Submarine pipeline state recognition method based on double attention and graph convolution fusion

Abstract

The invention relates to the technical field of underwater oil pipeline state detection, and provides a submarine pipeline state recognition method based on double attention and graph convolution fusion. Comprising the following steps: step 1, acquiring multi-beam point cloud data and shallow profile data of a submarine pipeline, and constructing a data set; 2, extracting local neighborhood features of the pipeline, and combining the local neighborhood features of the pipeline with the global features; 3, learning neighborhood point deep features in a local neighborhood through a space attention and channel attention module, and endowing pipeline edge features with different weights and taking the pipeline edge features as new features; 4, the steps 2 and 3 are repeated, the depth of the neural network is deepened, the pipeline features are enhanced, the pipeline state probability is calculated, and whether the pipeline is exposed, suspended or buried or not is judged according to the probability result, the judgment speed is high, and the problems of hysteresis, mechanical performance and the like existing in existing manual pipeline observation and manual pipeline state recognition can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater oil pipeline status detection, and provides a submarine pipeline status recognition method based on dual attention and graph convolution fusion. Background Art

[0002] In recent decades, global offshore oil and gas development has surged. Submarine pipelines are the fastest, safest, most reliable, and most economical means of continuously and massively transporting oil and gas resources. They shoulder the crucial task of offshore oil and gas gathering and transportation and are often called the "lifeline" of offshore oil and gas projects. Since 1954, when the US-based Brown & Root Offshore Engineering Company laid the first submarine pipeline in the Gulf of Mexico, a vast submarine pipeline network has emerged in major waters around the world. Since China completed its first submarine oil pipeline in the Chengbei Oilfield in the Bohai Sea in 1985, it has laid over 6,000 km of submarine pipelines in various sea areas.

[0003] Submarine pipelines, as crucial production facilities in offshore oil and gas fields, are characterized by high investment and high risk. Submarine pipelines are expensive, costing between $300,000 and $1 million per kilometer. The marine environment in which they operate is extremely complex and subject to numerous uncertainties. As pipelines are laid longer and operated for longer periods, the probability of damage increases, and accidents become more frequent. The oil and gas transported by submarine pipelines are harmful to humans. Leaks or damage to submarine pipelines can have serious impacts on the surrounding environment and personnel. At the very least, leaks can lead to wasteful resources. In severe cases, crude oil or natural gas leaks can cause explosions, resulting in casualties, property damage, and severe damage to the surrounding ecological environment. Furthermore, large-scale leaks can halt oil and gas field production, directly resulting in significant economic losses. The causes of submarine pipeline damage are complex and multifaceted, but corrosion, anchoring, and impacts with foreign objects are the primary causes.

[0004] As global land-based energy consumption increases and oil resources decrease, oil production is shifting to the marine environment, making underwater pipelines increasingly important. Monitoring the condition of submarine pipelines is crucial not only to the supply of oil and energy but also to the protection of the submarine ecosystem. Due to the complex and changing hydrodynamic environment, shallow geological conditions, and shallow sedimentary conditions, as well as the increasing number of offshore structures, changes in the marine dynamic environment have occurred. Intensified currents or high-speed vortices can occur in certain areas, causing severe erosion of the seabed near platforms and causing submarine pipelines to become suspended to varying degrees.

[0005] When the span of an exposed, suspended submarine pipeline reaches a certain length, the pipeline is prone to bending and deformation. Under the influence of ocean currents and waves, suspended pipelines are prone to vortex-induced resonance (VIR). When this reaches a certain level, stress concentration can cause cracks in the pipeline, leading to pipeline failure and major marine accidents such as oil and gas leaks. This not only causes significant economic losses to offshore oil and gas production, but also pollutes the marine environment due to large-scale oil and gas leaks, resulting in adverse social impacts. Therefore, it is imperative to promptly manage suspended submarine pipelines, and to conduct timely inspections and investigations of the management results to ensure the normal and safe operation of the submarine pipelines.

[0006] Traditional detection methods use multi-beam side-scan sonar and shallow profiling to analyze underwater terrain, generating a point cloud dataset of underwater oil pipelines. After visualizing the pipelines, manual monitoring and analysis are then performed. However, this method is prone to misjudgments and missed detections, posing safety risks to underwater pipeline protection. Furthermore, this method is mechanically backward and requires engineers and technicians to constantly observe the pipelines during construction, which challenges their attention and endurance. Therefore, using neural networks to process the continuous changes in underwater pipelines not only effectively analyzes topographic features, but also performs faster than human visual recognition. Therefore, the development of a neural network-based submarine oil and gas pipeline buried status system can, to some extent, address the lag and mechanical limitations of existing manual pipeline observation and identification methods. Summary of the Invention

[0007] The purpose of the present invention is to address the shortcomings of the existing technology and propose a submarine pipeline status recognition method based on dual attention and graph convolution fusion. It is a submarine oil and gas pipeline buried status recognition method based on dual attention mechanism and graph convolution fusion.

[0008] Solve the following technical problems:

[0009] Conventional submarine pipeline status judgment relies on manual judgment, which has high misjudgment and missed judgment rates, low efficiency, lag, and poor accuracy.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] A submarine pipeline state recognition method based on dual attention and graph convolution fusion includes the following steps:

[0012] Step 1: Obtain multi-beam point cloud data and shallow profile data of the submarine pipeline and construct a data set;

[0013] Step 2: Extract pipeline local neighborhood features and combine them with global features.

[0014] Step 3: The spatial attention and channel attention modules are used to learn the deep features of the neighborhood points in the local neighborhood, and different weights are assigned to the pipeline edge features as new features.

[0015] Step 4: Repeat steps 2) and 3) to deepen the neural network, enhance the pipeline features, and calculate the pipeline state probability. Based on the probability results, determine whether the pipeline is exposed, suspended, or buried.

[0016] Furthermore, the pipeline local feature extraction in step 2 includes:

[0017] (1) For any point u in the data set i , get u through the nearest neighbor algorithm i k neighbor points N(u i )={u i1 ,u i2 ,u i3 ,...,u ik};

[0018] (2) with u i is the center of the sphere, u i to N(u i ) Construct a sphere with the longest distance of neighbor points as radius r; in the sphere, find b points as core points u b ; For each core point u b There is a weight matrix W respectively b Correspondingly, for any v ij , define the kernel function p, and the calculation method is shown in formula (1):

[0019]

[0020] Among them, B represents the total number of core points; v ij Indicates the relative distance between the center point and the neighboring points; the weight coefficient q(v ij ,u b ) is defined as shown in formula (2):

[0021]

[0022] Among them, θ is the influence distance of the core point, and its size is determined by the input density of the point cloud;

[0023] (3) Use the weight matrix to calculate the local feature v ij Perform feature transformation to obtain e ij , as shown in formula (3);

[0024] e ij =p(v ij )v ij (3).

[0025] Furthermore, in step 2, local features and global features are connected through a multi-layer perceptron to obtain edge features.

[0026] Furthermore, the edge feature is q ij , as follows:

[0027] q ij =k β (u i , e ij ) (4).

[0028] Furthermore, the step three includes:

[0029] (1) Preprocessing edge features: Use the previously calculated edge feature vector q ij , which is F, performs a global pooling on it, and then forms a neural network Y by the pooling layer, the fully connected layer and the activation layer j ;

[0030] (2) Calculate the weight matrix of spatial attention: Use the reduced edge feature vector F and generate two new feature vectors O and I using two independent convolution operations; perform matrix multiplication on the new feature vectors to obtain the normalized spatial attention weight matrix X ij ;

[0031] (3) Calculate the enhanced output feature A S :The neural network Y j With attention weight X ij Perform matrix multiplication operation, and finally multiply it with the scale parameter and the input feature F i Perform element-by-element summation to obtain the enhanced output feature A S ;

[0032] (4) Calculate the feature vector A after the channel attention mechanism C :The enhanced output feature A S After dimensionality reduction, maximum pooling and average pooling are performed respectively, and then a multi-layer perceptron MLP with shared weights is added to obtain the channel attention weight. Finally, the weight matrix is ​​matrix multiplied with the input feature vector F to obtain the feature vector A after the channel attention mechanism. C .

[0033] Furthermore, the weight matrix of the spatial attention is X ij , calculated by the softmax function, the calculation formula is as follows:

[0034] X ij =soft max(O i , I i) (5);

[0035] Where, O i and I j are the feature representations of point i and point j in O and I respectively.

[0036] Furthermore, the enhanced output feature A S , the calculation formula is as follows:

[0037]

[0038] Where α is a learnable scale parameter with an initial value of 0.

[0039] Furthermore, the feature vector A after the channel attention mechanism C The calculation is as follows:

[0040] A C =σ(MLP(MaxPool(F))+MLP(AvgPool(F)))F (7);

[0041] Where σ represents the Sigmoid activation function.

[0042] Furthermore, the step 4 includes:

[0043] (1) The eigenvector A C and A S , connect through pooling, re-input edge convolution, repeat steps 2 and 3, enhance pipeline features, extend network depth, increase the number of parameters in the neural network, and improve prediction accuracy;

[0044] (2) Input the enhanced pipeline features into the MLP, perform softmax normalization on the results, and obtain the pipeline state probability matrix R;

[0045] (3) The obtained probabilities of each state are compared, and the optimal probability is selected as the neural network judgment result, which is the state of the pipeline.

[0046] Furthermore, the pipeline state probability matrix R is calculated as follows:

[0047] R=softmax(MLP(A C )+MLP(A S )) (8).

[0048] The beneficial effects of the present invention are as follows:

[0049] The present invention analyzes and calculates multi-beam side-scan data and shallow profile data that are easily collected on the ground, uses edge convolution to extract pipeline point cloud features, and performs data enhancement on the point cloud dataset. At the same time, a multi-layer perceptron is used to connect the local features of the pipeline with the global features of the pipeline to obtain the pipeline edge features. The edge features are enhanced through spatial attention to increase their computational share. After inputting them into channel attention, the maximum pooling operation is used to obtain the aggregated local neighborhood features, which are used as the new features of the center point. The present invention combines dual attention with graph dynamic convolution to greatly improve the early warning accuracy of underwater pipeline rupture accidents, and to a certain extent solves the problems of hysteresis and mechanicality in existing manual observation of pipelines and manual pipeline status identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a workflow diagram of the buried state recognition method for submarine oil and gas pipelines based on dual attention mechanism and graph convolution fusion;

[0051] Figure 2 This is the structure diagram of the spatial attention algorithm;

[0052] Figure 3 This is the structure diagram of the channel attention algorithm.

[0053] Figure 4 This is the overall structural diagram of the network based on the dual attention mechanism and graph convolution fusion algorithm. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0055] Example 1:

[0056] A submarine oil and gas pipeline buried status recognition system based on dual attention and graph convolution fusion includes a data acquisition module, a pipeline feature extraction and enhancement module, and a pipeline status recognition module.

[0057] Among them, the data acquisition module includes high-frequency multi-beam sonar and three-dimensional profiler, which are used to obtain multi-beam point cloud data and shallow profile data of submarine pipelines.

[0058] The pipeline feature extraction and enhancement module extracts local pipeline features through convolutional neural networks and combines the local pipeline features with global features through multi-layer perceptrons.

[0059] A multilayer perceptron, or MLP, is a feedforward artificial neural network model that maps multiple input datasets to a single output dataset.

[0060] The pipeline state recognition module includes a self-attention module, which calculates the pipeline state probability and completes the pipeline state judgment through the combination of CAM (channel attention) and SAM (spatial attention).

[0061] Example 2:

[0062] A submarine pipeline state recognition method based on dual attention and graph convolution fusion includes the following steps:

[0063] 1) Data collection:

[0064] Multi-beam point cloud data and shallow profile data of submarine oil and gas pipelines are collected through multi-beam sonar systems, shallow profile sonars and other data acquisition equipment.

[0065] 2) Use edge convolution (EdgeConv) to extract local neighborhood features of the pipeline and combine the extracted features with global features:

[0066] (1) For any point u in the data set i , and obtain u by K-Nearest Neighbor (KNN) algorithm i k neighbor points N(u i )={u i1 ,u i2 ,u i3 ,...,u ik}.

[0067] (2) with u i is the center of the sphere, u i to N(u i ) Construct a sphere with the longest distance of neighbor points as radius r. In the sphere, find b points as core points u b For each core point u b There is a weight matrix W respectively b Correspondingly, for any v ij , define the kernel function p, and the calculation method is shown in formula (1):

[0068]

[0069] Among them, B represents the total number of core points; v ij Indicates the relative distance between the center point and the neighboring points; the weight coefficient q(v ij ,u b ) is defined as shown in formula (2):

[0070]

[0071] Among them, θ is the influence distance of the core point, and its size is determined by the input density of the point cloud.

[0072] (3) Use the weight matrix to calculate the local feature v ij Perform feature transformation to obtain e ij , as shown in formula (3);

[0073] e ij =p(v ij )v ij (3)

[0074] (4) Connect local features and global features through a multilayer perceptron (MLP) to obtain edge features q ij , as shown in formula (4):

[0075] q ij =k β (u i , e ij ) (4)

[0076] Among them, k β :R F ×R F →R F′ It is a nonlinear function with a set of learnable parameters β.

[0077] 3) The spatial attention and channel attention modules are used to learn the deep features of the neighborhood points in the local neighborhood, and different weights are assigned to the pipeline edge features as new features:

[0078] (1) Preprocess the edge features and use the previously calculated edge feature vector q ij , which is F, and a global pooling is performed on it to reduce its dimension. Then, through the neural network composed of the pooling layer, the fully connected layer (FC) and the activation layer (ReLU), Y is obtained. j .

[0079] (2) Calculate the weight matrix of spatial attention. Use the reduced edge feature vector F and generate two new feature vectors O and I using two independent convolution (ConV1, ConV2) operations. Perform matrix multiplication on O and I and apply the softmax function to obtain the normalized spatial attention weight matrix X. ij , calculated as shown in formula (5):

[0080] X ij =soft max(O i , I i ) (5)

[0081] Among them, O i and I jare the feature representations of point i and point j in O and I respectively.

[0082] (3) Calculate new features and convert Y j With attention weight X ij Perform matrix multiplication (Scale) operation, and finally multiply it with the scale parameter a and the input feature F i Perform element-by-element summation to obtain the enhanced output feature A S , as shown in formula (6):

[0083]

[0084] Here, α is a learnable scale parameter with an initial value of 0.

[0085] (4) Calculate new features and convert A S After dimensionality reduction (Pooling), Max Pooling (MaxPool) and Average Pooling (AvgPool) are performed respectively, and then a shared weight MLP is passed. After summing, the channel attention weight is obtained through the Sigmold function. Finally, the weight matrix and the input feature vector F are matrix multiplied (Scale) to obtain the feature vector A after the channel attention mechanism. C , calculated as shown in formula (7);

[0086] A C =σ(MLP(MaxPool(F))+MLP(AvgPool(F)))F (7)

[0087] Among them, σ represents the Sigmoid activation function.

[0088] 4) Repeat steps 2) and 3) to deepen the neural network, enhance pipeline features, and calculate pipeline state probabilities. Based on the probability results, determine whether the pipeline is exposed, suspended, or buried:

[0089] (1) The new feature vector A obtained by the above operation C and A S , connect through pooling, re-input edge convolution (EdgeConv), repeat steps 2) and 3), enhance pipeline features, extend the network depth, increase the number of parameters in the neural network, and improve prediction accuracy.

[0090] (2) Input the enhanced pipeline features into the MLP, perform softmax normalization on the results, and calculate the probability of the pipeline state. The calculation is shown in formula (8):

[0091] R=softmax(MLP(AC )+MLP(A S )) (8)

[0092] Among them, R is the pipeline state probability matrix, which records the probabilities of different states.

[0093] (3) The obtained probabilities of each state are compared, and the optimal probability is selected as the neural network judgment result, which is the state of the pipeline. When the pipeline is in a suspended and exposed state, necessary measures must be taken to ensure the normal operation of the oil pipeline.

[0094] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A submarine pipeline state recognition method based on dual attention and graph convolution fusion, characterized by: The steps include: Step 1: Obtain multi-beam point cloud data and shallow profile data of the submarine pipeline and construct a data set; Step 2: Extract pipeline local neighborhood features and combine them with global features. Step 3: The spatial attention and channel attention modules are used to learn the deep features of the neighborhood points in the local neighborhood, and different weights are assigned to the pipeline edge features as new features. Step 4: Repeat steps 2) and 3) to deepen the neural network, enhance the pipeline features, and calculate the pipeline state probability. Based on the probability results, determine whether the pipeline is exposed, suspended, or buried.

2. The submarine pipeline state recognition method based on dual attention and graph convolution fusion according to claim 1 is characterized in that: The pipeline local feature extraction in step 2 includes: (1) For any point u in the data set i , get u through the nearest neighbor algorithm i k neighbor points N(u i )={u i1 ,u i2 ,u i3 ,…,u ik }; (2) Take u i is the center of the sphere, u i to N(u i ) Construct a sphere with the longest distance of neighbor points as radius r; in the sphere, find b points as core points u b ; For each core point u b There is a weight matrix W respectively b Correspondingly, for any v ij , define the kernel function p, and the calculation method is shown in formula (1): Among them, B represents the total number of core points; v ij Indicates the relative distance between the center point and the neighboring points; the weight coefficient q(v ij ,u b ) is defined as shown in formula (2): Among them, θ is the influence distance of the core point, and its size is determined by the input density of the point cloud; (3) Use the weight matrix to calculate the local feature v ij Perform feature transformation to obtain e ij , as shown in formula (3); yes ij =p(v ij )v ij (3)。 3. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 2 is characterized in that: The second step connects the local features and the global features through a multi-layer perceptron to obtain edge features.

4. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 3 is characterized in that: The edge feature is q ij , as follows: q ij =k β (u i ,e ij ) (4)。 5. A submarine pipeline state recognition method based on dual attention and graph convolution fusion according to any one of claims 1 to 4, characterized in that: The step three includes: (1) Preprocessing edge features: Use the previously calculated edge feature vector q ij , which is F, performs a global pooling on it, and then forms a neural network Y by the pooling layer, the fully connected layer and the activation layer j ; (2) Calculate the weight matrix of spatial attention: Use the reduced edge feature vector F and generate two new feature vectors O and I using two independent convolution operations; perform matrix multiplication on the new feature vectors to obtain the normalized spatial attention weight matrix X ij ; (3) Calculate the enhanced output feature A S :The neural network Y j With attention weight X ij Perform matrix multiplication operation, and finally multiply it with the scale parameter and the input feature F i Perform element-by-element summation to obtain the enhanced output feature A S ; (4) Calculate the feature vector A after the channel attention mechanism C :The enhanced output feature A S After dimensionality reduction, maximum pooling and average pooling are performed respectively, and then a multi-layer perceptron MLP with shared weights is added to obtain the channel attention weight. Finally, the weight matrix is ​​matrix multiplied with the input feature vector F to obtain the feature vector A after the channel attention mechanism. C .

6. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 5 is characterized in that: The weight matrix of the spatial attention is X ij , calculated by the softmax function, the calculation formula is as follows: X ij =softmax(O i ,I i ) (5); Where, O i and I j are the feature representations of point i and point j in O and I respectively.

7. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 5 is characterized in that: The enhanced output feature A S , the calculation formula is as follows: Where α is a learnable scale parameter with an initial value of 0.

8. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 5 is characterized in that: The feature vector A after the channel attention mechanism C The calculation is as follows: A C =σ(MLP(MaxPool(F))+MLP(AvgPool(F)))F (7); Where σ represents the Sigmoid activation function.

9. A submarine pipeline state recognition method based on dual attention and graph convolution fusion according to any one of claims 6 to 8, characterized in that: The fourth step includes: (1) The eigenvector A C and A S , connect through pooling, re-input edge convolution, repeat steps 2 and 3, enhance pipeline features, extend network depth, increase the number of parameters in the neural network, and improve prediction accuracy; (2) Input the enhanced pipeline features into the MLP, perform softmax normalization on the results, and obtain the pipeline state probability matrix R; (3) The obtained probabilities of each state are compared, and the optimal probability is selected as the neural network judgment result, which is the state of the pipeline.

10. The method for submarine pipeline state recognition based on dual attention and graph convolution fusion according to claim 9 is characterized in that: The pipeline state probability matrix R is calculated as follows: R=softmax(MLP(A C )+MLP(A S ))(8)。

Images