An underground target dynamic electromagnetic detection method and system based on a graph neural network

Abstract

The application provides a kind of underground target dynamic electromagnetic detection method and system based on graph neural network, it is related to detection technical field, method includes: obtaining the pose information of detector on multiple detection points and secondary field feedback signal;With the pose information of each detection point and secondary field feedback signal as input feature, construct graph neural network;Obtain underground target detection data set;Using underground target detection data set, the graph neural network is trained;According to the graph neural network of training completion, dynamic electromagnetic detection is carried out to underground target.In the present application, the information sequence of dynamic pose and feedback signal is used to help the graph neural network to learn the continuity and time sequence characteristics in the movement process of detector, so as to capture dynamic change and meet the demand of real-time detection;At the same time, the graph neural network can effectively process the deviation of feedback signal caused by the movement or inclination of detector, so as to improve the positioning accuracy of underground target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of detection, in particular to a method and system for dynamic electromagnetic detection of underground targets based on a graph neural network. BACKGROUND

[0002] Electromagnetic detection of underground targets has wide applications in many fields due to its ability to effectively obtain spatial information of underground metal targets. However, traditional methods of underground target detection face many challenges. As the demand for detection continues to increase, the accuracy and real-time requirements of target detection have also gradually increased. How to improve detection accuracy in complex geographical environments and dynamic detection conditions has become a problem to be solved.

[0003] Currently, the main process of electromagnetic detection of underground targets includes moving an electromagnetic detector in the target area and collecting data through the sensed secondary field electromagnetic signals. During the entire detection process, the detector usually collects data at different positions and heights. Combined with the electromagnetic response model of the target body, the position, size and material of the target can be inverted. In terms of data processing, traditional methods mostly use offline processing, that is, after the detection is completed, all collected data are uniformly analyzed and processed to derive the related parameters of the underground target.

[0004] Although the above traditional method can achieve certain detection effect under certain conditions, it still has many limitations. First, due to the change of the posture of the detector during the detection process, the secondary field induction data will be affected by the movement and tilt of the detector, resulting in deviation of the measured data and affecting the positioning accuracy of the target. In addition, the time sequence characteristics between the detection data are usually ignored during data processing, and the dynamic information in the detection process is not fully utilized, which makes the detection result unable to reflect the real-time change of the target and difficult to meet the real-time detection demand. SUMMARY

[0005] In order to solve the technical problems that the existing method is affected by the movement and tilt of the detector due to the change of the posture of the detector during the detection process, resulting in deviation of the measured data and affecting the positioning accuracy of the target, and the time sequence characteristics between the detection data are usually ignored during data processing, and the dynamic information in the detection process is not fully utilized, which makes the detection result unable to reflect the real-time change of the target and difficult to meet the real-time detection demand, the present application provides a method and system for dynamic electromagnetic detection of underground targets based on a graph neural network.

[0006] The technical scheme provided by the embodiments of the present application is as follows:

[0007] First aspect:

[0008] The underground target dynamic electromagnetic detection method based on a graph neural network provided by the embodiment of the present application comprises the following steps:

[0009] S1: acquiring pose information of a detector at multiple detection points and a secondary field feedback signal;

[0010] S2: taking the pose information of each detection point and the secondary field feedback signal as input features to construct a graph neural network;

[0011] S3: acquiring an underground target detection data set;

[0012] S4: training the graph neural network by using the underground target detection data set;

[0013] S5: performing dynamic electromagnetic detection on an underground target according to the trained graph neural network.

[0014] The second aspect is:

[0015] The underground target dynamic electromagnetic detection system based on a graph neural network provided by the embodiment of the present application comprises the following steps:

[0016] A processor;

[0017] A memory, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the underground target dynamic electromagnetic detection method based on a graph neural network according to the first aspect.

[0018] The third aspect is:

[0019] The computer readable storage medium provided by the embodiment of the present application stores a computer program, and the program is executed by the processor to implement the underground target dynamic electromagnetic detection method based on a graph neural network according to the first aspect.

[0020] The technical scheme provided by the embodiment of the present application has at least the following beneficial effects:

[0021] (1) In the present application, the pose information of a detector at multiple detection points and a secondary field feedback signal are acquired, the pose information of each detection point and the secondary field feedback signal are taken as input features to construct a graph neural network, and the graph neural network can effectively process the feedback signal deviation caused by the movement or inclination of the detector, thereby improving the positioning accuracy of the underground target.

[0022] (2) In the present application, the graph neural network is trained by using the underground target detection data set, and the underground target is dynamically electromagnetically detected according to the trained graph neural network. The relationship between nodes can be captured through the graph neural network, the continuity and time sequence characteristics in the motion process of the detector are learned by using the information sequence of the dynamic pose and the feedback signal, so that the dynamic change is captured, and the demand of real-time detection is met. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1 A flowchart of a graph neural network-based underground target dynamic electromagnetic detection method provided by the embodiment of the present application is shown in the figure.

[0025] Figure 2 A structure diagram of a graph neural network-based underground target dynamic electromagnetic detection system provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0026] The technical solutions in the present application will be described below with reference to the drawings.

[0027] In the embodiments of the present application, the words such as "example", "for example" are used to represent an example, illustration or explanation. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0028] In the embodiments of the present application, "image" and "picture" can be used interchangeably. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent. "Of", "corresponding" and "corresponding" can be used interchangeably. It should be pointed out that when the distinction is not emphasized, the meanings expressed are consistent.

[0029] In the embodiments of the present application, sometimes the subscript such as W1 may be mistakenly used as a non-subscript such as W1. When the distinction is not emphasized, the meanings expressed are consistent.

[0030] In order to make the technical problems, technical solutions and advantages to be solved by the present application more clear, the following will be described in detail in combination with the drawings and specific embodiments.

[0031] Referring to the drawings Figure 1 , a flowchart of a method for detecting a dynamic electromagnetic target underground based on a graph neural network is shown.

[0032] The embodiment of the present application provides a method for detecting a dynamic electromagnetic target underground based on a graph neural network, which can be realized by a device for detecting a dynamic electromagnetic target underground based on a graph neural network. The device can be a terminal or a server. The processing flow of the method for detecting a dynamic electromagnetic target underground based on a graph neural network can include the following steps:

[0033] S1: Obtain the pose information of the detector at a plurality of detection points and the secondary field feedback signal.

[0034] The secondary field feedback signal refers to the reflection or scattering signal received by the detector after emitting electromagnetic waves (primary field) due to the interaction between the electromagnetic waves and the underground target. These signals reflect the electromagnetic properties and position of the target.

[0035] It should be noted that the dynamic detection path of the underground target is composed of N detection points. The detector moves in turn at the N detection points and obtains the secondary field feedback signal output by the corresponding sensor at each point.

[0036] In the present application, moving in turn at a plurality of detection points and collecting the secondary field feedback signal can improve the detection and identification capability of the target and meet the detection requirements in dynamic and complex environments.

[0037] S2: Take the pose information of each detection point and the secondary field feedback signal as input features to construct a graph neural network.

[0038] The graph neural network (Graph Neural Networks, GNNs) is a deep learning model for processing graph structure data. Graph structure data widely exists in many practical applications, such as social networks, molecular structures, knowledge graphs and traffic networks. The main purpose of GNNs is to effectively perform node classification, graph classification, link prediction and other tasks by utilizing the structural information of the graph.

[0039] In one possible implementation, S2 specifically includes:

[0040] S201: Take each detection point as a node of the graph neural network, and construct a corresponding node feature matrix for each detection point according to the pose information and the secondary field feedback signal at each detection point.

[0041] wherein the node feature matrix is a core concept in the graph neural network, and is used to represent the feature information of each node in the graph. In the graph neural network, the feature matrix of each node provides the attributes and related data of the node, and is the input of the graph structure data.

[0042] In a possible implementation, S201 specifically includes:

[0043] S2011: splice the pose information of the current detection point position acquired by each detector with the secondary field feedback signal to obtain a one-dimensional vector corresponding to each detector.

[0044] S2012: longitudinally splice the one-dimensional vectors corresponding to each detector to obtain the node feature matrix.

[0045] Specifically, taking each detection point position as a node of the graph neural network, the node initial feature X is composed of two parts of the pose information and the corresponding Bs feedback signal. The pose information is composed of the spatial coordinates D x , D y , D z and the attitude angle D γ , D θ , The secondary field feedback signal is composed of array signals from different sensors. The array signal value output by each sensor is preprocessed into a one-dimensional vector with a size of 1*Q. The sensor book signal is the sampling (B s,1 , B s,2 , …, B s,Q ) of Q times of the secondary field feedback signal within the detection time T, which can represent the intensity and attenuation characteristics of the secondary field feedback signal. Assuming that the number of sensors on the detector is M, the pose information is spliced with the array signal vector of each sensor, and then M one-dimensional vectors with a size of 1*(6+Q) are obtained. Longitudinally splicing the M one-dimensional vectors, a matrix with a size of M*(6+Q) is obtained, which is the node initial feature X_(M×(6+Q)). The node initial feature X contains the pose information of the detector at the detection point position and the secondary field feedback signal of each sensor. Generating the node initial feature for each detection point position, N node initial features can be obtained.

[0046] In the present application, by fusing the detection point position and the sensor feedback signal in one feature matrix, the network can comprehensively utilize multiple aspects of information for learning and prediction, further optimizing the performance of the model. At the same time, by splicing the pose information of the detection point position and the secondary field feedback signal into the node feature matrix, the spatial position and electromagnetic response of the target are comprehensively considered. This multi-dimensional feature integration provides a more comprehensive description of the target, which helps to improve the detection accuracy.

[0047] S202: Construct an adjacency feature matrix corresponding to each probe point according to the node feature matrix.

[0048] The adjacency feature matrix is an important concept in graph neural networks, which is used to represent the relationship or connection mode between nodes in the graph. Its main role is to help the graph neural network capture the correlation information between nodes, support effective processing and analysis of graph structure data.

[0049] In one possible implementation, S202 specifically includes:

[0050] S2021: Input the node feature matrix of the current node into the convolution layer to extract the feature information of the current node itself.

[0051] S2022: Input the feature information of the current node itself into the activation function to obtain the node feature:

[0052] Y = σ(W a ·X + b a )

[0053] Where Y represents the node feature, σ represents the activation function, X represents the node feature matrix, W a represents the weight matrix of the feature information itself, and b a represents the bias term of the feature information itself.

[0054] S2023: Generate a relationship matrix by multiplying the node feature of the current node with the transpose of the node feature of the i-th node, input the relationship matrix into the convolution layer, extract the correlation information between the current node and the i-th node, and perform standard normalization and activation operations on the correlation information to obtain the updated relationship matrix:

[0055]

[0056] Where, represents the updated relationship matrix, σ represents the activation function, BN represents the normalization layer, Y1 represents the node feature of the current node, Y i represents the node feature of the i-th node, () T represents the transpose matrix, W b represents the node relationship modeling weight matrix, and b b represents the node relationship modeling bias term.

[0057] S2024: Perform one-dimensional convolution on the updated relationship matrix to compress the dimension and generate the adjacency feature matrix:

[0058]

[0059] wherein, represents the adjacency feature matrix of the current node and the i-th node, represents the convolution kernel weight matrix, represents the relationship matrix, represents the bias term in one-dimensional convolution.

[0060] For example, taking the 1st node as an example, the construction process of the adjacency feature matrix is shown:

[0061] In order to extract the feature information of each node itself, first output the node initial feature X to a convolution layer to obtain the node feature information, and then input the feature to the activation function sigmoid to remove the redundant information and obtain the node feature Y. This operation does not change the feature size, which is still M*(6+Q). Let the node feature Y of the i-th node be Y i , by multiplying the node feature of the first node and the transpose of the node feature of the i-th node, a relationship matrix with a dimension of M*M is generated; then input the relationship matrix into a new convolution layer to extract the correlation information of the two; and then perform standard normalization operation and activation operation on the correlation information to obtain a refined relationship matrix with a dimension of M*M In order to facilitate the subsequent input of the graph convolution layer to generate new node information, one-dimensional convolution is used for dimension compression to generate which is the edge vector representation of the 1st node and the i-th node. By the same operation, the correlation relationship between the 1st node and the remaining N-1 nodes is calculated, and then N-1 edge vector representations of the 1st node are obtained.

[0062] In the present application, by calculating the relationship matrix, the interaction and correlation between nodes can be effectively captured. This helps the network to understand the complex relationship between nodes and improves the expressiveness of the graph neural network. At the same time, this method ensures that the graph neural network can fully utilize the feature information of nodes and edges, thereby improving the overall performance and accuracy of the underground target detection system.

[0063] S203: combining the node feature matrix and the adjacency feature matrix using the graph convolution layer to output the final node feature.

[0064] In one possible implementation, S203 specifically includes:

[0065] S2031: dimensionally splicing the adjacency feature matrices of multiple nodes to obtain global edge features.

[0066] Among them, the global edge feature is an important concept in the graph neural network for representing the feature information of the edge in the graph. When processing the graph structure data, the global edge feature is used to capture the comprehensive information of all edges in the graph, which helps the model better understand the relationship between nodes and the overall structure of the graph.

[0067] In the present application, the global edge feature provides the connection information between nodes and the characteristics of the edge, effectively captures these relationships through graph convolution operation, and helps to build more accurate node representation.

[0068] S2032: Multiply the global edge feature with the weight matrix to obtain an intermediate matrix:

[0069] Z (6+Q)*M =W (6+Q)*(N-1) ·F (N-1)*M

[0070] Wherein, Z (6+Q)*M represents an intermediate matrix with a dimension of (6+Q)*M, W (6+Q)*(N-1) represents a weight matrix with a size of (6+Q)*(N-1), and F (N-1)*M represents a global edge feature with a size of (N-1)*M.

[0071] S2033: Multiply the node feature with the intermediate matrix to obtain a feature matrix containing node feature information and inter-node relationship information, and perform feature activation through an activation function. Multiply the node feature with its transpose to obtain a self-feature matrix. Element-wise sum the activated feature and the self-feature matrix to obtain an intermediate node feature:

[0072]

[0073] Wherein, Y M ′ *M represents an intermediate node feature with a dimension of M*M, Y M*(6+Q) represents a node feature with a dimension of (6+Q)*M, () T represents a transpose matrix, and sigma represents an activation function. Z (6+Q)*M represents an intermediate matrix with a dimension of (6+Q)*M.

[0074] In the present application, the combination of node features and intermediate matrices can effectively integrate inter-node relationship information and node self-features. Through element-wise summation and activation processing of the features, more rich and fine node feature representation can be generated.

[0075] S2034: Input the intermediate node feature into the value convolution layer and the Relu activation layer in sequence to output the final node feature:

[0076] Z M*M = Relu (We·Y' + b) M*M +b e )

[0077] wherein, Z M*M represents the final node feature with the dimension of M*M, Relu() represents the Relu activation function, W e represents the node feature aggregation weight matrix, and b e represents the node feature aggregation bias term.

[0078] Optionally, the same operation is performed on N detection points, and multiple node features can be obtained.

[0079] In the present application, the node feature matrix and the adjacency feature matrix are combined, so that the node's own features and the relationship information between nodes can be comprehensively considered. This comprehensive feature representation enables the graph neural network to more comprehensively understand the characteristics of the node and its position in the graph.

[0080] Further, through detailed feature processing and fusion, the graph neural network is provided with rich input features, which significantly improves the ability to process graph structure data.

[0081] S204: According to the final node feature, one-dimensional graph convolution operation and splicing operation are sequentially performed to obtain the graph global feature.

[0082] Among them, one-dimensional graph convolution (1D Graph Convolution) is an important operation in graph neural network (GNNs), and one-dimensional graph convolution is mainly used to process the relationship between nodes and edges in graph data, especially in processing feature propagation and updating in the graph.

[0083] In one possible implementation, S204 specifically includes:

[0084] S2041: The final node feature is subjected to adjacency feature matrix and graph convolution operation for the second time to obtain a second node feature.

[0085] S2042: One-dimensional convolution operation is performed on the second node feature to obtain a third node feature.

[0086] S2043: The multiple third node features are spliced to obtain a graph global feature containing the underground target feature information obtained by all detection points:

[0087] H N*M = concat[G 1 ,G 2 ,…,G N ]

[0088] wherein, HN*M denotes the global feature of the graph with dimension N*M, N denotes the total number of nodes, G i denotes the i-th third node feature, i = 1, 2…N, and concat[] denotes a concatenation operation.

[0089] In the present application, by passing the final node feature through one-dimensional graph convolution and concatenation operation, the feature information of each node can be effectively integrated. This enables the model to comprehensively capture the features of each node in the graph, thereby obtaining more rich global information. At the same time, the global feature can be used as the feature input of the graph for the graph classification task, which can accurately capture the overall information of the graph and perform classification or regression tasks, thereby improving the detection accuracy.

[0090] S205: outputting the detection result of the underground target according to the global feature of the graph.

[0091] In one possible implementation, S205 is specifically:

[0092] inputting the global feature of the graph into a convolutional layer for feature extraction, and sequentially inputting the extracted features into a fully connected layer to output the detection result of the underground target:

[0093] H N*M = Fc(σ(W f ·H N*M +b f ))

[0094] wherein H N*M denotes the global feature of the graph with dimension N*M, σ denotes an activation function, Fc() denotes a fully connected layer operation, W f denotes an output result mapping weight matrix, and b f denotes an output result mapping bias term.

[0095] In one possible implementation, the detection result includes: centroid spatial coordinates, material, and morphology.

[0096] In one possible implementation, the centroid spatial coordinates are specifically: (T x , T y , T z ), T x denotes the spatial position coordinate of the target on the x-axis, T y denotes the spatial position coordinate of the target on the y-axis, and T z denotes the spatial position coordinate of the target on the z-axis.

[0097] The category of the material includes: no target, metal target, and non-metal target.

[0098] The morphology includes: the base radius, height, pitch angle, and yaw angle of the cylindrical target.

[0099] Specifically, the result output module of the underground target detection task is composed of a centroid positioning module, a material judgment module and a shape output module. As shown in formula 8, the construction process of the centroid positioning module is as follows: the global feature H of the image is input into a convolutional layer for feature extraction, and then input into a fully connected layer in sequence, and the output dimension of the fully connected layer is set to 3, that is, the spatial coordinates of the centroid of the target (T N*M , T x , T y ). z

[0100] The construction process of the material judgment module and the shape output module is similar to that of the centroid positioning module, except that the output dimension of the material judgment module is set to 1, that is, corresponding to (no target, metal target, non-metal target) three categories; the output dimension of the shape output module is set to 4, that is, corresponding to the radius, height, pitch angle and yaw angle of the cylindrical target.

[0101] In the present application, by inputting the global feature of the image into the convolutional layer and the fully connected layer, the accurate features of the target can be extracted, so as to accurately calculate the spatial coordinates of the centroid of the underground target, and improve the positioning accuracy and recognition ability of the underground target.

[0102] S206: complete the construction of the graph neural network.

[0103] S3: obtain an underground target detection data set.

[0104] Optionally, the underground target detection data set is obtained in the following manner:

[0105] A batch of targets with different materials and shapes are taken as detection samples, which are placed in different positions underground in different postures, and the material, shape and posture information of the target, that is, the target information, is recorded. A detector is used to perform multiple underground target detection experiments according to different detection points, and the corresponding detector posture information and the array signals output by different sensors, that is, the detector information, are recorded, so as to obtain the underground target detection data set.

[0106] In the present application, the target detection data set with different materials and shapes ensures that the data set can cover a wide range of target types and characteristics. This diversity enables the data set to more comprehensively reflect different situations in actual detection. At the same time, the target is placed in different positions underground in different postures to simulate the distribution of targets in real environment, enhancing the representativeness and authenticity of the data.

[0107] S4: train the graph neural network using the underground target detection data set.

[0108] ​Specifically, the graph neural network is trained based on the underground target detection data set until the loss function converges, and the training of the graph neural network is completed.

[0109] The loss function is a function used to measure the gap between the model prediction and the actual target in machine learning and deep learning. It is a core component of training the model, and by optimizing the loss function, the model parameters are adjusted to improve the prediction accuracy of the model.

[0110] In the present application, by training based on the underground target detection data set, the model can learn to extract useful features from the detection data, thereby improving the prediction accuracy of the underground target. At the same time, the loss function is used to quantify the gap between the model prediction and the actual target, and optimizing the loss function can make the prediction results of the model closer to the true value, improving the overall accuracy.

[0111] S5: According to the trained graph neural network, the underground target is dynamically electromagnetically detected.

[0112] The technical scheme provided by the embodiments of the present application has at least the following beneficial effects:

[0113] (1) In the present application, the pose information of the detector at multiple detection points and the secondary field feedback signal are obtained, and the pose information of each detection point and the secondary field feedback signal are used as input features to construct a graph neural network. The graph neural network can effectively handle the deviation of the feedback signal caused by the movement or tilt of the detector, thereby improving the positioning accuracy of the underground target.

[0114] (2) In the present application, the graph neural network is trained using the underground target detection data set, and the underground target is dynamically electromagnetically detected according to the trained graph neural network. The graph neural network can capture the relationship between nodes and use the information sequence of dynamic pose and feedback signal to help the network learn the continuity and timing characteristics in the movement process of the detector, thereby capturing dynamic changes and meeting the real-time detection requirements.

[0115] Reference is made to the accompanying drawings Figure 2 , which shows a structural schematic diagram of an underground target dynamic electromagnetic detection system based on a graph neural network provided by the present application.

[0116] The underground target dynamic electromagnetic detection system 20 based on the graph neural network provided by the embodiments of the present application is applied to the above-mentioned underground target dynamic electromagnetic detection method based on the graph neural network, and comprises:

[0117] a processor 201;

[0118] The memory 202 has computer readable instructions stored thereon, and the computer readable instructions are executed by the processor 201 to implement the underground target dynamic electromagnetic detection method based on the graph neural network according to the first aspect.

[0119] The underground target dynamic electromagnetic detection system 20 based on the graph neural network provided by the present application can execute the underground target dynamic electromagnetic detection method based on the graph neural network, and achieve the same or similar technical effects. To avoid repetition, the present application will not be described again.

[0120] The technical scheme provided by the embodiments of the present application has at least the following beneficial effects:

[0121] (1) In the present application, the pose information of the detector at multiple detection points and the secondary field feedback signal are obtained, and the pose information of each detection point and the secondary field feedback signal are used as input features to construct a graph neural network. The graph neural network can effectively process the feedback signal deviation caused by the movement or tilt of the detector, thereby improving the positioning accuracy of the underground target.

[0122] (2) In the present application, the graph neural network is trained using an underground target detection data set, and the underground target is dynamically electromagnetically detected according to the trained graph neural network. The graph neural network can capture the relationship between nodes, and use the information sequence of the dynamic pose and the feedback signal to help the network learn the continuity and time sequence features in the movement process of the detector, thereby capturing dynamic changes and meeting the real-time detection requirements.

[0123] It should be understood that the processor in the embodiments of the present application can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0124] It should also be understood that the memory in the embodiments of the present application can be volatile or nonvolatile memory, or can include both volatile and nonvolatile memory. The nonvolatile memory can be read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically EPROM (EEPROM), or flash memory. The volatile memory can be random access memory (RAM) used as external cache. By way of example, and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0125] The above-described embodiments can be implemented in whole or in part by software, hardware (e.g., circuitry), firmware, or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center through wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0126] It should be understood that the term "and / or" herein merely describes an association relationship of associated objects, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after it, but it can also represent an "and / or" relationship, which can be understood according to the context before and after it.

[0127] In the present application, "at least one" means one or more, and "multiple" means two or more. "At least one of the following" or the like means any combination of the items, including any combination of single or multiple items. For example, at least one of a, b, or c can represent a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be single or multiple.

[0128] It should be understood that in various embodiments of the present application, the size of the sequence number of the above-described processes does not mean the order of execution, and the execution order of the processes should be determined according to their functions and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0129] Those skilled in the art can clearly understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0130] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the devices, apparatuses and units described above can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0131] In several embodiments provided by the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0132] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0133] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit.

[0134] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0135] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the underground target dynamic electromagnetic detection method based on a graph neural network.

[0136] The computer readable storage medium provided by the present application can realize the steps and effects of the underground target dynamic electromagnetic detection method based on the graph neural network of the method embodiment, and the present application will not be repeated here to avoid repetition.

[0137] The technical solutions provided by the embodiments of the present application have at least the following beneficial effects:

[0138] (1) In the present application, the pose information of the detector at a plurality of detection points and the secondary field feedback signal are obtained, and the pose information of each detection point and the secondary field feedback signal are used as input features to construct a graph neural network. The graph neural network can effectively process the feedback signal deviation caused by the movement or tilt of the detector, thereby improving the positioning accuracy of the underground target.

[0139] (2) In the present application, the graph neural network is trained using an underground target detection data set, and the underground target is dynamically electromagnetically detected according to the trained graph neural network. The graph neural network can capture the relationship between nodes, use the information sequence of the dynamic pose and the feedback signal, help the network learn the continuity and time sequence features in the movement process of the detector, and thus capture the dynamic changes to meet the real-time detection requirements.

[0140] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0141] The following points need to be explained:

[0142] (1) The drawings of the embodiments of the present application only relate to the structures involved in the embodiments of the present application, and other structures can be referred to the general design.

[0143] (2) In the drawings used to describe the embodiments of the present application, the thickness of a layer or region is exaggerated or reduced for clarity, i.e., the drawings are not drawn according to the actual scale. It can be understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, it can be "directly" on or under the other element or there can be an intermediate element.

[0144] (3) In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other to obtain new embodiments.

[0145] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for underground target dynamic electromagnetic detection based on a graph neural network, characterized in that, The method comprises the following steps: S1: acquiring pose information of a detector at multiple detection points and a secondary field feedback signal; S2: taking the pose information and the secondary field feedback signal of each detection point as input features to construct a graph neural network; S3: acquiring an underground target detection data set; S4: training the graph neural network by using the underground target detection data set; S5: dynamically electromagnetically detecting an underground target according to the trained graph neural network.

2. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 1, characterized in that, The S2 specifically comprises: S201: taking each detection point as a node of the graph neural network, and constructing a corresponding node feature matrix for each detection point according to the pose information and the secondary field feedback signal at the detection point; S202: constructing a corresponding adjacency feature matrix for each detection point according to the node feature matrix; S203: combining the node feature matrix and the adjacency feature matrix by using a graph convolution layer to output a final node feature; S204: sequentially performing one-dimensional graph convolution operation and splicing operation according to the final node feature to obtain a graph global feature; S205: outputting a detection result of an underground target according to the graph global feature; S206: completing the construction of the graph neural network.

3. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 2, characterized in that, The S201 specifically comprises: S2011: splicing the pose information and the secondary field feedback signal of the current detection point acquired by each detector to obtain a one-dimensional vector corresponding to each detector; S2012: longitudinally splicing the one-dimensional vectors corresponding to the detectors to obtain the node feature matrix.

4. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 3, characterized in that, The S202 specifically comprises: S2021: inputting the node feature matrix of a current node into a convolution layer to extract feature information of the current node itself; S2022: inputting the feature information of the current node itself into an activation function to obtain a node feature: Y = σ(W a • X + b a ) wherein Y represents a node feature, sigma represents an activation function, X represents a node feature matrix, W a represents a self-feature information weight matrix, b a represents a self-feature information bias term; S2023: multiplying the node feature of the current node and the transpose of the node feature of the i-th node to generate a relationship matrix, inputting the relationship matrix into the convolution layer to extract correlation information of the current node and the i-th node, and performing standard normalization and activation operation on the correlation information to obtain an updated relationship matrix: wherein, represents an updated relationship matrix, σ represents an activation function, BN represents a normalization layer, Y1 represents a node feature of a current node, Y i represents a node feature of the i-th node, () T represents a transposed matrix, W b represents a node relationship modeling weight matrix, b b represents a node relationship modeling bias term; S2024: performing one-dimensional convolution on the updated relationship matrix to compress the dimension of the relationship matrix to generate an adjacency feature matrix: wherein, represents an adjacency feature matrix of the current node and the i-th node, represents a convolution kernel weight matrix, represents a relation matrix, represents a bias term in one-dimensional convolution.

5. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 4, characterized in that, The S203 specifically comprises: S2031: dimensionally splicing the adjacency feature matrices of multiple nodes to obtain a global edge feature; S2032: multiplying the global edge feature and a weight matrix to obtain an intermediate matrix: Z (6+Q)*M = W (6+Q)*(N-1) • F (N-1)* M wherein Z (6+Q)*M represents an intermediate matrix of dimension (6+Q)*M, W (6+Q)*(N-1) represents a weight matrix of size (6+Q)*(N-1), F (N-1)*M represents a global edge feature of size (N-1)*M; S2033: multiplying the node feature and the intermediate matrix to obtain a feature matrix containing feature information of the node itself and relationship information between nodes, performing feature activation by using an activation function, multiplying the node feature and the transpose of the node itself to obtain a self-feature matrix, and performing element-wise summation on the activated feature and the self-feature matrix to obtain an intermediate node feature: where Y M ′ *M represents an intermediate node feature with a dimension of M*M, Y M*(6+Q) represents a node feature with a dimension of (6+Q)*M, () T represents a transpose matrix, σ represents an activation function, Z (6+Q)*M represents an intermediate matrix with a dimension of (6+Q)*M; S2034: inputting the intermediate node feature into the convolution layer and the Relu activation layer in sequence to output a final node feature: Z M*M = Relu(W e · Y′ M*M + b e ) wherein Z M*M represents the final node feature with the dimension of M*M, Relu() represents the Relu activation function, W e represents the node feature aggregation weight matrix, b e represents the node feature aggregation bias term.

6. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 5, characterized in that, The S204 specifically comprises: S2041: The final node feature is subjected to an adjacency feature matrix and a graph convolution operation for the second time to obtain a second node feature; S2042: One-dimensional convolution operation is performed on the second node feature to obtain a third node feature; S2043: The plurality of third node features are spliced to obtain a graph global feature containing underground target feature information obtained by the plurality of probe points; H N*M = concat[G 1 ,G 2 ,…,G N ] wherein H N*M represents the global feature of the graph with dimension N*M, N represents the total number of nodes, G i represents the i-th third node feature, i=1,2…N, and concat[] represents a concatenation operation.

7. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 6, characterized in that, The S205 is specifically: The graph global feature is input into a convolution layer for feature extraction, and the extracted features are sequentially input into a fully connected layer to output a detection result of the underground target: H N*M = Fc(σ(W f ·H N*M +b f )) wherein H N*M represents the global feature of the image with dimension N*M, σ represents an activation function, Fc() represents a fully connected layer operation, W f represents an output result mapping weight matrix, b f represents an output result mapping bias term.

8. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 7, characterized in that, The detection result includes: centroid spatial coordinates, material, and shape.

9. The underground target dynamic electromagnetic exploration method based on a graph neural network according to claim 8, characterized in that, The centroid space coordinates are specifically: (T x , T y , T z ), T x represents the space position coordinates of the target on the x-axis, T y represents the space position coordinates of the target on the y-axis, and T z represents the space position coordinates of the target on the z-axis. The material category includes: no target, metal target, and non-metal target; The shape includes: the base radius, height, pitch angle, and yaw angle of the cylindrical target.

10. A system for underground target dynamic electromagnetic detection based on a graph neural network, characterized in that, It includes: A processor; A memory, the memory has computer readable instructions stored thereon, and the computer readable instructions are executed by the processor to implement the underground target dynamic electromagnetic detection method based on the graph neural network in any one of claims 1 to 9.

Images