A non-buried submarine pipeline identification method and system based on multi-beam point cloud

Abstract

The present invention relates to the field of ocean exploration technology, and in particular to a method and system for identifying non-buried submarine pipelines based on multi-beam point clouds. The method comprises acquiring three-dimensional point cloud data of a submarine pipeline area and preprocessing the acquired point cloud data; constructing a neighborhood based on the preprocessed point cloud data, setting a masked attention mechanism according to the neighborhood structure, performing geometric encoding and feature enhancement using the output of the masked attention mechanism, decoding the enhanced features, restoring the feature representation of the complete point cloud, and calculating and verifying the span height based on the segmentation results. The present invention adopts a decoding strategy combining inverse distance weighted interpolation with jump connections to accurately propagate the enhanced features to the original point cloud, and fitting the seabed reference surface through inverse distance weighted interpolation to quickly locate high-risk span areas.

Description

Technical Field

[0001] The present invention relates to the field of ocean detection technology, and in particular to a method and system for identifying non-buried submarine pipelines based on multi-beam point clouds. Background Art

[0002] As a key infrastructure for marine oil and gas transportation, the spatial distribution state, overhang and exposure of submarine pipelines directly affect the long-term stability and operational safety of the pipelines. Under the long-term influence of ocean currents, the stress of the soil near the seabed gradually changes, which triggers erosion and scouring, causing the sediment covering the submarine pipeline to gradually disappear, eventually resulting in the top of the pipeline being exposed on the seabed. As the scouring action continues, the seabed around the submarine pipeline is constantly eroded, and obvious scour pits will form in the bottom area of the pipeline. The continuous deepening and expansion of the scour pits will gradually increase the height difference between the submarine pipeline and the surrounding seabed, eventually causing the bottom of the pipeline to completely separate from the seabed, resulting in overhangs.

[0003] Traditional methods for detecting the state of subsea pipelines have problems such as strong subjectivity, low efficiency, and incomplete coverage, making it difficult to achieve high-precision monitoring of the entire pipeline range. Traditional detection methods mainly include: manual exploration, which is only applicable to pipeline detection in shallow water areas and has great risks; secondly, ROV (Remotely Operated Vehicle) equipped with optical detection equipment Underwater vehicles (unmanned remotely operated vehicles) are limited by environmental factors such as water transparency, waves, and ocean currents, resulting in low detection efficiency and difficulty in achieving efficient and continuous pipeline status monitoring over large sea areas. Furthermore, side-scan sonar detection is significantly affected by beam grazing angle and the complexity of seabed topography, easily creating shadow effects in rugged or high-contrast areas, affecting complete target identification. Furthermore, the sonar images they acquire still require manual interpretation, which is significantly influenced by subjective factors and prone to interpretation errors. Finally, shallow subsurface profilers suffer from low resolution and can only provide cross-sectional data along the survey line, making it difficult to provide complete three-dimensional status information. Although pipeline status detection technology has matured, problems such as low detection efficiency, incomplete detection, and significant subjective factors still exist. Currently, a method and system for identifying non-buried submarine pipelines based on multi-beam point clouds is needed. Summary of the Invention

[0004] In order to solve the problems of low detection efficiency, incomplete detection and significant influence of subjective factors in submarine pipeline identification, the present invention provides a non-buried submarine pipeline identification method and system based on multi-beam point cloud.

[0005] In a first aspect, the present invention provides a method for identifying non-buried submarine pipelines based on multi-beam point clouds, which adopts the following technical solutions:

[0006] A non-buried submarine pipeline identification method based on multi-beam point cloud, comprising:

[0007] Acquire 3D point cloud data of the submarine pipeline area and pre-process the acquired point cloud data;

[0008] Neighborhood construction is performed based on preprocessed point cloud data, including obtaining K key points through random sampling, calculating local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density;

[0009] Setting up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and performing multi-stage normalization on the similarity matrix;

[0010] The output of the masked attention mechanism is used for geometric encoding and feature enhancement, including calculating the neighboring points of each key point through the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the neighboring points;

[0011] Decode the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer;

[0012] The span height is calculated and verified based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the datum surface using inverse distance weighted interpolation based on the seabed points.

[0013] Furthermore, the obtained point cloud data is preprocessed, including merging point cloud data from different sources into a unified time axis, generating a surface from the point cloud data based on roll, pitch, and compass calibration survey lines, deleting noise points using CUBE surface filtering, subtracting the overlapping areas of the water depth surfaces generated by the main survey line and the connecting survey line, deleting excessive noise points, and finally performing Lambert law correction on the original backscatter data to obtain scalar backscatter intensity data.

[0014] Furthermore, the neighborhood is constructed based on the preprocessed point cloud data, and multiple key points are selected from the preprocessed point cloud data using random sampling. The global density mean of the point cloud is calculated, and the sampling probability of random sampling is adjusted according to the global density mean. The local density of each key point is estimated using a Gaussian kernel function, and the neighborhood radius is adjusted according to the local density normalization result. The local density calculation formula is:

[0015] ;

[0016] in, Indicates the total number of neighborhood points involved in the calculation, Represents the kernel function bandwidth, which controls the range of the Gaussian kernel. Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

[0017] Furthermore, the neighborhood radius is adjusted according to the local density normalization result, including constructing a covariance matrix using the k-nearest neighbor point set of the key point, performing eigendecomposition on the covariance matrix to obtain eigenvalues, calculating the curvature value based on the eigenvalues, and then normalizing the local density. The neighborhood radius is calculated using the curvature value and the normalized local density. The neighborhood radius calculation formula is:

[0018] ;

[0019] in, Expressed as the basic radius parameter, Expressed as the normalized local density, Expressed as the global maximum local density value, Represented as key points The curvature value of Expressed as a geometric structure sensitivity factor, it balances the weight of the influence of density and curvature on the neighborhood radius.

[0020] Furthermore, the use of a multilayer perceptron to calculate a neighborhood feature similarity matrix includes obtaining a set of key points and a set of neighborhood radiuses, screening neighborhood points based on the neighborhood radius and key points, and calculating a feature similarity matrix using the screened neighborhood points and key points. The calculation formula of the feature similarity matrix is:

[0021] ;

[0022] in, represents the neighborhood radius, and Represent the point cloud semantic features of point i and point j respectively, Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, and Represents a multilayer perceptron network with one-dimensional convolutional layer, normalization layer and GELU activation function layer, which is used to map two input point clouds to a multilayer perceptron of specified dimension. Represents a feature concatenation operation.

[0023] Furthermore, the multi-stage normalization of the similarity matrix includes using the Softmax function to normalize the similarity matrix. The first stage normalization is performed to make the sum of the attention weights of each key point on the neighboring points 1. The matrix after Softmax normalization is square root transformed and normalized again in the second stage. By suppressing extreme weights, the importance distribution of points in the neighborhood is balanced. Finally, the matrix after the second stage normalization is normalized in the column direction. The first stage normalization formula is:

[0024] ;

[0025] Among them, i represents the key point, N represents the total number of neighborhood points, and j and k represent different neighborhood points respectively.

[0026] Furthermore, the output results of the masked attention mechanism are used for geometric encoding and feature enhancement, including screening K neighborhood points of each key point based on the masked attention weight matrix, calculating the relative coordinates of the key point and its neighboring points, and generating relative position encoding by nonlinearly mapping the relative coordinates. The key point coordinates, neighbor point coordinates, relative coordinates and Euclidean distance are spliced into relative position features, and then the semantic features of the neighboring points and the key points are processed by dimensionality increase and the difference is calculated to obtain relative semantic features. The relative position features are fused with the relative semantic features, and feature extraction and aggregation are performed through a convolutional neural network to obtain high-dimensional enhanced point cloud features. The relative semantic feature calculation formula is:

[0027] ;

[0028] in, Represent the normal vector features of the point cloud, represents the backscattering intensity characteristics of the point cloud, is the semantic category of the point cloud.

[0029] Furthermore, the decoding operation of the enhanced features includes feature propagation based on the inverse distance weighted average interpolation method, propagating the enhanced features from the sampling points to all points of the original point cloud, jump-joining the features from the encoding stage with the features obtained by interpolation to obtain the initial features, constructing the UnitPointNet module with the initial features as input, and iteratively updating using the Unit PointNet module to obtain the complete point cloud feature representation, wherein the UnitPointNet module consists of a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a maximum pooling layer. The feature propagation calculation formula is:

[0030] ;

[0031] in, represents the weighting coefficient, Represents key points After enhancement, It is represented as the feature vector obtained by propagating the original point through weighted averaging, and m represents the number of selected neighboring key points.

[0032] Furthermore, the extraction of seabed points from the non-pipeline point cloud and fitting of the datum plane using inverse distance weighted interpolation based on the seabed points include extracting a seabed point cloud data set based on the complete point cloud feature representation, fitting the datum plane to the seabed point cloud based on the inverse distance weighted interpolation algorithm, and then calculating the vertical distance from the suspended pipeline point cloud data to the fitted seabed datum plane. The calculation formula of the inverse distance weighted interpolation is:

[0033] ;

[0034] in, is the estimated value of the point to be interpolated, is the measured value of the known point i, is the distance from the known point i to the interpolation point, is a weight index used to control the influence of distance, and n is the number of neighboring known points used for interpolation calculation.

[0035] In the second aspect, a non-buried submarine pipeline identification system based on multi-beam point cloud includes:

[0036] The data acquisition module is configured to acquire three-dimensional point cloud data of the submarine pipeline area and pre-process the acquired point cloud data;

[0037] The neighborhood module is configured to: construct a neighborhood based on the preprocessed point cloud data, including obtaining K key points using random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density;

[0038] The attention module is configured to: set up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and perform multi-stage normalization on the similarity matrix;

[0039] The feature module is configured to use the output of the masked attention mechanism for geometric encoding and feature enhancement, including calculating the nearest neighbors of each key point using the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the nearest neighbors.

[0040] The decoding module is configured to: decode the enhanced features and restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer;

[0041] The output module is configured to calculate and verify the span height based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the reference surface based on the seabed points using inverse distance weighted interpolation.

[0042] In summary, the present invention has the following beneficial technical effects:

[0043] 1. This method dynamically adjusts the neighborhood radius through random sampling combined with kernel density estimation. Compared with traditional fixed-radius neighborhood construction methods, it can adaptively adjust the neighborhood range based on the local density and geometric curvature of the point cloud. In scenarios with large density variations, such as bends in submarine pipelines and areas with complex terrain, this method can avoid the problems of noise introduced by too large a neighborhood or feature loss due to too small a neighborhood, thereby more accurately capturing the local geometric and semantic features of the pipeline and improving the integrity and accuracy of point cloud feature expression.

[0044] 2. The present invention uses a multi-layer perceptron to calculate the neighborhood feature similarity matrix and optimizes the attention weights through multi-stage normalization, enabling the model to focus on key neighborhood points related to the pipeline. Compared with the traditional feature aggregation method based on Euclidean distance, this mechanism effectively integrates the semantic information of the point cloud (such as backscatter intensity and normal vector) and the geometric structure. In the identification of submarine pipelines and similar background areas such as sediments and rocks, it significantly improves the model's sensitivity to pipeline boundaries and material differences, thereby reducing the missegmentation rate.

[0045] 3. This invention effectively preserves the spatial structure and semantic attributes of point clouds by jointly encoding relative position features and relative semantic features, combined with convolutional neural networks for feature extraction and aggregation. In submarine pipeline identification tasks, this method can accurately distinguish subtle geometric differences between pipelines and their surroundings (such as raised structures and suspended forms). It also leverages semantic features such as normal vectors and backscatter intensity to enhance recognition of pipelines of different materials, improving segmentation accuracy and model robustness.

[0046] 4. This invention adopts a decoding strategy that combines inverse distance weighted interpolation with skip connections to accurately propagate enhanced features to the original point cloud. The feature representation is iteratively updated through the Unit PointNet module, which ensures feature integrity while reducing information loss. This enables the model to output more refined pipeline segmentation results. Based on the segmentation results, the seabed datum is fitted through inverse distance weighted interpolation, which can quickly locate and quantitatively determine the distribution and scale of high-risk overhanging areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure is a schematic diagram of the overall process of a method for identifying non-buried submarine pipelines based on multi-beam point clouds according to an embodiment of the present invention.

[0048] Figure 2 This is a data preprocessing flow chart of a method for identifying non-buried submarine pipelines based on multi-beam point clouds according to an embodiment of the present invention.

[0049] Figure 3This is a structural diagram of a masked attention algorithm for a non-buried submarine pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention.

[0050] Figure 4 This is a diagram of the algorithm structure of a point cloud segmentation network coding layer of a non-buried submarine pipeline identification method based on multi-beam point clouds in an embodiment of the present invention.

[0051] Figure 5 3 is a randomly sampled point cloud density distribution diagram of a non-buried submarine pipeline identification method based on a multi-beam point cloud according to an embodiment of the present invention; wherein (a) is the original point cloud data; and (b) is the point cloud data after random sampling. DETAILED DESCRIPTION

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] Example 1

[0054] Reference Figure 1 In this embodiment, a method for identifying non-buried submarine pipelines based on multi-beam point clouds includes:

[0055] S1. Acquire three-dimensional point cloud data of the submarine pipeline area and pre-process the acquired point cloud data;

[0056] S2. Neighborhood construction is performed based on the preprocessed point cloud data, including obtaining K key points using random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density;

[0057] S3. Setting up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and performing multi-stage normalization on the similarity matrix;

[0058] S4. Use the output of the masked attention mechanism for geometric encoding and feature enhancement, including calculating the neighboring points of each key point through the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the neighboring points;

[0059] S5. Decode the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer;

[0060] S6. Calculate and verify the span height based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the datum surface using inverse distance weighted interpolation based on the seabed points.

[0061] Specifically, a non-buried submarine pipeline identification method based on multi-beam point cloud includes the following contents:

[0062] S1. Acquire three-dimensional point cloud data of the submarine pipeline area and pre-process the acquired point cloud data;

[0063] like Figure 1 、 Figure 2 As shown in the figure, in the 3D point cloud segmentation problem, point cloud classification is a prerequisite for segmentation. Actual multibeam bathymetric data has the following characteristics: ① The point cloud data volume is large; ② The seabed point cloud data accounts for a large proportion, while the pipeline point cloud data only accounts for a small portion; ③ It contains a large amount of water, sidelobes, and water noise caused by marine organisms. Therefore, multibeam point cloud data needs to be preprocessed before input into model training.

[0064] Import the collected original multi-beam data *.pds file into the newly created HIPS project file, create and set the parameters of the ship type, and enter the TPU parameters and relative positions of each sensor for subsequent filtering operations. Merge the time-based data into space-based data, that is, the water depth data, time-based attitude data, navigation data, and heading data in the multi-beam sonar coordinate system collected based on time are merged into the space-based water depth data under the same time axis. Attitude and navigation data correction: set the attitude data deviation value according to the surface generated by the roll, pitch, and compass calibration lines laid out during the measurement process to avoid water depth errors caused by instrument installation parameters, browse and edit navigation data such as heading and speed, check for abnormal points, and delete and smooth them.

[0065] Tidal changes have a significant impact on depth measurements at depths less than 200 meters. To avoid this impact, the tidal data from the multi-beam measurement period needs to be imported into the software for tidal correction. Using CARIS interpolation, the overlapping areas of the depth surfaces generated by the main and interconnecting survey lines are subtracted. The number of difference points exceeding 1% of the average water depth is counted, and excessive noise points are deleted. The original backscatter data is corrected for Lambert's law, propagation loss, terrain, beam footprint area, and time-varying gain, ultimately obtaining scalar backscatter intensity data related only to the physical properties of the seabed.

[0066] Export the preprocessed depth data to ASCII format, typically with a *.xyz or *.txt extension. Export the raw HIPS data for each pipeline to a *.txt file without weighted depth gridding or thinning to prevent the point cloud of spanning submarine pipelines from being omitted or smoothed due to weighted averaging and thinning. The file contains five columns of data: coordinate information (x, y, z), depth information, and backscatter intensity information. The coordinate information is projected.

[0067] After normalization preprocessing of the point cloud data, the seabed point cloud, exposed pipeline point cloud, and suspended pipeline point cloud were manually annotated using CloudCompare point cloud processing software. The specific steps are as follows:

[0068] (1) Open the point cloud file: Since the projected coordinate information is too large for the CloudCompare software, the original data needs to be translated and scaled, the center of the point cloud data needs to be redefined, and the relative coordinates of other point clouds relative to the data center need to be calculated.

[0069] (2) Point cloud preprocessing: ① Point cloud denoising. The accuracy of point cloud data has a direct impact on the model. By rotating the point cloud to check for undeleted noise data, the noise data can be deleted to improve data accuracy. ② Point cloud rendering: The exported original data has no color information. It is extremely difficult to distinguish between the seabed terrain point cloud and the submarine pipeline point cloud during the annotation process. The elevation information (water depth information) is selected to render the point cloud. ③ Point cloud cropping: Deep learning learns the features of point cloud data, which is to learn the spatial position and the relationship between adjacent points of the entire point cloud. However, the volume of the submarine point cloud is huge, and the submarine pipeline point cloud only accounts for a very small part. Learning the entire point cloud data may not accurately learn the features, which affects the segmentation accuracy of the model. Therefore, the multi-beam point cloud is segmented into small areas along the routing direction. On the one hand, it can reduce the data volume of each sample and enhance the feature learning ability of the model; on the other hand, it can also increase the number of samples to improve the learning ability.

[0070] (3) Point cloud annotation: The most important but most tedious operation in deep learning is to annotate the point cloud with semantic category labels. A large amount of manual visual interpretation is required to annotate the information. Based on the presence of the acoustic transmission area of the pipeline in the multi-beam sonar image and the side-scan sonar image, it is judged whether the submarine pipeline is suspended or exposed, and the length of the shadow area is measured to calculate the suspended height. At the same time, the point cloud is marked with the corresponding semantic category at the same coordinate position of the three-dimensional point cloud. By manually circling the submarine pipeline point cloud and the submarine point cloud, the point cloud is segmented and the Scalar Fields attribute is added. In this paper, the submarine point cloud is set to 0, the suspended submarine pipeline point cloud is set to 1, and the exposed point cloud is set to 2. The format of the point cloud is (x, y, z, reflection intensity, category attribute).

[0071] The normal vector feature is the component perpendicular to the local surface of each point cloud, reflecting the curvature, smoothness and other characteristics of the point cloud surface. Introducing normal vector information helps to distinguish geometric shapes. Suspended submarine pipelines have cylindrical characteristics, and exposed submarine pipelines have semicircular characteristics. Therefore, calculating the normal vector information of the point cloud is conducive to improving the model's ability to recognize pipelines. The normal vector calculation tool in the CloudCompare software is used to calculate the normal vector. The point cloud format after calculation is (x, y, z, , reflection intensity, category attribute), where, They represent the normal vector features of point clouds in different directions respectively, and finally merge the three types of annotated point cloud data into a whole point cloud.

[0072] Export the annotated point cloud data and save it in ASCII code format.

[0073] S2. Neighborhood construction is performed based on the preprocessed point cloud data, including obtaining K key points using random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density;

[0074] like Figure 5 As shown, representative key points are selected from the preprocessed point cloud data to avoid the high computational cost of traditional farthest point sampling and improve the probability of retaining key points in low-density areas (such as pipelines). In this embodiment, the global density mean of the point cloud is first calculated based on the preprocessed point cloud data. , as a benchmark for judging high and low density areas: ,in, The initial density of all points is estimated by Gaussian kernel density, with bandwidth h = 0.1. The global density mean of the point cloud is used as the threshold to judge low-density points and high-density points. ( ), the random sampling probability is increased to:

[0075] ;

[0076] in, is the correction factor, is the initial density of all points, is the global density mean of the point cloud, N is the total number of sampling points, Expressed as the basic sampling probability term, in traditional uniform random sampling, the probability of each point being selected is the inverse of the total number of points, that is, the probability of all points being selected is equal. represents the set of key points selected from the preprocessed point cloud by random sampling, i.e. , k is the number of key points, , for high-density points, that is , the sampling probability remains uniformly random: ,K key points are obtained based on random sampling of low-density and high-density points, and the proportion of key points in low-density areas such as pipelines is increased through density correction to avoid feature omission.

[0077] Local density is calculated based on the acquired key points. In 3D point cloud data processing, point density is often uneven, especially in complex environments with high-density areas (such as flat seabeds) and low-density areas (such as pipeline point clouds or noise points). Traditional K-nearest neighbor clustering algorithms require a larger neighborhood range in low-density areas to cover the same number of points, while smaller neighborhood ranges are applied in high-density areas. This may result in similar characteristics and unrecognizable local feature differences in the neighborhood, leading to local information loss. To this end, this application proposes a density expansion ball query method that enables the query neighborhood to be adaptively adjusted according to the density of the local point cloud. A smaller neighborhood range is applied to low-density point cloud areas, while a larger neighborhood range should be set for high-density point cloud areas to expand the variance of feature differences within the area to capture more contextual information.

[0078] In order to avoid the high computational cost of directly calculating the number of neighbor points, this application uses kernel density estimation and uses the Gaussian kernel function to calculate the local density:

[0079] ;

[0080] in, Indicates the total number of neighborhood points involved in the calculation, Represents the bandwidth of the adaptive kernel function, which controls the range of the Gaussian kernel. Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

[0081] For kernel function bandwidth , dynamically adjusted according to the local density, when When it is a high-density area, the kernel function bandwidth Expressed as:

[0082] ;

[0083] in, represents the optimization factor, Expressed as the global maximum initial density, ensuring that the bandwidth of high-density areas is reduced, is the original bandwidth, when When it is a low-density area, By adaptively adapting the kernel function bandwidth, the bandwidth is reduced in high-density areas to reduce the smoothing effect and avoid over-aggregation of adjacent seabed terrain point clouds. In low-density areas, the bandwidth is kept wide to capture the global distribution of sparse points and improve the stability of density estimation.

[0084] Dynamically adjust the neighborhood radius according to the local density and geometric curvature, so that the neighborhood can adapt to the density and structural complexity of the point cloud at the same time. , obtain the neighborhood point set through the k-nearest neighbor algorithm In this embodiment, k=20 is taken and the covariance matrix is constructed based on the obtained neighborhood point set:

[0085] ;

[0086] in, is the neighborhood point mean, Indicates the selection of key points The set of k nearest neighbor points of Represents the three-dimensional coordinates of the j-th point in the neighborhood, performs eigenvalue decomposition on the covariance matrix, and obtains the eigenvalue , the curvature value is defined as the proportion of the minimum eigenvalue: The curvature value range is ([0, 1]). A larger value indicates a more complex local geometric structure, such as sharp bends at the edge of a pipe or a span. The neighborhood radius is calculated by combining the normalized density and curvature. The neighborhood radius calculation formula is:

[0087] ;

[0088] in, Expressed as the basic radius parameter, Expressed as the normalized local density, Expressed as the global maximum local density value, Represented as key points The curvature value of It is expressed as a geometric structure sensitivity factor, balancing the influence weights of density and curvature on the neighborhood radius. Based on the calculation formula of the neighborhood radius, the neighborhood radius automatically increases in high-density areas and remains small in low-density areas, forming a directed graph with a variable neighborhood. It builds a more reasonable connection relationship in the point cloud and enhances the ability to disseminate information. It can obtain richer contextual information, thereby accelerating the flow of remote information and improving the perception of global features.

[0089] S3. Setting up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and performing multi-stage normalization on the similarity matrix;

[0090] like Figure 3As shown, the traditional ball query method treats all neighboring points equally when constructing a local neighborhood, that is, all points within the search ball are included in the calculation, regardless of their actual contribution to the task. This method works well in evenly distributed point cloud scenarios, but in areas with large density variations (exposed, suspended pipe point clouds, and transition areas between submarine terrain point clouds), it may lead to incorrect feature aggregation and affect segmentation accuracy. In order to enhance the model's ability to perceive local irregularities, a masked attention mechanism is proposed that enables it to adaptively select valid neighboring points and assign different importance weights. Different from the conventional attention mechanism, the traditional Transformer encoder uses a padding mask to process sequences of different lengths to ensure consistent input format, and the decoder uses a sequence mask to prevent the omission of true values. The masked attention mechanism used in this application is set in the encoder, specifically targeting the neighborhood relationship of point cloud data, for local feature learning.

[0091] According to the key point set obtained in step S2 Neighborhood radius corresponding to each key point , for each key point , filter neighborhood points by Euclidean distance calculation:

[0092] ;

[0093] in, Indicates a point With neighboring points If the distance is less than , then determine belong Neighborhood point set , corresponding to the matrix position Set to 1, otherwise set to 0. This process forms a binary mask matrix by judging the spatial position relationship point by point , where k represents the number of key points and N represents the total number of points in the point cloud. Based on the point cloud annotation in step S1, for each point , the semantic features With three-dimensional coordinates Concatenate into input feature vector:

[0094] ;

[0095] Among them, d represents the semantic feature dimension, R represents the dimensional space, represents the semantic features of point j, Represents the three-dimensional coordinates of the j-th point. After splicing, the multi-layer perceptron is used to construct a feature similarity matrix for the point cloud of the neighborhood. The calculation formula of the feature similarity matrix is:

[0096] ;

[0097] in, represents the neighborhood radius, and Represent the point cloud semantic features of point i and point j respectively, Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, and Represents a multilayer perceptron network with one-dimensional convolutional layer, normalization layer and GELU activation function layer, which is used to map two input point clouds to a multilayer perceptron of specified dimension. represents the feature concatenation operation, and The structures are the same, consisting of a one-dimensional convolutional layer (Conv1D) + a batch normalization layer (BN1D) + a GELU activation function. The calculation formula of the GELU activation function is:

[0098] ;

[0099] in, Represents the basic linear term, providing linear scaling so that the output range is positively correlated with the input, while compressing the overall function value to about half of the input value. represents the hyperbolic tangent function term, represents the cubic term, introduces a slight nonlinear curvature, fits the tail characteristics of the Gaussian distribution, and the constant part is the empirically optimized value.

[0100] Through two independent MLP networks and Encode the splicing features, output the feature vector of the specified dimension, and use MLP to learn the feature space and The potential and alignment relationship between them, in order to identify both semantic information and geometric information, the point cloud features and point cloud coordinates are encoded simultaneously. In the actual training process, -∞ represents the points with minimal correlation with the neighborhood. In order to ensure that these points will not be given non-zero attention weights in the subsequent Softmax process, resulting in numerical instability, they are uniformly set to 10 9 , It reflects the cosine similarity between key points and neighborhood points in the feature space. The larger the value, the more similar the semantic or geometric features of the two are.

[0101] Then, multi-stage normalization is used to suppress extreme weights, balance the importance of points in the neighborhood, and avoid a single neighborhood point dominating feature aggregation. First, the first stage normalization is performed, specifically Softmax global normalization:

[0102] ;

[0103] Among them, i represents the key point, N represents the total number of neighborhood points, j and k represent different neighborhood points, Represents the feature similarity matrix, normalizes the attention weight of each key point i to the neighboring point j into a probability distribution, and the non-neighboring point After the exponential operation, it approaches 0 and is automatically masked (weight is 0), realizing hard screening of the neighborhood range.

[0104] Then, the square root of each element of the Softmax normalized weight matrix is taken to suppress excessive weights (for example, the square root of a weight of 0.8 is 0.89, and the square root of a weight of 0.2 is 0.45, and the gap is narrowed). The square root renormalization formula is:

[0105] ;

[0106] in, represents the weight after Softmax normalization in the first stage, Represents the number of neighborhood points of the i-th key point. The weight distribution is smoothed through the second stage normalization to avoid individual neighborhood points from obtaining too high weights due to high feature similarity. The contribution of each point in the local neighborhood is balanced. Finally, the norm normalization is performed. The specific formula is:

[0107] ;

[0108] Among them, K represents the total number of key points generated by random sampling, Represents the weight after the second stage square root renormalization, column by column (each neighborhood point j For all key points i weights) to ensure that the neighborhood points j The sum of the attention weights at different key points is 1, which further balances the weight distribution across key points and prevents a point from dominating the global features due to being activated with high weights in the neighborhood of multiple key points.

[0109] S4. Use the output of the masked attention mechanism for geometric encoding and feature enhancement, including calculating the neighboring points of each key point through the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the neighboring points;

[0110] like Figure 4 As shown, for each key point ,according to The weight values are sorted from high to low, and the K neighborhood points with the largest weights are selected. The neighborhood points with higher weights have higher feature similarity with the key points and make greater contributions to pipeline identification. They are retained first to obtain each key point. The corresponding set of K neighboring points , and its corresponding attention weight , for each neighbor point , calculation and key points The relative coordinates of: And Euclidean distance:

[0111] ;

[0112] in, It is expressed as the three-dimensional coordinates of the kth nearest neighbor of the key point, that is, ,in addition, 、 and Indicates key points The three-dimensional coordinate value of the key point coordinate , neighboring point coordinates , relative coordinates and Euclidean distance Splicing into relative position features:

[0113] ;

[0114] in, Indicates concatenating symbols by dimension. Represents the Euclidean distance, and the relative position features are nonlinearly transformed through the multi-layer perceptron (MLP) to generate relative position encoding:

[0115] ;

[0116] in, Represented as relative position feature, Represented as a one-dimensional convolutional layer, represents the batch normalization layer, Represents the activation function layer, and the final output dimension is 128 dimensions, which enhances the expressiveness of position features. Then, relative semantic feature calculation is performed. First, the semantic feature dimension is increased. The semantic features of key points and neighboring points are increased in dimension through a one-dimensional convolution layer, and the difference between the increased-dimensional semantic features of neighboring points and key points is calculated:

[0117] ;

[0118] in, Represents the original semantic features of key points, Represents the original semantic features of the neighboring points. The semantic features are relative semantic features that combine the normal vector, backscatter intensity (1 dimension), and semantic category (1 dimension). The relative semantic feature calculation formula is:

[0119] ;

[0120] in, Represent the normal vector features of point clouds in different directions, represents the backscattering intensity characteristics of the point cloud, is the semantic category of the point cloud.

[0121] Finally, the relative position encoding is concatenated with the relative semantic features (256 dimensions) to form a high-dimensional fusion feature. The fusion features of K neighboring points are aggregated through a two-dimensional convolutional layer (Conv2D) to extract high-order features of the local area. The maximum value of each dimension is retained through the maximum pooling operation, and the feature dimension is further improved by the multi-layer perceptron to output the final enhanced point cloud features.

[0122] S5. Decode the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer;

[0123] The decoding layer is mainly used to propagate the global features gradually extracted during the encoding process from the sparse point cloud back to the original point cloud and restore the feature representation of the complete point cloud. Because in the segmentation task, it is necessary to predict a label for each input point, and the downsampling in the encoding process leads to a reduction in the number of points, and the information needs to be propagated back to each original point through the decoding layer.

[0124] For each original point , find the nearest neighbor key point in the key point set S through the K nearest neighbor algorithm , calculate the original point Euclidean distance to each neighboring key point , and calculate the weighting coefficient based on the inverse distance:

[0125] ;

[0126] in, Indicates the original point and neighboring key points The Euclidean distance, m represents the number of neighboring key points selected by the K nearest neighbor algorithm. The closer the distance, the higher the weight of the key points. Expressed as a weight index, the default is 2, Indicates the original point and neighboring key points Euclidean distance between of In the decoding stage, when the enhanced features of the key points are propagated to the original point cloud, Adjust the contribution of neighboring points to avoid noise introduced by distant points, while enhancing the consistency of local features, ensuring local consistency of feature propagation, and propagating the enhanced features of key points to the original points through weighted averaging:

[0127] ;

[0128] in, represents the weighting coefficient, Indicates key points After enhancement, The feature vector is expressed as the weighted average of the original points and then propagated. It is used for subsequent segmentation or span height calculation. m represents the number of neighboring key points selected by the K-nearest neighbor algorithm. Ultimately, the feature matrix of all original points is generated. Compared with traditional nearest neighbor interpolation or uniformly weighted interpolation, this method avoids feature blurring caused by discrete jumps and global averaging through inverse distance weighting and K-nearest neighbor search. This method can more accurately preserve detailed features, especially in high-density change areas such as pipe edges. During the encoding process, intermediate features at each level are stored. After interpolation, the features from the encoding stage are jump-joined with the interpolated features (with the same number of feature points) to preserve rich local information:

[0129] ;

[0130] in, represents channel dimension splicing, Represents the feature matrix propagated from key point enhancement features to the original point cloud through inverse distance weighted interpolation in the decoding stage, Represented as the encoding phase Low-level intermediate features with the same spatial resolution.

[0131] This ensures that the spatial distribution and other geometric information of the original point cloud of low-level features can be preserved, and the semantic information of high-level features can be passed to the low-resolution point cloud. The spliced features are input into a Unit PointNet module, which consists of a one-dimensional convolution layer, a one-dimensional batch normalization layer, a ReLU (Rectified Linear Unit) activation function layer activation function, and a maximum pooling layer. The feature vector of each point is updated to enhance local geometry and global semantic information. This process is repeated until the features are propagated to the original point set. The expression of the Unit PointNet module is:

[0132] ;

[0133] in, Represented as a maximum pooling operation, Expressed as a linear rectifier activation function, Represented as a one-dimensional batch normalization layer, Represented as a one-dimensional convolutional layer, represents the initial features obtained after jump splicing. The batch normalization layer normalizes the convolution output to stabilize the feature distribution. The formula is:

[0134] ;

[0135] in, and is the batch statistic, x is the input feature, and As a learnable parameter, the inverse distance weighted average avoids the discrete jump of neighbor interpolation and makes the pipeline edge feature transition natural.

[0136] S6. Calculate and verify the span height based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the datum surface using inverse distance weighted interpolation based on the seabed points.

[0137] Based on the three-dimensional point cloud segmentation model used in the embodiment, that is, based on the three-dimensional point cloud segmentation model composed of the encoding layer in step S4 and the decoding layer in step S5, the suspended pipeline point cloud is segmented to accurately identify the suspended pipeline point cloud (label 1) from the massive point cloud and distinguish it from the seabed point cloud (label 0) and other ground feature point clouds. The segmented suspended pipeline point cloud and seabed point cloud are used for subsequent reference plane fitting and span height calculation, respectively. After the segmentation is completed, the suspended pipeline point cloud and the seabed point cloud are exported separately, and a plane fitting is performed on the seabed point cloud to construct a reference reference plane. Subsequently, the distance between the suspended pipeline point cloud and the fitted seabed plane is calculated in the vertical direction to accurately obtain span height information. Inverse distance weighting is applicable to surface generation of point cloud data. Its principle is that the closer the known points are to the interpolation point, the greater their influence on it, while the farther away the points are, the smaller their influence. IDW calculates the value of the target interpolation point by assigning a weight to each known point. The weight depends on the distance between the point and the interpolation point. The calculation formula of the inverse distance weighted interpolation is:

[0138] ;

[0139] in, is the estimated value of the point to be interpolated, is the measured value of the known point i, is the distance from the known point i to the interpolation point, is a weight index used to control the influence of distance, and n is the number of neighboring known points used for interpolation calculation.

[0140] The IDW interpolation method is a deterministic interpolation technique based on distance weighting. Its calculation process does not rely on complex probabilistic models or statistical analysis, resulting in low computational overhead. It is particularly suitable for the rapid interpolation of large-scale point cloud data. In flat seabed areas, where the bathymetric point cloud is relatively evenly distributed, IDW can better capture topographic variations and achieve accurate fitting. In scour trenches, IDW uses elevation-weighted interpolation of seabed points on both sides to maintain topographic continuity and avoid bias caused by missing data in global models (such as least squares). Furthermore, IDW has strong noise detection capabilities, preventing abnormal fluctuations around subsea pipelines caused by fragmented data in obscured areas.

[0141] The span height of a pipeline refers to the vertical distance between the submarine pipeline and the seabed reference plane. The core of the calculation is to calculate the spatial distance between the pipeline point cloud and the seabed reference plane. Based on the seabed reference plane fitted by inverse distance weighting (IDW), the calculation of the span height can be converted into the calculation of each suspended pipeline point cloud ( ) to the fitting surface S(x, y):

[0142] ;

[0143] in, is the span height of the submarine pipeline, The water depth of the segmented spanning pipeline point cloud is represented by D, and the diameter of the pipeline is represented by D. Since the multi-beam acoustic waves first contact the top of the submarine pipeline, the span height is calculated by subtracting the diameter D of the pipeline from the distance the pipeline point cloud is above the fitted surface. The calculated span height is assigned to the point cloud attribute of the suspended pipeline. Subsequently, the pipeline point cloud data containing the span height information is loaded into the ArcGIS platform, and color rendering is performed based on the span height attribute to intuitively display the spatial suspension characteristics of the pipeline.

[0144] Example 2

[0145] This embodiment differs from Embodiment 1 in that it provides a non-buried submarine pipeline identification system based on a multi-beam point cloud, including:

[0146] The data acquisition module is configured to acquire three-dimensional point cloud data of the submarine pipeline area and pre-process the acquired point cloud data;

[0147] The neighborhood module is configured to: construct a neighborhood based on the preprocessed point cloud data, including obtaining K key points using random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density;

[0148] The attention module is configured to: set up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and perform multi-stage normalization on the similarity matrix;

[0149] The feature module is configured to use the output of the masked attention mechanism for geometric encoding and feature enhancement, including calculating the nearest neighbors of each key point using the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the nearest neighbors.

[0150] The decoding module is configured to: decode the enhanced features and restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer;

[0151] The output module is configured to calculate and verify the span height based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the reference surface based on the seabed points using inverse distance weighted interpolation.

[0152] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying non-buried submarine pipelines based on multi-beam point clouds, characterized in that: include: Acquire 3D point cloud data of the submarine pipeline area and pre-process the acquired point cloud data; Neighborhood construction is performed based on preprocessed point cloud data, including obtaining K key points through random sampling, calculating local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; Setting up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and performing multi-stage normalization on the similarity matrix; The output of the masked attention mechanism is used for geometric encoding and feature enhancement, including calculating the neighboring points of each key point through the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the neighboring points; Decode the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer; The span height is calculated and verified based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the datum surface using inverse distance weighted interpolation based on the seabed points.

2. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that: The obtained point cloud data is preprocessed, including merging point cloud data from different sources into a unified time axis, using CUBE surface filtering to delete noise points from the surface generated by the point cloud data based on roll, pitch and compass calibration survey lines, subtracting the overlapping areas of the water depth surfaces generated by the main survey line and the connecting survey line, deleting excessive noise points, and finally performing Lambert law correction on the original backscatter data to obtain scalar backscatter intensity data.

3. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that: The neighborhood is constructed based on the preprocessed point cloud data. Multiple key points are selected from the preprocessed point cloud data using random sampling. The global density mean of the point cloud is calculated, and the sampling probability of random sampling is adjusted according to the global density mean. The local density of each key point is estimated using the Gaussian kernel function. The neighborhood radius is adjusted according to the local density normalization result. The local density calculation formula is: ； Among them, among them, Indicates the total number of neighborhood points involved in the calculation, Represents the kernel function bandwidth, which controls the range of the Gaussian kernel. Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

4. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 3, characterized in that: The neighborhood radius is adjusted according to the local density normalization result, including constructing a covariance matrix using the k-nearest neighbor point set of the key point, performing eigendecomposition on the covariance matrix to obtain eigenvalues, calculating the curvature value based on the eigenvalues, and then normalizing the local density. The neighborhood radius is calculated using the curvature value and the normalized local density. The neighborhood radius calculation formula is: ； in, Expressed as the basic radius parameter, Expressed as the normalized local density, Expressed as the global maximum local density value, Represented as key points The curvature value of Expressed as a geometric structure sensitivity factor, it is used to balance the influence of density and curvature on the neighborhood radius.

5. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that: The method of calculating the neighborhood feature similarity matrix using a multi-layer perceptron includes obtaining a set of key points and a set of neighborhood radius, screening the neighborhood points based on the neighborhood radius and the key points, and calculating the feature similarity matrix using the screened neighborhood points and the key points. The calculation formula of the feature similarity matrix is: ； in, represents the neighborhood radius, and Represent the point cloud semantic features of point i and point j respectively, Represents the three-dimensional coordinates of the key points, represents the three-dimensional coordinates of the j-th point in the neighborhood, and Represents two multi-layer perceptron networks consisting of a one-dimensional convolutional layer, a normalization layer, and a GELU activation function layer, which are used to map two input point clouds to a multi-layer perceptron of a specified dimension. Represents a feature concatenation operation.

6. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 5, characterized in that: The multi-stage normalization of the similarity matrix includes using the Softmax function to normalize the similarity matrix. The first stage normalization is performed to make the sum of the attention weights of each key point on the neighboring points 1. The matrix after Softmax normalization is square root transformed and normalized again in the second stage. By suppressing extreme weights, the importance distribution of points in the neighborhood is balanced. Finally, the matrix after the second stage normalization is normalized in the column direction. The first stage normalization formula is: ； Among them, i represents the key point, N represents the total number of neighborhood points, and j and k represent different neighborhood points respectively.

7. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that: The output results of the masked attention mechanism are used for geometric encoding and feature enhancement, including screening K neighborhood points of each key point based on the masked attention weight matrix, calculating the relative coordinates of the key point and its neighboring points, and generating relative position encoding by nonlinear mapping of the relative coordinates. The key point coordinates, neighbor point coordinates, relative coordinates and Euclidean distance are spliced into relative position features. After that, the semantic features of the neighboring points and the key points are processed by dimension increase and the difference is calculated to obtain relative semantic features. The relative position features are fused with the relative semantic features, and feature extraction and aggregation are performed through a convolutional neural network to obtain high-dimensional enhanced point cloud features. The relative semantic feature calculation formula is: ； in, Represent the normal vector features of the point cloud, represents the backscattering intensity characteristics of the point cloud, is the semantic category of the point cloud.

8. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 7, characterized in that: The decoding operation of the enhanced features includes feature propagation based on the inverse distance weighted average interpolation method, propagating the enhanced features from the sampling points to all points of the original point cloud, jump-joining the features from the encoding stage with the features obtained by interpolation to obtain the initial features, constructing the Unit PointNet module with the initial features as input, and iteratively updating the Unit PointNet module to obtain the complete point cloud feature representation. The Unit PointNet module consists of a convolution layer, a batch normalization layer, a ReLU activation function layer, and a maximum pooling layer. The feature propagation calculation formula is: ； in, represents the weighting coefficient, Represents key points After enhancement, It is represented as the feature vector obtained by propagating the original point through weighted averaging, and m represents the number of selected neighboring key points.

9. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 8, characterized in that: The method extracts seabed points from the non-pipeline point cloud and fits the datum surface using inverse distance weighted interpolation based on the seabed points, including extracting a seabed point cloud data set based on the complete point cloud feature representation, fitting the datum surface to the seabed point cloud based on the inverse distance weighted interpolation algorithm, and then calculating the vertical distance from the suspended pipeline point cloud data to the fitted seabed datum surface. The calculation formula of the inverse distance weighted interpolation is: ； in, is the estimated value of the point to be interpolated, is the measured value of the known point i, is the distance from the known point i to the interpolation point, is a weight index used to control the influence of distance, and n is the number of neighboring known points used for interpolation calculation.

10. A non-buried submarine pipeline identification system based on multi-beam point cloud, executing the method according to claim 1, characterized in that: include: The data acquisition module is configured to acquire three-dimensional point cloud data of the submarine pipeline area and pre-process the acquired point cloud data; The neighborhood module is configured to: construct a neighborhood based on the preprocessed point cloud data, including obtaining K key points using random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; The attention module is configured to: set up a masked attention mechanism based on the neighborhood structure, including using a multi-layer perceptron to calculate the neighborhood feature similarity matrix and perform multi-stage normalization on the similarity matrix; The feature module is configured to use the output of the masked attention mechanism for geometric encoding and feature enhancement, including calculating the nearest neighbors of each key point using the K-nearest neighbor algorithm, and performing relative feature calculation and feature fusion based on the nearest neighbors. The decoding module is configured to: decode the enhanced features and restore the feature representation of the complete point cloud, including propagating the features to the original point cloud, and output the segmentation results by the Unit PointNet layer; The output module is configured to calculate and verify the span height based on the segmentation results, including extracting seabed points from the non-pipeline point cloud and fitting the reference surface based on the seabed points using inverse distance weighted interpolation.

Images