as follows:

[0055] X = [δp n , δp e , δp u , δv n , δv e , δv u , φ, θ, ψ, ε x , ε γ , ε_z, ▽ ax , ▽ aγ ,▽ a_ z, ε' x , ε' γ , ε'_z, ▽' ax , ▽' aγ , ▽' a_ z] T

[0056] in:

[0057] δp n , δp e , δp u These represent the positional errors in the north, east, and celestial directions, respectively.

[0058] δv n , δv e , δv u These represent the velocity errors in the north, east, and sky directions, respectively.

[0059] φ, θ, ψ: respectively, roll, pitch, and yaw attitude errors;

[0060] ε x , ε γ ε_z: Gyroscope zero bias error;

[0061] ▽ ax , ▽ aγ , ▽ a_ z: Accelerometer zero bias error;

[0062] ε' x , ε' γ ε'_z: gyroscope random walk error;

[0063] ▽' ax , ▽' aγ , ▽' a_ z: Accelerometer random walk error.

[0064] The aforementioned 21-dimensional state vector fully encompasses five error terms: position, velocity, attitude, sensor bias, and random walk. It can completely describe the dynamic error characteristics of the tightly coupled BeiDou / INS system, ensuring the convergence and stability of the filtering algorithm under complex sea conditions.

[0065] The observation equation integrates pseudorange differential observations and carrier phase double-difference observations. To address abrupt changes in observation noise caused by multipath effects on the sea surface, the algorithm introduces an adaptive factor based on innovation covariance matching, setting a relative deviation threshold of 3σ (σ being the theoretical standard deviation of BeiDou positioning observation noise) between the innovation sequence covariance and the theoretical covariance. When the deviation between the innovation sequence covariance and the theoretical covariance exceeds the 3σ threshold, the measurement noise covariance matrix is ​​dynamically adjusted. This threshold is determined using the 3σ criterion commonly used in satellite navigation and adaptive Kalman filtering, corresponding to the 99.7% confidence interval boundary of the normal distribution. It distinguishes between normal observation fluctuations and abnormal disturbances caused by multipath effects and sudden changes in sea state, avoiding false adjustments triggered by normal observation fluctuations while ensuring rapid response to abnormal noise, thus guaranteeing the convergence and stability of the filtering algorithm under complex sea conditions. The calculated 3D coordinate accuracy of the pipeline monitoring points is better than 5 cm, providing a high-precision position reference for subsequent visualization updates.

[0066] Implementation, for example Figure 2As shown, in the 3D rendering engine construction phase, a WebGPU-based graphics architecture is built. WebGPU device objects and command encoding queues are created, and vertex buffers, index buffers, and unified buffer resources are allocated. The pipeline 3D model adopts a cylindrical segmented modeling method, with each pipeline segment consisting of a geometric mesh of 64 vertices. Vertex attributes include position coordinates, normal direction, and texture coordinates. The seabed topography digital elevation model data is uploaded to the GPU memory in floating-point texture format, and the topography geometry is dynamically generated through the heightmap sampling function in the vertex shader. The ocean water scalar field data adopts a regular mesh data structure, including temperature, salinity, pressure, and ocean current velocity vector fields, and is rendered using an adaptive ray casting algorithm executed by the computation shader. The rendering pipeline adopts a deferred shading architecture, generating depth textures, normal textures, and base color textures during the geometry processing phase, and uniformly calculating the contributions of directional light, point light sources, and ambient occlusion during the lighting processing phase. For large-scale pipeline network rendering, the engine implements frustum-based instantiation culling and detail level control, setting distance thresholds: 0-500m is the near distance range centered on the viewpoint, 500-2000m is the medium distance range, and greater than 2000m is the far distance range. Pipeline segments in the near distance range (0-500m) are loaded with complete geometric details and high-resolution textures, the medium distance range (500-2000m) uses medium-precision geometry and standard-resolution textures, and the far distance range (>2000m) uses simplified geometry and low-resolution textures. This significantly reduces rendering computational overhead while ensuring visual fidelity, adapting to the high frame rate real-time rendering requirements of ultra-long-distance underwater pipeline scenes.

[0067] Implementation, for example Figure 3 As shown, in the intelligent deformation prediction stage, a two-stage graph neural network integrating physical information and an attention mechanism is constructed. This network is based on a traditional temporal graph convolutional network architecture, and its performance is improved through three core innovations: physically enhanced graph convolutional layers, a four-head adaptive graph attention mechanism, and a parallel spatiotemporal feature fusion module.

[0068] The pipeline network topology is abstracted as an undirected graph structure, where monitoring nodes correspond to graph vertices and pipeline segments between adjacent nodes correspond to graph edges. The model input features consist of three parts: a location feature matrix composed of the BeiDou positioning coordinate sequence of each monitoring point over the past 60 seconds, an environmental feature matrix composed of time-series fluid pressure data collected by pressure sensors, and a temperature feature matrix composed of time-series pipe wall temperature data collected by temperature sensors.

[0069] Based on the aforementioned infrastructure, this invention proposes the following three core improvements:

[0070] (1) Physically enhanced graph convolutional layer

[0071] This invention proposes a physically enhanced graph convolutional layer that introduces material mechanics constraints during graph convolution. In the adjacency matrix construction stage, adjacency relationships are defined based on pipeline topology connections, and the mechanical coupling coefficient is used as a weighting factor for the adjacency matrix. The formula for calculating the mechanical coupling coefficient is:

[0072]

[0073] Where E is the elastic modulus of the pipe, I is the moment of inertia of the section, and L... ij This is the pipe segment length between adjacent monitoring nodes i and j. This coefficient characterizes the mechanical sensitivity of deformation transmission between adjacent nodes.

[0074] (2) Four-head adaptive graph attention mechanism

[0075] This invention introduces a four-head adaptive graph attention mechanism after the physically enhanced graph convolutional layer. This mechanism calculates the attention coefficients between nodes and dynamically adjusts the information aggregation weights.

[0076] The attention coefficient is calculated by taking into account both node feature similarity and physical distance.

[0077]

[0078] Where h i with h j W represents the feature vectors of node i and node j, respectively. q With W k It is a linear transformation matrix. (d ij ) is based on physical distance d ij Radial basis function encoding, where || denotes vector concatenation operation. This is a trainable attention vector. The attention coefficients are normalized using the softmax function to obtain the final aggregated weights. The four-head mechanism concatenates or averages the features output by multiple independent attention heads, enhancing the model's ability to represent different feature subspaces.

[0079] (3) Spatiotemporal feature fusion module

[0080] Traditional temporal graph convolutional networks stack graph convolutional layers and recurrent neural networks in a series manner, resulting in the separation of spatial feature extraction and temporal feature extraction processes. This invention designs a parallel spatiotemporal feature fusion module to simultaneously capture spatiotemporal correlation features.

[0081] The parallel spatiotemporal feature fusion module contains two parallel branches:

[0082] Spatial-first branch: First, spatial features are extracted through physically enhanced graph convolutional layers, and then temporal features are extracted by inputting bidirectional gated recurrent units.

[0083] Time-first branch: First, the independent temporal features of each node are extracted through a bidirectional gated recurrent unit, and then spatial information is exchanged through a physically enhanced graph convolutional layer.

[0084] The output features of the two branches are adaptively weighted and fused through a gated fusion unit:

[0085]

[0086] Where F spatial For space-first branch output, F temporal For time-priority branch output, W z Let be a trainable weight matrix, and σ be the sigmoid activation function. The gated fusion unit dynamically adjusts the contribution weights of the two branches based on the input features.

[0087] The model training uses historical monitoring data to construct a training set. The loss function includes a root mean square error term, a structural similarity index term, and a physical consistency loss term. The physical consistency loss term includes deformation compatibility loss and force balance loss, ensuring that the prediction results meet numerical accuracy and maintain reasonable spatial distribution, while also conforming to the basic laws of continuum mechanics.

[0088] During the dynamic visualization update phase, high-precision positioning data calculated in real time and deformation data output by the prediction model are injected into the rendering engine. Each frame update of the rendering engine first parses the BeiDou positioning data stream to obtain the real spatial coordinates of each monitoring point at the current moment; secondly, it reads the displacement vector output by the deformation prediction model. This displacement vector is the three-dimensional spatial offset output by the model, in meters, with a value range determined by the mechanical properties of the pipeline material and the boundary conditions of the environmental load. A single offset does not exceed 0.5 meters. The predicted offset is applied to the coordinates of the monitoring points to obtain an optimized spatial position that balances real-time positioning and short-term deformation prediction. Then, a constraint-based solution method based on position dynamics is used to update the spatial position of all vertices of the pipeline mesh. The position dynamics solver uses BeiDou positioning data as rigid position constraints, transforms the elastic modulus and moment of inertia of the pipeline material into bending and tensile constraints, and solves for the optimal vertex position that satisfies all constraints through Gauss-Seidel iteration. The converged mesh data is written to the vertex buffer, triggering the next frame rendering. The seabed surface rendering module uses seabed image data collected by multibeam sonar as a diffuse texture, overlays it with a normal map generated by a digital elevation model, and calculates the final pixel color through the lighting model in the fragment shader.

[0089] During the data security interaction phase, a secure transmission and desensitized display module based on national cryptographic algorithms is deployed. Before 3D visualization rendering of the original BeiDou positioning coordinates, a grid scrambling process is first performed: the target sea area is divided into a geographic grid with a side length of 100 meters, and the same pseudo-random offset is applied to all coordinate points within each grid. The offset is generated by using the grid's unique index and the daily updated SM4 key as input, encrypting the data using the SM4 block cipher algorithm to obtain a fixed-length encrypted string, extracting the high 32 bits and low 32 bits of the encrypted string and converting them into longitude and latitude offset components, respectively, and then normalizing them to generate a planar offset. This is then superimposed with a fixed elevation offset to form a 3D pseudo-random offset. The offset is generated by the SM4 algorithm based on the grid index and the daily key. The scrambled coordinates maintain the relative topology of the pipeline network, but the absolute position offset is between 50 meters and 100 meters, effectively preventing the leakage of precise coordinates.

[0090] The sensor data stream is encrypted using the SM9 identifier cryptography algorithm before transmission, and the receiving end decrypts and restores the original data using the identifier private key. The WebSocket communication link between the browser and the server is encrypted using a transport layer security protocol, and the key negotiation process embeds the SM2 elliptic curve key exchange algorithm.

Claims

1. A method for visualizing and updating dynamic data of subsea pipelines using BeiDou positioning, characterized in that, Includes the following steps: Step S1: Construct a marine digital twin data architecture consisting of a perception layer, a data layer, a fusion layer, a model layer, and an application layer; The perception layer integrates BeiDou positioning terminals, pressure sensors, temperature sensors, and sonar detection equipment deployed along the subsea pipeline to collect pipeline position coordinate sequences, environmental parameters, and seabed topography data in real time. The data layer standardizes, cleans, and encrypts multi-source heterogeneous data for storage. The fusion layer uses an adaptive Kalman filter algorithm to tightly couple BeiDou positioning data and inertial navigation data to obtain the pipeline's dynamic spatial position with centimeter-level accuracy. The model layer uses a two-stage graph neural network that integrates physical information and an attention mechanism as its core carrier. Through three core components—physically enhanced graph convolutional layers, multi-head adaptive graph attention mechanisms, and parallel spatiotemporal feature fusion modules—it provides accurate deformation prediction data support for visualization updates. The application layer relies on a lightweight 3D visualization rendering engine built on WebGPU. It receives centimeter-level high-precision real-time positioning data from the fusion layer and deformation prediction data from the model layer as the core driver, and finally realizes dynamic visualization and interaction of the spatial location, operation status and structural morphology of the subsea pipeline, adapting to the visualization needs of ultra-long-distance pipeline scenarios. Step S2: Build a 3D visualization rendering engine based on the WebGPU graphics interface. Implement parallel decoding and vertex transformation of the digital elevation model of the seabed topography and the 3D model of the pipeline through the computation shader. Use an improved ray casting algorithm to perform volume rendering of the scalar field of the ocean water body to generate a real ocean environment scene containing ocean current gradient, temperature field and pressure field. The improved ray casting algorithm adopts a volume rendering strategy that combines adaptive step size control with early ray termination: dynamically adjusts the ray sampling step size according to the seabed topography and pipeline spatial distribution, and performs fine sampling of pipeline areas and coarse sampling of blank areas. And sampling is terminated early when the cumulative light opacity reaches the standard, reducing redundant calculations and achieving efficient and realistic volume rendering of the scalar field of ocean water; Step S3: Construct a two-stage graph neural network that integrates physical information and attention mechanism as a pipeline deformation prediction model, which includes a physically enhanced graph convolutional layer combined with a four-head adaptive graph attention mechanism and a parallel spatiotemporal feature fusion module. The historical location sequence obtained by the Beidou positioning terminal and the environmental load data collected by the sensor are used as input features. The graph convolutional layer is used to capture the spatial correlation features between each monitoring point of the pipeline. Finally, the temporal evolution features are extracted through the gated recurrent unit, and the predicted deformation displacement and stress distribution cloud map of the pipeline at future time are output. Step S4: Inject the high-precision pipeline positioning data calculated in real time in step S1 and the predicted deformation data generated in step S3 into the 3D visualization rendering engine to drive the real-time update of the geometric mesh vertices of the pipeline model and realize the visualization mapping of the dynamic deformation process; at the same time, overlay the seabed image data collected by multibeam sonar with the digital elevation model to generate realistic seabed surface texture. Step S5: Establish a data security desensitization and encryption transmission channel based on SM4. By dynamically deflecting and remapping the geographic coordinates of classified pipelines, ensure that the visualization terminal does not display the original precise coordinate data; use the SM9 identifier cryptography algorithm to encrypt the sensor data stream in the transmission link end-to-end to achieve secure interaction of sensitive data.

2. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning as described in claim 1, characterized in that, The adaptive Kalman filter algorithm used in the fusion layer in step S1 includes a noise adaptive estimation module based on information covariance matching. This module adjusts the filter gain matrix in real time to suppress positioning noise caused by multipath effects and sea state changes, and outputs a differential positioning result with better stability than single-point positioning.

3. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning as described in claim 1, characterized in that, In step S2, the volume rendering based on WebGPU adopts an adaptive step size control and an early ray termination strategy. The sampling step size is dynamically adjusted according to the seabed topography and pipeline spatial distribution. A fine sampling step size is used for regions of interest containing pipelines, and a coarse sampling step size is used for water areas without targets to reduce computational overhead.

4. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 1, characterized in that, The dual-stage graph neural network constructed in step S3, which integrates physical information and attention mechanism, includes three layers of physically enhanced graph convolutional layers, a four-head adaptive graph attention mechanism, and two layers of parallel spatiotemporal feature fusion modules. The graph convolutional layers capture the spatial correlation features between pipeline nodes by normalizing the Laplacian matrix with mechanical coupling weights. The bidirectional gated recurrent unit extracts the forward and backward correlation features of the time series of each monitoring point. Finally, the multidimensional deformation prediction results are output through the fully connected layer.

5. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 1, characterized in that, In step S4, the pipeline model's geometric mesh vertex update adopts a deformation constraint solution method based on position dynamics. BeiDou positioning data is used as the position constraint term, and the material mechanics constitutive model is used as the energy constraint term. A smooth deformation surface that satisfies physical laws is generated through iterative solution.

6. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 1, characterized in that, In step S5, the dynamic deflection processing adopts a grid scrambling algorithm based on geofencing. While maintaining the topological connectivity of the pipeline network, the same pseudo-random deflection amount is applied to the coordinate points falling into the same grid cell to achieve reversible spatial location desensitization.

7. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 4, characterized in that, The physically enhanced graph convolutional layer introduces material mechanics constraints during the graph convolution process; During the adjacency matrix construction phase, adjacency relationships are defined based on pipeline topology connections, and the mechanical coupling coefficient is used as a weighting factor for the adjacency matrix. The formula for calculating the mechanical coupling coefficient is: Where E is the elastic modulus of the pipe, I is the moment of inertia of the section, and L... ij The length of the pipe segment between adjacent monitoring nodes i and j; this coefficient characterizes the mechanical sensitivity of deformation transmission between adjacent nodes.

8. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 4, characterized in that, An adaptive graph attention mechanism is introduced after the physically enhanced graph convolutional layer; the adaptive graph attention mechanism calculates the attention coefficients between nodes and dynamically adjusts the information aggregation weights. The attention coefficient is calculated by taking into account both node feature similarity and physical distance. Where h i with h j W represents the feature vectors of node i and node j, respectively. q With W k It is a linear transformation matrix. (d ij ) is based on physical distance d ij Radial basis function encoding, where || denotes vector concatenation operation. The attention vector is trained; the attention coefficients are normalized using the softmax function to obtain the final aggregate weights. This mechanism concatenates or averages the features output by multiple independent attention heads, enhancing the model's ability to express different feature subspaces.

9. The method for visualizing and updating dynamic data of submarine pipelines integrating BeiDou positioning according to claim 4, characterized in that, The parallel spatiotemporal feature fusion module includes a space-first branch and a time-first branch; The spatial priority branch first extracts spatial features through a physically enhanced graph convolutional layer, and then inputs them into a bidirectional gated recurrent unit to extract temporal features; The time-first branch first extracts the independent temporal features of each node through a bidirectional gated recurrent unit, and then performs spatial information interaction through a physically enhanced graph convolutional layer. The output features of the two branches are adaptively weighted and fused through a gated fusion unit: Where F spatial For space-first branch output, F temporal For time-priority branch output, W z σ is a trainable weight matrix, and σ is the sigmoid activation function; the gated fusion unit dynamically adjusts the contribution weights of the two branches according to the input features.

Images