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

Abstract

The invention relates to the technical field of ocean exploration, in particular to a non-buried submarine pipeline identification method and system based on multi-beam point clouds, and the method comprises the steps: obtaining three-dimensional point cloud data of a submarine pipeline region, and carrying out the preprocessing of the obtained point cloud data; performing neighborhood construction based on the preprocessed point cloud data, setting a mask attention mechanism according to a neighborhood structure, performing geometric coding and feature enhancement by using an output result of the mask attention mechanism, performing decoding operation on the enhanced features, recovering feature representation of the complete point cloud, and performing suspended span height calculation and verification based on a segmentation result. According to the method, a decoding strategy combining reverse distance weighted interpolation and jump connection is adopted, enhanced features are accurately propagated to the original point cloud, the seabed datum plane is fitted through the reverse distance weighted interpolation, and the high-risk suspended cross area can be rapidly positioned.

Description

Technical Field

[0001] The present invention relates to the technical field of marine exploration, and particularly 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 occurrence state, sag and exposure of submarine pipelines directly affect the long-term stability and operation safety of the pipelines. Under the long-term action of ocean currents scouring, the stress of the soil near the seabed gradually changes, resulting in erosion scouring. As a result, the sediment covering the submarine pipeline gradually erodes, and finally the top of the pipeline is exposed above the seabed. With the continuous progress of scouring, the seabed around the submarine pipeline is continuously scoured and eroded, and obvious scouring pits will be formed in the area at the bottom of the pipeline. The continuous deepening and expansion of the scouring pits gradually increase the height difference between the submarine pipeline and the surrounding seabed, and finally lead to the complete detachment of the bottom of the pipeline from the seabed, resulting in sag.

[0003] Traditional methods for detecting the occurrence state of submarine 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 probing is only applicable to pipeline detection in shallow water areas and has great risks. Secondly, the ROV (Remotely Operated Underwater Vehicle) equipped with optical detection equipment is limited by environmental factors such as water transparency, waves and ocean currents, and its detection efficiency is low, making it difficult to achieve efficient and continuous monitoring of the pipeline occurrence state in a large area of the sea. In addition, side-scan sonar detection is greatly affected by the beam grazing angle and the complexity of the seabed topography, and it is easy to produce shadow effects in rugged or high-contrast areas, affecting the complete identification of the target object. At the same time, the sonar images obtained still need to rely on manual interpretation, which is greatly affected by subjective factors and is prone to interpretation errors. Finally, the shallow stratum profiler has the problem of low resolution and can only provide cross-sectional data along the survey line, making it difficult to provide complete three-dimensional occurrence state information. Although the technology for detecting the occurrence state of submarine pipelines has become increasingly mature, there are still problems such as low detection efficiency, incomplete detection, and great influence of subjective factors. At present, a method and system for identifying non-buried submarine pipelines based on multi-beam point clouds are needed. Summary of the Invention

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

[0005] In a first aspect, a method for identifying non-buried submarine pipelines based on multi-beam point clouds provided by the present invention adopts the following technical solution: A method for identifying non-buried submarine pipelines based on multi-beam point clouds, comprising: Obtain three-dimensional point cloud data of the submarine pipeline area, and preprocess the obtained point cloud data; Construct neighborhoods based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; Set a masked attention mechanism according to the neighborhood structure, including calculating the neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; Use the output result 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; Perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting the segmentation result by the Unit PointNet layer; Calculate and verify the sag height based on the segmentation result, including extracting seabed points from non-pipeline point clouds and fitting a reference plane using inverse distance weighted interpolation according to the seabed points.

[0006] Further, the preprocessing of the obtained point cloud data includes merging point cloud data from different sources to 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, performing a difference operation on the overlapping regions of the water depth surfaces generated by the main survey line and the connecting survey line, deleting out-of-limit noise points, and finally correcting the original backscatter data using Lambert's law to obtain scalar backscatter intensity data.

[0007] Further, in the neighborhood construction based on the preprocessed point cloud data, multiple key points are selected from the preprocessed point cloud data by 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. For each key point, its local density 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: ; Wherein, represents the total number of neighboring points participating in the calculation, represents the kernel bandwidth, which controls the range of action of the Gaussian kernel, represents the three-dimensional coordinates of the key point, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

[0008] Further, adjusting the neighborhood radius according to the local density normalization result includes constructing a covariance matrix using the k-nearest neighbor point set of the key point, performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues, calculating the curvature value based on the eigenvalues, then performing normalization processing on the local density, and calculating the neighborhood radius using the curvature value and the normalized local density. The formula for the neighborhood radius is: ; where, represents the base radius parameter, represents the normalized local density, represents the global maximum local density value, represents the key point curvature value of, represents the geometric structure sensitivity factor, which balances the influence weights of density and curvature on the neighborhood radius.

[0009] Further, calculating the neighborhood feature similarity matrix using the multi-layer perceptron includes obtaining the key point set and the neighborhood radius set, 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 formula for the feature similarity matrix is: ; where, represents the neighborhood radius, and respectively represent the point cloud semantic features of point i and point j, represents the three-dimensional coordinates of the key point, represents the three-dimensional coordinates of the j-th point in the neighborhood, and represent the multi-layer perceptron network of the one-dimensional convolutional layer, the normalization layer and the GELU activation function layer, which is used to map two input point clouds to a multi-layer perceptron of a specified dimension, represents the feature concatenation operation.

[0010] Further, performing multi-stage normalization on the similarity matrix includes performing the first-stage normalization on the similarity matrix using the Softmax function, which is used to make the sum of the attention weights of each key point to the neighborhood points equal to 1, performing a square root transformation on the matrix after Softmax normalization and then performing the second-stage normalization again, balancing the importance distribution of the points in the neighborhood by suppressing the extreme weights, and finally performing column-wise norm normalization on the matrix after the second-stage normalization. The formula for the first-stage normalization is: ; where, i represents the key point, N represents the total number of neighborhood points, and j and k respectively represent different neighborhood points.

[0011] Furthermore, using the output result of the masked attention mechanism for geometric encoding and feature enhancement includes screening K neighboring points of each key point based on the masked attention weight matrix, calculating the relative coordinates between the key point and its neighboring points, and generating relative position encoding by non-linearly mapping the relative coordinates. Concatenating the key point coordinates, neighboring point coordinates, relative coordinates, and Euclidean distance to form relative position features. Then, after dimensionality increasing processing of the semantic features of the neighboring points and the key point, calculating the difference to obtain relative semantic features. Fusing the relative position features and the relative semantic features, and performing feature extraction and aggregation through a convolutional neural network to obtain high-dimensional enhanced point cloud features. The formula for the relative semantic features is: ; Wherein, respectively represent the normal vector features of the point cloud, represents the backscattering intensity feature of the point cloud, is the semantic category of the point cloud.

[0012] Furthermore, decoding the enhanced features includes performing feature propagation based on the inverse distance weighted average interpolation method, propagating the enhanced features from the sampled points to all points of the original point cloud, performing skip connection splicing on the features from the encoding stage and the interpolated features to obtain initial features, constructing a UnitPointNet module with the initial features as the input, and using the Unit PointNet module for iterative update to obtain the complete point cloud feature representation. Among them, the UnitPointNet module consists of a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer. The formula for the feature propagation is: ; Wherein, represents the weighting coefficient, represents the key point enhanced features, represents the feature vector propagated by weighted average of the original points, and m represents the number of selected neighboring key points.

[0013] Furthermore, extracting seabed points from the non-pipeline point cloud and fitting a reference plane using the inverse distance weighted interpolation based on the seabed points includes extracting the seabed point cloud data set from the complete point cloud feature representation, performing reference plane fitting on 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 reference plane. The formula for the inverse distance weighted interpolation is: ; Wherein, 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 the weight exponent used to control the influence degree of the distance, and n is the number of neighboring known points used for interpolation calculation.

[0014] In a second aspect, a non-buried submarine pipeline recognition system based on multi-beam point cloud includes: A data acquisition module configured to acquire three-dimensional point cloud data of the submarine pipeline area and preprocess the acquired point cloud data; A neighborhood module configured to: construct a neighborhood based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; An attention module configured to: set a masked attention mechanism according to the neighborhood structure, including calculating a neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; A feature module configured to: perform geometric encoding and feature enhancement using the output result of the masked attention mechanism, including calculating the neighboring points of each key point by the K-nearest neighbor algorithm and performing relative feature calculation and feature fusion based on the neighboring points; A decoding module configured to: perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting a segmentation result by the Unit PointNet layer; An output module configured to: calculate and verify the spanning height based on the segmentation result, including extracting seabed points from the non-pipeline point cloud and fitting a reference plane using inverse distance weighted interpolation according to the seabed points.

[0015] In summary, the present invention has the following beneficial technical effects: 1. By combining random sampling with kernel density estimation to dynamically adjust the neighborhood radius, the present invention can adaptively adjust the neighborhood range according to the local density and geometric curvature of the point cloud compared with the traditional fixed-radius neighborhood construction method. In scenarios with large density changes such as the bending of submarine pipelines and complex terrain areas, this method can avoid the problems of introducing noise due to too large a neighborhood or losing features due to too small a neighborhood, so as to more accurately capture the local geometry and semantic features of the pipeline and improve the integrity and accuracy of the point cloud feature expression.

[0016] 2. The present invention calculates the neighborhood feature similarity matrix using a multi-layer perceptron and optimizes the attention weights through multi-stage normalization, enabling the model to focus on key neighborhood points related to the pipeline. Compared with traditional feature aggregation methods based on Euclidean distance, this mechanism effectively integrates the semantic information (such as backscattering intensity, normal vector) and geometric structure of the point cloud. In the identification of areas where the seabed pipeline is similar to the background of sediments, rocks, etc., it significantly improves the sensitivity of the model to pipeline boundaries and material differences and reduces the mis-segmentation rate.

[0017] 3. Through the joint encoding of relative position features and relative semantic features, combined with a convolutional neural network for feature extraction and aggregation, the present invention can effectively retain the spatial structure and semantic attributes of the point cloud. In the seabed pipeline identification task, this method can accurately distinguish the subtle geometric differences between the pipeline and the surrounding environment (such as pipeline bulge structures, suspended forms), and at the same time use semantic features such as normal vectors and backscattering intensity to enhance the identification ability of pipelines with different materials, improving the segmentation accuracy and model robustness.

[0018] 4. The present invention adopts a decoding strategy that combines inverse distance weighted interpolation and skip connections to accurately propagate the enhanced features to the original point cloud, and iteratively updates the feature representation through the Unit PointNet module. While ensuring the integrity of the features, it reduces information loss, enabling the model to output a more refined pipeline segmentation result. Based on the segmentation result, the seabed reference plane is fitted by inverse distance weighted interpolation, which can quickly locate and quantitatively determine the distribution and scale of high-risk suspension areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is the overall flow diagram of a non-buried seabed pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention.

[0020] Figure 2 is the data preprocessing flow chart of a non-buried seabed pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention.

[0021] Figure 3 is the mask attention algorithm structure diagram of a non-buried seabed pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention.

[0022] Figure 4 is the algorithm structure diagram of the point cloud segmentation network encoding layer of a non-buried seabed pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention.

[0023] Figure 5 is the random sampling point cloud density distribution diagram of a non-buried seabed pipeline identification method based on multi-beam point cloud according to an embodiment of the present invention; wherein, (a) is the original point cloud data; (b) is the point cloud data after random sampling. DETAILED DESCRIPTION OF THE INVENTION

[0024] The present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Embodiment 1 Referring to Figure 1 , a method for identifying non-buried submarine pipelines based on multi-beam point clouds in this embodiment includes: S1. Obtain three-dimensional point cloud data of the submarine pipeline area and preprocess the obtained point cloud data; S2. Construct neighborhoods based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; S3. Set a masked attention mechanism according to the neighborhood structure, including calculating a neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; S4. Perform geometric encoding and feature enhancement using the output result of the masked attention mechanism, 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; S5. Perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting the segmentation result by the Unit PointNet layer; S6. Calculate and verify the sag height based on the segmentation result, including extracting seabed points from the non-pipeline point cloud and fitting a reference plane using inverse distance weighted interpolation according to the seabed points.

[0026] Specifically, a method for identifying non-buried submarine pipelines based on multi-beam point clouds includes the following: S1. Obtain three-dimensional point cloud data of the submarine pipeline area and preprocess the obtained point cloud data; As Figure 1 , Figure 2 shown, in the problem of three-dimensional point cloud segmentation, the premise of segmenting the point cloud requires classifying the point cloud. The actual multi-beam bathymetric data has the following characteristics: ① a large amount of point cloud data; ② a large proportion of seabed point cloud data, and only a small part of the pipeline point cloud data; ③ containing a large amount of water body, sidelobes, and water body noise caused by marine organisms. Therefore, it is necessary to preprocess the multi-beam point cloud data before inputting it into the model for training.

[0027] 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 input the TPU parameters and relative positions of each sensor for subsequent filtering operations. Combine the time-axis-based data into space-based data, that is, the water depth data in the multi-beam sonar coordinate system collected based on time, the attitude data based on time, navigation data, and heading data, and merge them into the space-based water depth data, attitude, and navigation data correction under the same time axis: Set the deviation value of the attitude data according to the surface generated by the roll, pitch, and compass calibration survey lines arranged during the measurement process to avoid water depth errors caused by instrument installation parameters, etc. Browse and edit the navigation data such as heading and speed, check for the existence of abnormal points, and delete and smooth them.

[0028] The change of tide level has a great influence on the water depth measurement with a water depth less than 200 meters. To avoid its influence, it is necessary to import the tide level data during the multi-beam measurement period into the software for tide level correction. Use CARIS difference to make a difference in the overlapping area of the water depth surfaces generated by the main survey line and the connecting survey line, count the number of difference points exceeding 1% of the average water depth, delete the over-limit noise points, and perform Lambert's law correction, propagation loss correction, terrain correction, beam footprint area correction, time-varying gain correction, etc. on the original backscatter data. Finally, obtain the scalar backscatter intensity data that only relates to the physical properties of the seabed.

[0029] Export the preprocessed water depth data in ASCII format, generally with the suffix *.xyz or *.txt. Export the original HIPS data of each pipeline as a *.txt file without performing weighted water depth gridding and thinning operations to prevent the underwater pipeline point clouds of the suspension span from being omitted and smoothed due to weighted averaging and thinning. The five columns of data saved in the file are coordinate information (x, y, z), water depth information, and backscatter intensity information, and the coordinate information is the projected coordinate information.

[0030] After the normalization preprocessing operation of the point cloud data, manually annotate the seabed point cloud, exposed submarine pipeline point cloud, and suspended submarine pipeline point cloud through the CloudCompare point cloud processing software. The specific steps are as follows: (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, redefine the center of the point cloud data, and calculate the relative coordinates of other point clouds relative to the data center.

[0031] (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 undetected noisy data and deleting it, the data accuracy can be improved. ② Point cloud rendering. The original exported data has no color information, making it extremely difficult to distinguish between seabed terrain point clouds and subsea pipeline point clouds during the annotation process. Select to apply elevation information (water depth information) to render the point cloud. ③ Point cloud cropping. Deep learning learns the features of point cloud data by learning the spatial positions and relationships between adjacent points of the entire point cloud. However, the volume of seabed point clouds is huge, and the subsea pipeline point clouds only account for a very small part. Learning from the entire point cloud data may not accurately learn the features, affecting the segmentation accuracy of the model. Therefore, small regional segmentation of the multibeam point cloud along the route direction can, on the one hand, reduce the data volume of each sample and enhance the model's feature learning ability; on the other hand, it can also increase the number of samples to improve the learning ability.

[0032] (3)Point cloud annotation: The most important but most cumbersome operation in deep learning is to annotate semantic category labels for point clouds. A large amount of manual visual interpretation is required to annotate information. According to the presence of the pipeline acoustic transmission area in the multibeam sonar image and the sidescan sonar image, determine whether the subsea pipeline is span and exposed, measure the length of the shadow area to calculate the span height, and at the same time mark the corresponding semantic categories for the point cloud at the same coordinate position in the three-dimensional point cloud. By manually enclosing the subsea pipeline point cloud and the seabed point cloud, the point cloud is segmented and the Scalar Fields attribute is added. In this paper, the seabed point cloud is set to 0, the suspended subsea 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).

[0033] The normal vector feature is the vertical direction component of each point cloud and the local surface, which reflects the curvature, smoothness and other characteristics of the point cloud surface. Introducing normal vector information helps to distinguish geometric shapes. The suspended subsea pipeline has a cylindrical characteristic, and the exposed subsea pipeline has a semi-circular characteristic. Therefore, calculating the normal vector information of the point cloud is beneficial to improving the model's recognition ability of the pipeline. The calculation of the normal vector is completed through the normal vector calculation tool in the CloudCompare software. After the calculation, the format of the point cloud is (x, y, z, , reflection intensity, category attribute), where, where, respectively represent the normal vector features of point clouds in different directions. Finally, the three types of annotated point cloud data are merged into an overall point cloud.

[0034] Export the annotated point cloud data and save it in ASCII code format.

[0035] S2. Neighborhood construction is carried out based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; As Figure 5 shown, representative key points are selected from the pre - processed point cloud data, avoiding the high computational cost of traditional farthest point sampling, and at the same time improving the retention probability of key points in low - density regions (such as pipelines). In this embodiment, according to the pre - processed point cloud data, the global density mean of the point cloud is first calculated , as the benchmark for judging high - and low - density regions: , where is the initial density of all points, preliminarily 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. For low - density points ([[]] ), the random sampling probability is increased to: ; where is the correction coefficient, is the initial density of all points, is the global density mean of the point cloud, N is the total number of sampling points, represents the basic sampling probability term. In traditional uniform random sampling, the probability of each point being selected is the reciprocal of the total number of points, that is, all points have an equal probability of being selected, represents the set of key points selected from the pre - processed point cloud through random sampling, that is , k is the number of key points, , for high - density points, that is , the sampling probability remains uniform random: . K key points are obtained according to the random sampling of low - density and high - density points, and the proportion of key points in low - density regions such as pipelines is increased through density correction to avoid feature omission.

[0036] Based on the obtained key points, the local density is calculated. In the processing of three - dimensional point cloud data, the density of points is usually non - uniform. Especially in complex environments, there are high - density regions (such as flat seabeds) and low - density regions (such as pipeline point clouds or noise points). Traditional K - nearest neighbor clustering algorithms require a larger neighborhood range in low - density regions to cover the same number of points, while a smaller neighborhood range is applied in high - density regions. This may lead to similar characteristics in the neighborhood region and unidentifiable local feature differences, resulting in the loss of local information. For this reason, this application proposes a density - inflated sphere query method, enabling 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 regions, while for high - density point cloud regions, in order to increase the variance of feature differences within the region, a larger neighborhood range should be set to capture more context information.

[0037] To avoid the high computational cost caused by directly calculating the number of neighbor points, this application adopts kernel density estimation and uses a Gaussian kernel function to calculate the local density: ; where represents the total number of neighborhood points participating in the calculation, represents the bandwidth of the adaptive kernel function, which controls the scope of the Gaussian kernel, represents the three-dimensional coordinates of the key point, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

[0038] Regarding the bandwidth of the kernel function, it is dynamically adjusted according to the local density. When , that is, in the high-density area, the bandwidth is expressed as: ; where represents the optimization factor, represents the global maximum initial density, ensuring that the bandwidth in the high-density area shrinks, is the original bandwidth. When , that is, in the low-density area, , by means of the adaptive kernel function bandwidth, when in the high-density area, the bandwidth is reduced to reduce the smoothing effect and avoid over-aggregating adjacent seabed terrain point clouds. In the low-density area, a wide bandwidth is maintained to capture the global distribution of sparse points, improving the stability of density estimation.

[0039] Dynamically adjust the neighborhood radius according to the local density and geometric curvature, so that the neighborhood adapts to both the density and structural complexity of the point cloud at the same time. For each key point , through the k-nearest neighbor algorithm, obtain the neighborhood point set . In this embodiment, k = 20, and based on the obtained neighborhood point set, construct a covariance matrix: ; where is the neighborhood point mean, represents the set composed of k nearest neighbor points selected for the key point , represents the three-dimensional coordinates of the j-th point in the neighborhood. Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues , and define the curvature value as the ratio of the minimum eigenvalue: . The range of the curvature value is (0, 1). The larger the value, the more complex the local geometric structure, such as sharp bends at the edges of pipelines and overhangs. Combine the normalized density and curvature to calculate the neighborhood radius. The formula for the neighborhood radius is: ​ ; Among them, is represented as the base radius parameter, is represented as the normalized local density, is represented as the global maximum local density value, is represented as the key point curvature value, is represented as the geometric structure sensitive factor, which balances 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 the high-density area and remains small in the low-density area, forming a directed graph with a variable neighborhood, constructing a more reasonable connection relationship in the point cloud, while enhancing the information propagation ability, being able to obtain richer context information, thereby accelerating the flow of remote information and improving the perception ability of global features.

[0040] S3. Set the masked attention mechanism according to the neighborhood structure, including calculating the neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; As Figure 3 shown, in the traditional ball query method, when constructing the local neighborhood, all neighbor points are treated equally, that is, all points located within the search ball will be included in the calculation without considering their actual contributions to the task. This method works well in the point cloud scenario with uniform distribution, but in the area with large density changes (the transition area between the exposed and suspended pipeline point cloud and the seabed terrain point cloud), it may lead to incorrect feature aggregation and affect the segmentation accuracy. In order to enhance the model's perception ability of local irregularities, a masked attention mechanism is proposed, enabling it to adaptively select effective neighbor points and assign different importance weights. Different from the conventional attention mechanism, the traditional Transformer encoder uses padding masks to process sequences of different lengths to ensure consistent input formats, and the decoder uses sequence masks to prevent the omission of real values. The masked attention mechanism adopted in this application is set in the encoder and is specifically for the neighborhood relationship of point cloud data and is used for local feature learning.

[0041] According to the set of key points obtained in step S2 and the neighborhood radius corresponding to each key point , for each key point ; Among them, represents the Euclidean distance between point and the neighborhood point . If the distance is less than , then it is determined that belongs to the neighborhood point set of , corresponding matrix position is 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 feature is concatenated with the three-dimensional coordinates to form an input feature vector: ; where d represents the dimension of the semantic feature, R represents the dimensional space, represents the semantic feature of point j, represents the three-dimensional coordinates of the j-th point. After concatenation, a multi-layer perceptron is used to construct a feature similarity matrix for the neighboring point cloud. The calculation formula of the feature similarity matrix is: ; where, 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 point, represents the three-dimensional coordinates of the j-th point in the neighborhood, and represent a multi-layer perceptron network of a one-dimensional convolutional layer, a normalization layer, and a GELU activation function layer, which is used to map two input point clouds to a multi-layer perceptron of a specified dimension, represents a feature concatenation operation, and have the same structure, both 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: ; where, represents the basic linear term, which provides linear scaling, makes the output range positively correlated with the input, and at the same time compresses the overall function value to about half of the input value, represents the hyperbolic tangent function term, represents the cubic term, which introduces a slight non-linear curvature to fit the tail characteristics of the Gaussian distribution. The constant part is an empirically optimized value.

[0042] Encode the concatenated features through two independent MLP networks and to output a feature vector of a specified dimension. Use the MLP to learn in the feature space and The potential and alignment relationships between them. To enable the recognition of both semantic and geometric information, the point cloud features and point cloud coordinates are encoded simultaneously. During the actual training process, -∞ represents points with extremely low correlation with the neighborhood. To ensure that these points will not be assigned non-zero attention weights during the subsequent Softmax process, leading to numerical instability, they are uniformly set to 10 9 , reflects the cosine similarity between the key point and the neighborhood points in the feature space. The larger the value, the more similar the semantic or geometric features of the two are.

[0043] After that, multi-stage normalization is used to suppress extreme weights, balance the importance of points within the neighborhood, and avoid a single neighborhood point dominating the feature aggregation. First, the first-stage normalization is performed, specifically Softmax global normalization: ; where i represents the key point, N represents the total number of neighborhood points, j and k represent different neighborhood points respectively, represents the feature similarity matrix, normalizes the attention weight of each key point i to the neighborhood point j into a probability distribution, and non-neighborhood points approach 0 after the exponential operation and are automatically masked (weight is 0), achieving a hard screening within the neighborhood range.

[0044] After that, the square root is taken element-wise for the weight matrix after Softmax normalization to suppress excessive weights (for example, the square root of weight 0.8 is 0.89, and the square root of weight 0.2 is 0.45, narrowing the gap). The square root renormalization formula: ; where, represents the weight after the first-stage Softmax normalization, 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 obtaining excessive weights due to highly similar features, balance the contributions of points within the local neighborhood, and finally, norm normalization is performed. The specific formula is: ; where K represents the total number of key points generated by random sampling, represents the weight after the second-stage square root renormalization, and is normalized column-wise (the weight of each neighborhood point j for all key points i ), ensuring that the sum of the attention weights of neighborhood points j at different key points is 1, further balancing the weight distribution across key points and avoiding a point dominating the global features due to being highly activated with high weights in the neighborhoods of multiple key points.

[0045] S4. Use the output results 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; As Figure 4 shown, for each key point , according to the weight values, sort them from high to low, and select the K neighboring points with the largest weights. The higher the weight of the neighboring point, the higher the feature similarity with the key point and the greater the contribution to pipeline recognition. Retain them preferentially to obtain the set of K neighboring points corresponding to each key point , and its corresponding attention weights . For each neighboring point , calculate the relative coordinates with the key point : and the Euclidean distance: ; Among them, represents the three-dimensional coordinates of the k-th neighboring point of the key point, that is . In addition, , and represent the three-dimensional coordinate values of the key point . Concatenate the key point coordinates , the neighboring point coordinates , the relative coordinates and the Euclidean distance to form the relative position feature: ; Among them, represents the concatenation symbol concatenated by dimension, represents the Euclidean distance. Perform a non-linear transformation on the relative position feature through a multi-layer perceptron (MLP) to generate the relative position encoding: ; Among them, represents the relative position feature, represents a one-dimensional convolutional layer, represents a batch normalization layer, represents an activation function layer. The final output dimension is 128 dimensions to enhance the expression ability of the position feature. Then perform relative semantic feature calculation. First, perform semantic feature dimensionality increase. For the semantic features of the key point and the neighboring point, perform dimensionality increase through a one-dimensional convolutional layer, and calculate the difference between the dimensionality-increased semantic features of the neighboring point and the key point: ; Among them, represents the original semantic feature of the key point, Represents the original semantic features of the neighboring points. The semantic features concatenate the normal vector, backscattering intensity (1D), and semantic category (1D) into relative semantic features. The calculation formula for the relative semantic features is as follows: ; Where, respectively represent the normal vector features of point clouds in different directions, represents the backscattering intensity feature of the point cloud, is the semantic category of the point cloud.

[0046] Finally, the relative position encoding is concatenated with the relative semantic features (256D) to form high-dimensional fusion features. The fusion features of K neighboring points are aggregated through a two-dimensional convolutional layer (Conv2D) to extract high-order features of the local region, and the maximum value of each dimension is retained through a max pooling operation. The feature dimension is further enhanced by a multi-layer perceptron to output the final enhanced point cloud features.

[0047] S5. Perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting the segmentation result by the Unit PointNet layer; 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, labels need to be predicted for each input point, and the number of points decreases during the downsampling in the encoding process, the information needs to be propagated back to each original point through the decoding layer.

[0048] For each original point , find its nearest neighbor key point in the key point set S through the K-nearest neighbor algorithm , calculate the Euclidean distance between the original point and each neighboring key point, and calculate the weighting coefficient based on the inverse distance: ; Where, represents the Euclidean distance between the original point and the neighboring key point , m represents the number of neighboring key points selected through the K-nearest neighbor algorithm. The closer the key point, the higher the weight. represents the weight exponent, defaulting to 2. represents the Euclidean distance between the original point and the neighboring key point to the power. During the decoding stage, when propagating the enhanced features of the key points to the original point cloud, through Adjust the contribution of adjacent points to avoid introducing noise from distant points, while enhancing the consistency of local features and ensuring local consistency of feature propagation. The enhanced features of key points are propagated to the original points through weighted averaging: ; Among them, represents the weighting coefficient, represents the key point after enhanced features, represents the feature vector obtained by propagating the original points through weighted averaging, which is used for subsequent segmentation or overhang height calculation. m represents the number of neighboring key points selected by the K-nearest neighbor algorithm. Finally, a feature matrix of all original points is generated. Compared with traditional nearest neighbor interpolation or uniform weighted interpolation, this method avoids feature blur caused by discrete jumps and global averaging through inverse distance weighting and K-nearest neighbor search. Especially in high-density change areas such as the pipe edge, it can more accurately retain detailed features. During the encoding process, intermediate features at each level are stored. After interpolation, the features from the encoding stage are concatenated with the interpolated features (with the same number of feature points) in a skip connection to retain rich local information: ; Among them, represents the channel dimension concatenation, represents the feature matrix that propagates from the enhanced features of key points to the original point cloud through inverse distance weighted interpolation in the decoding stage, represents the low-level intermediate features in the encoding stage with the same spatial resolution as

[0049] This can ensure that geometric information such as the spatial distribution of the original point cloud of low-level features can be retained, and the semantic information of high-level features can be transmitted to the low-resolution point cloud. The concatenated features are input into a Unit PointNet module, which consists of a one-dimensional convolutional layer, a one-dimensional batch normalization layer, a ReLU (Rectified Linear Unit) activation function layer, and a max pooling layer to update the feature vector of each point to enhance local geometric and global semantic information. Repeat this process until the features are propagated to the original point set. The expression of the Unit PointNet module is: ; Among them, represents the max pooling operation, represents the rectified linear activation function, represents the one-dimensional batch normalization layer, represents the one-dimensional convolutional layer, Denote the initial features obtained after jump stitching. The batch normalization layer normalizes the convolutional output to stabilize the feature distribution. The formula is: ; Among them, and are batch statistics, x is the input feature, and are learnable parameters. Inverse distance weighted averaging avoids the discrete jumps of nearest neighbor interpolation and makes the transition of pipeline edge features natural.

[0050] S6. Calculate and verify the suspension height based on the segmentation results, including extracting seabed points from non-pipeline point clouds and fitting a reference plane using inverse distance weighted interpolation based on the seabed points.

[0051] Based on the 3D point cloud segmentation model adopted in the embodiment, that is, the 3D point cloud segmentation model composed of the encoding layer in step S4 and the decoding layer in step S5, segment the suspended pipeline point cloud to accurately identify the suspended pipeline point cloud (label 1) from the massive point clouds and distinguish it from the seabed point cloud (label 0) and other feature point clouds. The segmented suspended pipeline point cloud and seabed point cloud are respectively used for subsequent reference plane fitting and suspension height calculation. After segmentation, export the suspended pipeline point cloud and seabed point cloud respectively, and perform plane fitting on the seabed point cloud to construct a reference datum plane. Subsequently, calculate the distance between the suspended pipeline point cloud and the fitted seabed plane in the vertical direction to accurately obtain the suspension height information. Inverse distance weighting is applicable to the surface generation of point cloud data. Its principle is that the closer a known point is to the interpolation point, the greater its influence, while the influence of points farther away is smaller. IDW calculates the value of the target interpolation point by assigning weights to each known point. The weight size depends on the distance between the point and the interpolation point: The calculation formula for the inverse distance weighted interpolation is: ; Among them, 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 the weight exponent used to control the influence degree of the distance, and n is the number of neighboring known points used for interpolation calculation.

[0052] The IDW interpolation method is a distance-weighted deterministic interpolation technique. Its calculation process does not rely on complex probability models or statistical analyses, thus having a relatively low computational cost. It is especially suitable for the rapid interpolation processing of large-scale point cloud data. In flat seabed areas, due to the relatively uniform distribution of bathymetric point clouds, IDW can better capture terrain changes and achieve accurate fitting. In scouring trench areas, IDW can use the elevation weighting interpolation of seabed points on both sides to maintain terrain continuity and avoid biases caused by data loss in global models (such as least squares). At the same time, IDW has a strong ability to identify noise and avoid abnormal undulations due to scattered data in the occluded areas around the submarine pipeline.

[0053] The sag height of the pipeline refers to the vertical distance between the submarine pipeline and the seabed reference plane. The core 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 sag height can be transformed into calculating the vertical distance from each suspended pipeline point cloud ( ) to the fitted surface S(x, y): ; Wherein, is the sag height of the submarine pipeline, is the water depth of the segmented suspended pipeline point cloud, D represents the pipeline diameter. Since the sound wave emitted by the multibeam first contacts the top of the submarine pipeline, when calculating the sag height, it is necessary to subtract the pipeline diameter D from the distance between the pipeline point cloud and the fitted surface. Assign the calculated suspended height to the point cloud attribute of the suspended pipeline. Subsequently, load the pipeline point cloud data containing sag height information in the ArcGIS platform and perform color rendering based on the sag height attribute to visually display the spatial suspension characteristics of the pipeline.

[0054] Embodiment 2 The difference between this embodiment and Embodiment 1 is that this embodiment provides a non-buried submarine pipeline recognition system based on multibeam point clouds, including: A data acquisition module, configured to acquire three-dimensional point cloud data of the submarine pipeline area and preprocess the acquired point cloud data; A neighborhood module, configured to: construct neighborhoods based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; An attention module, configured to: set a masked attention mechanism according to the neighborhood structure, including calculating the neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; The feature module is configured to: perform geometric encoding and feature enhancement using the output result of the masked attention mechanism, 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; The decoding module is configured to: perform a decoding operation on the enhanced feature to restore the feature representation of the complete point cloud, including propagating the feature to the original point cloud and outputting the segmentation result by the Unit PointNet layer; The output module is configured to: calculate and verify the suspension height based on the segmentation result, including extracting seabed points from the non-pipeline point cloud and fitting a reference plane using inverse distance weighted interpolation according to the seabed points.

[0055] The above are all preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A method for identifying non-buried submarine pipelines based on multi-beam point clouds, characterized in that Including: Obtain the three-dimensional point cloud data of the subsea pipeline area, and preprocess the obtained point cloud data; Construct neighborhoods based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; Set a masked attention mechanism according to the neighborhood structure, including calculating the neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; Perform geometric encoding and feature enhancement using the output result of the masked attention mechanism, 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; Perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting the segmentation result by the Unit PointNet layer; Calculate and verify the suspension height based on the segmentation result, including extracting seabed points from the non-pipeline point cloud and fitting a reference plane using inverse distance weighted interpolation according to the seabed points.

2. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, wherein The preprocessing of the obtained point cloud data includes merging point cloud data from different sources to 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, performing a difference operation on the overlapping regions of the water depth surfaces generated by the main survey line and the connecting survey line, deleting out-of-limit noise points, and finally correcting the original backscatter data using Lambert's law 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, wherein In the neighborhood construction based on the preprocessed point cloud data, multiple key points are selected from the preprocessed point cloud data by random sampling. By calculating the global density mean of the point cloud and adjusting the sampling probability of random sampling according to the global density mean, for each key point, a Gaussian kernel function is used to estimate its local density, and the neighborhood radius is adjusted according to the local density normalization result. The formula for the local density is: ； Among them, represents the total number of neighborhood points participating in the calculation, represents the kernel function bandwidth, which controls the action range of the Gaussian kernel, represents the three-dimensional coordinates of the key point, represents the three-dimensional coordinates of the j-th point in the neighborhood, represents the Euclidean distance.

4. A method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 3, characterized in that, The adjustment of the neighborhood radius according to the local density normalization result includes constructing a covariance matrix using the k-nearest neighbor point set of the key point, performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues, calculating curvature values based on the eigenvalues, then performing normalization processing on the local density, and calculating the neighborhood radius using the curvature values and the normalized local density. The formula for the neighborhood radius is: ； Among them, is represented as the base radius parameter, is represented as the normalized local density, is represented as the global maximum local density value, is represented as the key point curvature value, is represented as the geometric structure sensitive factor, balancing the influence weights of density and curvature on the neighborhood radius.

5. A method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that, The calculation of the neighborhood feature similarity matrix using a multi-layer perceptron includes obtaining the key point set and the neighborhood radius set, 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 formula for the feature similarity matrix is: ； Among them, represents the neighborhood radius, and respectively represent the point cloud semantic features of point i and point j, represents the three-dimensional coordinates of the key point, represents the three-dimensional coordinates of the j-th point within the neighborhood, and represent two multi-layer perceptron networks composed 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 the feature concatenation operation.

6. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 5, wherein The multi-stage normalization of the similarity matrix includes using the Softmax function to normalize the similarity matrix for the first-stage normalization, which is used to make the sum of the attention weights of each key point to its neighborhood points equal to 1. Perform a square root transformation on the matrix after Softmax normalization and then perform a second-stage normalization. By suppressing extreme weights, balance the importance distribution of points within the neighborhood. Finally, perform column-wise norm normalization on the matrix after the second-stage normalization. The formula for the first-stage normalization is: ； where i represents the key point, N represents the total number of neighborhood points, and j and k represent different neighborhood points respectively.

7. A method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 1, characterized in that Performing geometric encoding and feature enhancement using the output result of the masked attention mechanism includes screening K neighboring points for each key point based on the masked attention weight matrix, calculating the relative coordinates between the key point and its neighboring points, and generating relative position encoding by non-linearly mapping the relative coordinates. Concatenating the key point coordinates, neighboring point coordinates, relative coordinates, and Euclidean distance to form relative position features. Then, performing dimensionality increase processing on the semantic features of the neighboring points and the key point and calculating the difference to obtain relative semantic features. Fusing the relative position features and the relative semantic features, and performing feature extraction and aggregation through a convolutional neural network to obtain high-dimensional enhanced point cloud features. The calculation formula for the relative semantic features is as follows: ； Among them, respectively represent the normal vector features of the point cloud, represents the backscattering intensity feature of the point cloud, is the semantic category of the point cloud.

8. A method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 7, characterized in that, Performing a decoding operation on the enhanced features includes performing feature propagation based on the inverse distance weighted average interpolation method, propagating the enhanced features from the sampled points to all points of the original point cloud, and performing skip connection splicing on the features from the encoding stage and the interpolated features to obtain initial features. Using the initial features as input to construct a Unit PointNet module and performing iterative update using the Unit PointNet module to obtain a complete point cloud feature representation. The Unit PointNet module consists of a convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer. The calculation formula for the feature propagation is as follows: ； Among them, represents the weighting coefficient, represents the key point the enhanced feature, is expressed as the feature vector obtained by propagating the original point through weighted average, and m represents the number of selected nearest neighbor key points.

9. The method for identifying non-buried submarine pipelines based on multi-beam point clouds according to claim 8, wherein Extracting seabed points from the non-pipeline point cloud and fitting a reference plane using the inverse distance weighted interpolation based on the seabed points includes extracting a seabed point cloud data set from the complete point cloud feature representation, performing reference plane fitting on 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 reference plane. The calculation formula for the inverse distance weighted interpolation is as follows: ； Among them, is the estimated value of the interpolation 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 the weight index used to control the influence degree of the distance, and n is the number of adjacent known points used for interpolation calculation.

10. A non-buried submarine pipeline recognition system based on multi-beam point cloud, which executes the method described in claim 1, characterized in that, Including: A data acquisition module configured to acquire three-dimensional point cloud data of a subsea pipeline area and preprocess the acquired point cloud data; A neighborhood module configured to construct a neighborhood based on the preprocessed point cloud data, including obtaining K key points by random sampling, calculating the local density using kernel density estimation, and dynamically adjusting the neighborhood radius according to the local density; An attention module configured to set a masked attention mechanism according to the neighborhood structure, including calculating a neighborhood feature similarity matrix using a multi-layer perceptron and performing multi-stage normalization on the similarity matrix; A feature module configured to perform geometric encoding and feature enhancement using the output result of the masked attention mechanism, 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; A decoding module configured to perform a decoding operation on the enhanced features to restore the feature representation of the complete point cloud, including propagating the features to the original point cloud and outputting a segmentation result by the Unit PointNet layer; An output module configured to calculate and verify the suspension height based on the segmentation result, including extracting seabed points from the non-pipeline point cloud and fitting a reference plane using the inverse distance weighted interpolation based on the seabed points.

Images