Underground vibration field array information intelligent restoration method and system

Abstract

The invention discloses an underground vibration field array information intelligent restoration method, which comprises the following steps of: S1, arranging a sensor and a virtual measuring point, and acquiring an explosion vibration signal by using the sensor so as to obtain actual data; meanwhile, by simulating underground three-layer media, performing forward modeling simulation by using an underground explosion propagation model to obtain simulation data; s2, the actual data and the simulation data are combined into a trace gather graph, and a target domain training data set and a source domain training data set are established based on the trace gather graph; s3, performing primary training on the deep interpolation model by using the source domain training data set; and finally, migrating the deep interpolation model after primary training to a target domain data set to perform secondary training in a self-supervised learning mode. And S4, interpolation of the virtual measuring point signals is carried out based on the depth interpolation model subjected to secondary training, so that restoration of the underground vibration field array information is completed. The invention further discloses a corresponding system, electronic equipment and a computer readable storage medium.

Description

Technical Field

[0001] The present invention relates to the fields of blasting vibration testing technology and damage power assessment technology, and in particular to an underground vibration field array information intelligent repair method and system. Background Art

[0002] The ground shock effect refers to the phenomenon that the explosion shock wave propagates in the rock and soil medium, causing strong vibrations in the structural foundation and leading to structural damage. The vibration wave field signal obtained by finite nodes is used to reconstruct the energy distribution of the underground explosion load in the test area and its distribution over time in an inversion manner to reflect the degree of damage caused by the underground explosion to underground and above-ground buildings. The spatiotemporal distribution of this explosion power field is an important test indicator for assessing the destructive power of this type of weapon. It is an important basis for analyzing the energy conduction mode of this type of ammunition explosion, optimizing the development plan of the weapon system, and evaluating the actual combat destructive effectiveness of ammunition.

[0003] Compared with the large-area, large-equivalent, large-depth, and long-term earthquake source positioning of deep earthquakes, deep coal mining, and oil exploration, the shallow underground space has the following differences: (1) The explosion point is relatively shallow, generally not exceeding 100m, which belongs to the problem of shallow space energy field reconstruction and imaging; (2) The shallow surface geological structure is complex, including sand, gravel, rock, mud and other media, and it is impossible to draw on the existing deep formation model structure, which belongs to the problem of near-surface complex space energy field reconstruction and imaging; (3) The constitutive properties of the soil in the near-field of the explosion are elastic-plastic, and the elastic waves are greatly affected by ground reflection and refraction. The wave group aliasing is complex and the seismic phase characteristics are not obvious, which belongs to the problem of underground explosion near-field energy field reconstruction and imaging.

[0004] In the process of underground energy field reconstruction, the completeness of shock wave test data directly affects the reconstruction accuracy. Domestic and foreign researchers mainly use two methods to improve the completeness of test data, including: (1) increasing the number of measurement nodes, reducing the sparsity of data through the redundancy of the layout, thereby improving the completeness of test data; (2) using interpolation algorithms (such as nonlinear interpolation, radial basis function, etc., combined with shock wave empirical formulas) to compensate for the amount of data in the test blind area.

[0005] As high-power ammunition technology continues to improve, the following problems still exist:

[0006] (1) The conventional fitting interpolation method can solve the problem of energy field reconstruction under isotropic conditions in the explosion field. However, in the shallow underground space, due to the complex underground structure, it is impossible to ignore the various different media at a shallow depth, and it is impossible to accurately establish the fitting interpolation equation based on the limited measurement points. Therefore, the applicability of this method is relatively low.

[0007] (2) The interpolated signal error is large;

[0008] (3) The numerical simulation method can invert the process of the earthquake source explosion in the underground through known signals, and can conduct theoretical analysis on the explosion mechanism and target effect. However, in the shallow underground space, the different values ​​of the state equation parameters, material model parameters and grid density often have a great influence on the inversion reasoning results, resulting in large deviations in the interpolation calculation results. Summary of the invention

[0009] The purpose of the present invention is to provide an underground vibration field array information intelligent repair method and system, which reconstructs the time domain signal of the virtual measuring point of the underground vibration field in a dual-drive manner by combining the underground explosion propagation model with actual data. First, a simulation training data set is constructed through the underground explosion propagation model: when the underground medium structure information is known, a simulation stratum medium model is established, and the signal of the preset point is solved by the finite difference method through the elastic wave equation, which is used as the source domain training data set of the deep interpolation model designed by the present invention and trained; then the actual data is used as the target domain data set, and the trained deep interpolation model is further trained: using the transfer learning method, the downsampling module and the position attention module in the trained deep interpolation model are frozen, and the actual collected data is used as the target domain training data set for training, so as to optimize the model's interpolation ability for the actual data and complete the interpolation of the limited virtual measuring point signals at the same time.

[0010] The first aspect of the present invention is to provide an underground vibration field array information intelligent repair method, comprising:

[0011] S1, arranging sensors and virtual measuring points, using the sensors to collect explosion vibration signals, and then obtaining actual data; at the same time, by simulating three layers of underground media, using the underground explosion propagation model forward simulation to obtain simulation data;

[0012] S2, combining the actual data and the simulated data into a gather map, and establishing a target domain training data set and a source domain training data set based on the gather map;

[0013] S3, training the deep interpolation model once using the source domain training data set; finally migrating the deep interpolation model after the once training to the target domain data set for secondary training in a self-supervised learning manner;

[0014] S4, interpolating virtual measuring point signals based on the second-trained deep interpolation model, thereby completing the restoration of the underground vibration field array information.

[0015] Preferably, the S1 comprises:

[0016] S11, arranging sensors according to requirements, and using the sensors to collect vibration signals of explosions, thereby obtaining actual data;

[0017] S12, arranging virtual measuring points according to requirements, simulating three layers of underground media, and obtaining simulation data by forward simulation of an underground explosion propagation model; the underground explosion propagation model is a mathematical model established by an elastic wave equation, and the simulation data obtained by simulation includes obtaining the simulation data by numerical calculation of the wave field using a finite difference method.

[0018] Preferably, the S12 includes:

[0019] (1) A structure with one layer of concrete, one layer of compacted gravel, and one layer of semi-infinite compacted clay was used for simulation, and a three-dimensional space velocity matrix was established to simulate the three layers of underground media;

[0020] (2) The source excitation function used in data simulation is set to Ricker wavelet, expressed as formula (1):

[0021]

[0022] Among them, f m represents the main frequency of the wave, which is usually set to 30 Hz, and t represents time. Using the finite difference method, t is replaced by kΔt and brought into the calculation, and the acoustic wave equation is used to perform numerical simulation of transverse and longitudinal waves respectively. The general form of the acoustic wave is shown in formula (2):

[0023]

[0024] In equation (2), there are second-order partial derivatives of the fluctuation u and second-order partial derivatives of the spatial coordinates x and y. For the second-order partial derivatives of the time term, the finite difference method is used to express equation (2) as equations (3)-(5) using the three-point difference format (FDM-2):

[0025]

[0026] Among them, υ(x,z) represents the speed of a discrete point in space; the subscript of u represents the index of the discrete point in space, the superscript represents the differential index in time, the k moment represents the current moment, k+1 represents the next moment, and k-1 represents the previous moment; in space, since Δx=Δy, it is commonly set to Δh=Δx=Δy, thereby obtaining the fluctuation value for a certain spatial sampling point at a certain moment as formula (6):

[0027]

[0028] (3) Set the location and interval of the earthquake source, set the interval of the sensor, and set the time sampling rate and fluctuation time at the same time, so as to obtain the fluctuation value at the sensor point at each moment; for a sensor, combine the fluctuation values ​​at all moments to obtain the signal received by this sensor.

[0029] Preferably, S2 includes:

[0030] S21, combining the actual data and the simulated data into a plurality of gather maps, including: interpolating target points and arranging virtual measurement points, assuming that the signal obtained at the virtual measurement points is 0 through explosion experiments, and combining the virtual measurement points with the actual data obtained to form a gather map;

[0031] (1) Preprocessing the actual data, analyzing the instantaneous frequency conversion characteristics of the signal, using VMD to select the peak value and number of layers of the threshold to decompose the signal, retaining the useful signal and removing the noise;

[0032] (2) Taking a row of sensors with a fixed y-axis coordinate as an example, all pre-processed real signals in this row and the signals of the virtual measurement points represented by 0 are arranged in a group according to the x-axis direction sequence, which is represented as a gather map; the vertical direction represents the complete time step, and the horizontal direction represents the sequence of the deployed detectors, and all gather maps are obtained;

[0033] S22, creating a target domain training dataset; including:

[0034] (1) For one of the multiple gather maps, set the shape and size, vertical step length and horizontal step length of a rectangular frame, perform sliding extraction on the gather map, and obtain multiple training samples, thereby obtaining all gather integration maps and all training samples; each training sample is recorded as S i , the shape is (1024×64), where represents the signal received by the jth sensor of the ith sample, as a column matrix;

[0035] (2) At this time, the column matrix of each training sample is a time domain signal received by a certain channel. In order to distinguish the channel of the actual received signal from the channel of the signal that needs to be reconstructed, the actual deployed sensor that can receive the actual vibration signal is set to 1, and the virtual sensor corresponding to the point where the signal needs to be reconstructed is set to 0. A one-dimensional mask matrix is ​​established for each sample, denoted as m i , with a length of 64, and then expanded into a two-dimensional mask matrix with a shape of (1024, 64), where i represents the sample sequence;

[0036] S23, creating a source domain training data set, including: obtaining source domain training samples from all gather images, all the obtained samples are source domain training samples, and creating a source domain training data set based on the source domain training samples.

[0037] Preferably, S3 includes:

[0038] S31, building a deep interpolation model framework, wherein the deep interpolation model framework used for training the source domain training data set and the target domain training data set is the same; the deep interpolation model framework is based on a U-net network and consists of 4 downsampling modules, 3 upsampling modules and three position attention modules, wherein the downsampling modules and the position attention modules are only trained by the source domain training data set;

[0039] S32, building a downsampling module, including: a downsampling residual block and a downsampling multi-scale convolution block, wherein the downsampling residual block includes multiple two-dimensional convolution layers and a rectified linear unit (ReLU), and multiple parallel convolution paths with kernels of different sizes are added after the downsampling residual block to form a multi-scale convolution module;

[0040] S33, building an upsampling module, including: a deconvolution layer, an upsampling residual block and an upsampling multi-scale convolution block, wherein the upsampling residual block and the upsampling multi-scale convolution block are designed the same as the downsampling module, except that the step size of the first layer of convolution is set to 1, and the other settings are the same;

[0041] S34, building a position attention module, including: setting the size of the input to the position attention module to C×H×W, reorganizing and transposing after passing through the first convolution layer to obtain a data shape of N×C, where N=H×W, performing matrix multiplication with the data reorganized after passing through the second convolution layer and obtaining an attention score with a shape of N×N through a softmax layer, performing matrix multiplication with the data reorganized after passing through the third convolution layer and reorganizing back to C×H×W, and finally adding element by element with the input to obtain an output feature;

[0042] S35, performing model training, including: using a transfer learning method to perform model training using the created simulated data set as a source domain training data set, and then using the real data set as a target domain training data set to perform model training after model adaptive adjustment on the trained interpolation model.

[0043] Preferably, the S35 includes:

[0044] (1) Use the created simulated dataset as the source domain training dataset for model training, including:

[0045] Before each round of training, all complete gather samples are randomly created into a one-dimensional mask matrix according to a certain number of seismic traces. "1" means retention and "0" means missing. The missing rate is controlled at 70%. All the created masks are multiplied by the samples using the broadcast mechanism to obtain the corresponding number of missing gather maps as network input. At the same time, the complete gather samples are used as labels. Iterate 100 rounds. During the process, 128 batches of samples are taken for training at a time. The learning rate is set to 0.001 and multiplied by 0.1 every 20 rounds. The loss function is set to MSE, and the optimizer uses the Adam algorithm to obtain the corresponding model weight parameters.

[0046] (2) Freeze the weight parameters of the four downsampling modules and the three position attention modules of the model trained on the source domain training dataset and migrate them to the target domain training dataset for self-supervised training, including:

[0047] When training in the target domain, the network parameters of the pre-trained model are introduced, and the weight parameters of the four downsampling modules and the three position attention modules are frozen. The self-supervised training method is used for training, including:

[0048] For a certain sample, further missing The remaining gather samples are used as the input of the network. During training, 100 rounds are iterated. During the training, 128 batches of samples are taken at a time. The learning rate is set to 0.0001 and multiplied by 0.1 every 20 rounds. The loss function is given by formula (7). They are all different random values;

[0049]

[0050] Among them, Out i represents the output value obtained when the i-th missing sample is input into the network, “·” represents the Hadamard product, and MSE represents the mean square error loss function.

[0051] Preferably, after the training of the deep interpolation model is completed, the samples obtained from one explosion are re-input into the deep interpolation model that has been trained twice, and the same signal in the output results of the laterally extracted gathers is homogenized; the same signal in the output results of the longitudinally extracted gathers is spliced ​​in time, and finally, the signal interpolated from the virtual measuring point is fitted into a smooth curve, thereby completing the repair of the array information of the underground vibration field.

[0052] The second aspect of the present invention is to provide an underground vibration field array information intelligent repair system for implementing the method of the first aspect, comprising:

[0053] The actual data and simulation data acquisition module (101) is used to arrange sensors and virtual measuring points, collect explosion vibration signals using the sensors, and obtain actual data; at the same time, simulate the three-layer underground medium and obtain simulation data using the underground explosion propagation model forward simulation;

[0054] A training data set construction module (102) is used to combine the actual data and the simulated data into a gather map, and to establish a target domain training data set and a source domain training data set based on the gather map;

[0055] A model training module (103) is used to perform a primary training on the deep interpolation model using the source domain training data set; and finally migrate the deep interpolation model after the primary training to the target domain data set for secondary training in a self-supervised learning manner;

[0056] The information repair module (104) is used to perform interpolation of virtual measurement point signals based on the depth interpolation model that has undergone the secondary training, thereby completing the repair of the underground vibration field array information.

[0057] A third aspect of the present invention provides an electronic device, comprising a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is configured to read the instructions and execute the method described in the first aspect.

[0058] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor and execute the method described in the first aspect.

[0059] Beneficial effects of the method and system of the present invention:

[0060] (1) The present invention simulates the real shallow underground medium, uses the finite difference method to solve the acoustic wave equation to generate the time domain signals collected by the sensor during the explosion, and combines them into a gather integration diagram to complete the establishment of the source domain data set; by moving the seismic source, the underground space is fully covered to ensure the validity of the sample, thereby enhancing the integrity, reliability and authenticity of the simulated data set.

[0061] (2) The present invention uses U-net as the basic architecture of the generative network model. By adding a multi-scale convolution module in the encoder, the receptive field is expanded while the model feature extraction capability is increased. By adding a position attention module at the jump connection to learn the spatial dependency of the feature, it can better fit the spatial relationship between the measurement points, reduce the interpolation error of the virtual measurement points, and make the reconstructed waveform closer to the real waveform.

[0062] (3) In order to solve the problem that the network cannot be effectively trained due to the small number of real data samples, the present invention uses transfer learning. By first training on a large number of samples of numerical simulation and then adaptively adjusting with a small number of real data samples, the model parameters trained on the simulation data can be effectively utilized to quickly adapt to the characteristics of the real data, thereby improving the network's robustness to signal interpolation.

[0063] (4) In order to solve the problem of incomplete data as labels in target domain training, the present invention uses a self-supervised training method, which uses random masks to construct training samples from missing data. Combined with the network previously trained using simulated data, it can train a spatial law that is more suitable for real collected signal data, so as to have higher interpolation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the related technologies, the drawings required for use in the specific embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 A flow chart of an underground vibration field array information intelligent repair method provided according to an embodiment of the present invention;

[0066] Figure 2 A schematic diagram of the layout of sensors and explosion points provided according to an embodiment of the present invention;

[0067] Figure 3 A schematic diagram of the layout of virtual measurement points provided according to an embodiment of the present invention;

[0068] Figure 4 A schematic diagram of a simulation layout provided according to an embodiment of the present invention;

[0069] Figure 5 It is an overall schematic diagram of a depth interpolation model provided according to an embodiment of the present invention;

[0070] Figure 6 A schematic diagram of a downsampling module provided according to an embodiment of the present invention;

[0071] Figure 7 A schematic diagram of a multi-scale convolution module provided according to an embodiment of the present invention;

[0072] Figure 8 A schematic diagram of an upsampling module provided according to an embodiment of the present invention;

[0073] Fig. 9A schematic diagram of a position attention module provided according to an embodiment of the present invention;

[0074] Fig.10 A diagram of a target domain training framework provided according to an embodiment of the present invention;

[0075] Fig.11 This is an architecture diagram of an underground vibration field array information intelligent repair system provided according to an embodiment of the present invention;

[0076] Fig.12 A structural diagram of an electronic device provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0079] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0080] Embodiment 1

[0081] like Figure 1 As shown, this embodiment provides an underground vibration field array information intelligent repair method, including:

[0082] S1, arranging sensors and virtual measuring points, using the sensors to collect explosion vibration signals, and then obtaining actual data; at the same time, by simulating three layers of underground media, using the underground explosion propagation model forward simulation to obtain simulation data;

[0083] As a preferred embodiment, the S1 includes:

[0084] S11, arranging sensors according to requirements, and using the sensors to collect vibration signals of explosions, thereby obtaining actual data;

[0085] In this embodiment, the implementation of step S11 is used to construct training samples of actual data, and the acquired real signals are used for training while interpolating the missing signals. Taking the actual underground airport runway as an example, sensors and virtual measuring points are deployed, and actual data is obtained by collecting signals excited to the surface at the source position, and then the acquired actual data can be used to establish real missing training samples, and the real missing training samples are used to construct the target domain data set;

[0086] In this embodiment, taking the following experiment as an example, a sensor array is arranged to obtain vibration signals through explosions.

[0087] like Figure 2 As shown, in order to improve the accuracy of the shallow underground explosion experiment and reduce the experimental error, this experimental design uses a certain equivalent of TNT explosives as the explosion source. The size of the experimental area is set to be 120 meters long, 120 meters wide, and 60 meters deep to ensure sufficient space to simulate and analyze the propagation characteristics of shock waves. The experimental coordinate system is set with the east direction of the experimental area as the positive direction of the x-axis, the north direction as the positive direction of the y-axis, and the positive direction of the z-axis from bottom to top. The plane 60 meters underground is the reference plane of z=0, and the number of detectors arranged in the experiment is set to N×N (taking N equal to 29 as an example).

[0088] The earthquake source is located at the center of this coordinate system, i.e. (60, 60, 30), which is used as the starting point of the explosion experiment. In order to make all signals in the constructed sample data have consistent spatial continuity, the sensors are evenly arranged in a square array, and the distance between each adjacent sensor along the x-axis and y-axis is 4 meters.

[0089] S12, arranging virtual measuring points according to requirements, simulating three layers of underground media, and obtaining simulation data by forward simulation of an underground explosion propagation model; the underground explosion propagation model is a mathematical model established by an elastic wave equation, and the simulation data obtained by simulation includes obtaining the simulation data by numerical calculation of the wave field using a finite difference method.

[0090] In this embodiment, the implementation of step S12 is used to simulate the underground geological structure, establish an underground velocity model, use the elastic wave equation to establish a mathematical model, and obtain simulation data by performing wave field numerical calculations using the finite difference method. Then, a simulated complete training sample can be established through the obtained simulated data, and the simulated complete training sample is used to construct a source domain data set;

[0091] In this embodiment, Figure 3 As shown in the figure, two virtual measurement points are evenly inserted between each actually deployed sensor, and a total of 85×85 sensors are deployed on a surface with a range of 120 m×120 m combined with the actual sensors.

[0092] The composition of the training data set is crucial to the success of training a neural network, and the quality of the training data set determines the separation performance of the network after training. The input and output of the deep interpolation model designed by the present invention are both seismic data, which are implicitly associated with the underlying geological model. Considering that the present invention is aimed at the reconstruction of shallow underground explosion shock wave signals, the simulation data should fully consider the complex propagation processes of the shallow underground wave field, such as reflection and transmission, and the reflection and refraction of elastic waves conform to the propagation law of underground wave fields. At the same time, the soil medium structure is consistent, and the medium velocity model can be simulated. Therefore, the use of elastic wave equations is more suitable for forward simulation of typical shallow underground media. The specific operations are as follows:

[0093] (1) Based on the underground airport runway model, a geological structure model is designed. Specifically, a layer of concrete, a layer of compacted gravel, and a layer of semi-infinite compacted clay are used for simulation. A three-dimensional space velocity matrix is ​​established to simulate the three layers of underground media. The shape and size of the matrix are (120, 30, 126). The first dimension represents the depth, represented by z, and the direction is from top to bottom; the second dimension represents the width of the ground surface, represented by y; the third dimension represents the length of the ground surface, represented by x. Each element in the matrix represents a grid with a depth of 1 meter and a length of 1 meter, which is Δh. The depth of 0-39 is set to 1000 m / s, 40-79 is set to 1500 m / s, and 80-120 is set to 2000 m / s.

[0094] (2) The source excitation function used in data simulation is set to Ricker wavelet, expressed as formula (1):

[0095]

[0096] Among them, f m represents the main frequency of the wave, usually set to 30 Hz, and t represents time. Using the finite difference method, t is replaced by kΔt and brought into the calculation, and the acoustic wave equation is used to perform numerical simulation of transverse and longitudinal waves respectively. The general form of the acoustic wave is shown in formula (2):

[0097]

[0098] In equation (2), there are second-order partial derivatives of the fluctuation u and second-order partial derivatives of the spatial coordinates x and y. For the second-order partial derivatives of the time term, the finite difference method is used to express equation (2) as equations (3)-(5) using the three-point difference format (FDM-2):

[0099]

[0100] Among them, υ(x,z) represents the speed of a discrete point in space; the subscript of u represents the index of the discrete point in space, the superscript represents the differential index in time, the k moment represents the current moment, k+1 represents the next moment, and k-1 represents the previous moment. In space, since Δx=Δy, it is commonly set to Δh=Δx=Δy, thus obtaining the fluctuation value for a certain spatial sampling point at a certain moment as formula (6):

[0101]

[0102] (3) Figure 4 As shown by the five-pointed star in the middle, the source is set at a depth of z: 60-100 with an interval of 4 meters, x direction: 40-80 with an interval of 1 meter, y: 10-20 with an interval of 10 meters; sensors are arranged between z=1, y=0, x: 5-116 with an interval of 1 meter, as shown in the figure. Figure 4 As shown by the four-pointed star in the figure, the time sampling rate is set to 0.0001 seconds, and the fluctuation time is set to 0.2048 seconds. Through the simulation formula in step (2), the fluctuation value at each sensor point can be obtained; for a sensor, the fluctuation values ​​at all times are combined to obtain the signal received by this sensor. Each source point can obtain a group (112) of sensor signals, and a total of 462 groups of sensor signals can be obtained.

[0103] S2, combining the actual data and the simulated data into a gather map, and establishing a target domain training data set and a source domain training data set based on the gather map;

[0104] As a preferred embodiment, the S2 includes:

[0105] S21, combining the actual data and the simulated data into a plurality of gather maps, including: interpolating target points and arranging virtual measurement points, assuming that the signal obtained at the virtual measurement points is 0 through explosion experiments, and combining the virtual measurement points with the actual data obtained to form a gather map;

[0106] (1) Preprocessing the actual data, analyzing the instantaneous frequency conversion characteristics of the signal, using VMD to select a threshold of 0.8 of the peak value and 8 layers to decompose the signal, retaining the useful signal and removing the noise;

[0107] (2) Taking a row of sensors with a fixed y-axis coordinate as an example, all preprocessed real signals in this row and the signals of the virtual measurement points (represented by 0) are arranged in a group according to the x-axis direction sequence, which is represented as a gather map; the vertical direction represents the complete time step, and the horizontal direction represents the sequence of deployed detectors. At this time, the shape of the gather map of this row is (2048×85). The same is true for a fixed x-axis. A total of 170 2048×85 gather maps can be obtained.

[0108] In this embodiment, the target point is interpolated. Taking the above experiment as an example, virtual measurement points are arranged. Through the explosion experiment, it is assumed that the signal obtained at the virtual measurement point is 0, and it is combined with the actual data obtained in S11 to form a track gather map.

[0109] S22, creating a target domain training dataset; including:

[0110] (1) For one of the multiple gather images, a rectangular frame is set with a shape size of (1024×64), a vertical step size of 1024, and a horizontal step size of 1. Sliding extraction is performed on the gather image to obtain 44 training samples, thereby obtaining a total of 170 gather integration images and a total of 7480 training samples; each training sample is denoted as S i , the shape is (1024×64), where represents the signal received by the jth sensor of the ith sample, as a column matrix;

[0111] (2) At this time, the column matrix of each training sample is a time domain signal received by a certain channel. In order to distinguish the channel of the actual received signal from the channel of the signal that needs to be reconstructed, the actual deployed sensor that can receive the actual vibration signal is set to 1, and the virtual sensor corresponding to the point where the signal needs to be reconstructed is set to 0. A one-dimensional mask matrix is ​​established for each sample, denoted as m i , with a length of 64, and then expanded into a two-dimensional mask matrix with a shape of (1024, 64), where i represents the sample sequence.

[0112] S23, create source domain training dataset

[0113] In this embodiment, source domain training samples are obtained based on 170 2048×85 gather images, and a total of 45,276 source domain training samples with a shape of (1024×64) are obtained. A source domain training data set is created based on the source domain training samples.

[0114] S3, training the deep interpolation model once using the source domain training data set; finally migrating the deep interpolation model after the once training to the target domain data set for secondary training in a self-supervised learning manner;

[0115] As a preferred implementation, S3 includes:

[0116] S31, building a deep interpolation model framework, wherein the deep interpolation model framework used to train the source domain training data set and the target domain training data set is the same; including:

[0117] The U-net network is used as the basic framework, which consists of 4 downsampling modules, 3 upsampling modules and three position attention modules. Figure 5 As shown in the figure, the downsampling module and the position attention module are only trained by the source domain training data set, the input sample size is 1×1024×64, firstly, the first downsampling module (downsampling module 1) is used to obtain 32×512×32 features, then the second downsampling module (downsampling module 2) is used to obtain 64×256×16 features, then the third downsampling module (downsampling module 3) is used to obtain 128×128×8 features, and then the fourth downsampling module (downsampling module 4) is used to obtain 256×64×4 features, the position The output and input sizes of the attention module are the same. The input obtained by the fourth downsampling module is input into the first upsampling module together with branch 1 to obtain 128×128×8 data, and then input into the second upsampling module together with branch 2 to obtain 64×256×16 data, and then input into the third upsampling module together with branch 3 to obtain 32×512×32 data. Finally, a deconvolution layer is passed to obtain an output of 1×1024×64. The deconvolution layer is set to a size of 4×4, a number of 1, a step size of 2, and a padding of 1.

[0118] S32, build the downsampling module, including:

[0119] like Figure 6 As shown in Figure 1, in order to avoid the degradation caused by the unsmooth forward and reverse information flow of the model during training, a residual block is added, which contains "convolution, rectified linear unit (ReLU)", and a 1×1 convolution is set at the residual connection to unify the number of channels. The input size of this module is set to C×H×W, where C is the number of channels, H is the sample length, and W is the sample width. The parameter structure is shown in Table 1, the residual block structure parameter table.

[0120] Table 1

[0121]

[0122] like Figure 7 As shown in Figure 2, in order to enable the model to capture features of different scales, multiple parallel convolution paths with kernels of different sizes are added after the residual block to form a multi-scale convolution module. The input size of the module is set to C×H×W, and the parameter structure is shown in Table 2 Multi-scale convolution module structure parameter table.

[0123] Table 2

[0124]

[0125]

[0126] S33, build upsampling module, including:

[0127] like Figure 8 As shown in Figure 3, it contains a deconvolution layer, a residual block and a multi-scale convolution block. The design of the residual block and the multi-scale convolution block is the same as the downsampling module, except that the step size of the first layer of convolution is set to 1, and the rest of the settings are the same. Taking the input size of C×H×W and the input size of C / 2×2H×2W as an example, the parameter structure of the module is shown in Table 3.

[0128] Table 3

[0129]

[0130] S34, building a position attention module. This is because in the downsampling operation, convolution cannot effectively utilize global features. The embodiment of the present invention introduces a self-attention mechanism to learn the spatial dependency of features, including:

[0131] like Fig. 9 As shown, the size of the input to the position attention module is set to C×H×W, and after passing through the first convolutional layer (convolutional layer 1), it is reorganized and transposed to obtain a data shape of N×C, where N=H×W, and then matrix multiplication is performed with the reorganized data after passing through the second convolutional layer (convolutional layer 2) and the attention score with a shape of N×N is obtained through the softmax layer, and then matrix multiplication is performed with the reorganized data after passing through the third convolutional layer (convolutional layer 3) and reorganized back to C×H×W, and finally the output features are obtained by adding element-by-element to the input.

[0132] S35, perform model training;

[0133] Since the data integrity of the source domain (simulation) and the target domain (real) is different, different training methods are adopted. The transfer learning method is used to use the created simulation dataset as the source domain training dataset for model training, and then the real dataset is used as the target domain training dataset to perform model training after model adaptive adjustment on the trained interpolation model.

[0134] include:

[0135] (1) Use the created simulated dataset as the source domain training dataset for model training, including:

[0136] Before each round of training, 45276 complete gather samples are randomly created into a one-dimensional mask matrix according to 64 seismic traces. "1" means retention and "0" means missing. The missing rate is controlled at 70%. The created 45276 masks are multiplied by the samples using the broadcast mechanism to obtain 45276 missing gather maps as network input. At the same time, the complete gather samples are used as labels. Iterate 100 rounds; take 128 batches of samples for training at a time; set the learning rate to 0.001 and multiply by 0.1 every 20 rounds; set the loss function to MSE, and use the Adam algorithm as the optimizer. In this way, the corresponding model weight parameters are obtained.

[0137] (2) Freeze the weight parameters of the four downsampling modules and the three position attention modules of the model trained on the source domain training dataset and migrate them to the target domain training dataset for self-supervised training, including:

[0138] The parameters of the model encoding part obtained by training with the source domain data set can better adapt to the characteristics of real data. To this end, when training in the target domain, the network parameters of the pre-trained model are introduced, and the network encoding part (the downsampling modules 1-4 and the three position attention modules in Figure 5) are frozen, and only the decoding part is trained. However, since the target domain data is missing and there is no complete data as a label to train the network, conventional training methods cannot be used. The embodiment of the present invention adopts a self-supervised training method, such as Fig.10 As shown, first, for a certain sample, further missing 20% ​​in the missing case (here you can use the mask representation ), the remaining gather samples are used as the input of the network. During training, 100 rounds are iterated. During the training, 128 batches of samples are taken at a time. The learning rate is set to 0.0001 and multiplied by 0.1 every 20 rounds. The loss function is given by formula (7). Different random values.

[0139]

[0140] Among them, Out i represents the output value obtained when the i-th missing sample is input into the network, “·” represents the Hadamard product, and MSE represents the mean square error loss function.

[0141] S4, interpolating virtual measuring point signals based on the second-trained deep interpolation model, thereby completing the restoration of the underground vibration field array information.

[0142] In this embodiment, after the training of the deep interpolation model is completed, the samples obtained from one explosion are re-input into the deep interpolation model that has been trained twice, and the signals belonging to the same signal in the output results of the horizontally extracted gathers are homogenized; the signals belonging to the same signal in the output results of the vertically extracted gathers are spliced ​​in time, and finally, the signals interpolated from the virtual measuring points are fitted into a smooth curve, thereby completing the repair of the array information of the underground vibration field.

[0143] Embodiment 2

[0144] like Fig.11 As shown, this embodiment provides an underground vibration field array information intelligent repair system, which is used to implement the method of embodiment 1, including:

[0145] The actual data and simulation data acquisition module 101 is used to arrange sensors and virtual measuring points, collect the vibration signals of the explosion using the sensors, and then obtain the actual data; at the same time, by simulating the three-layer underground medium, the simulation data is obtained by forward simulation of the underground explosion propagation model;

[0146] A training data set construction module 102 is used to combine the actual data and the simulated data into a gather map, and to establish a target domain training data set and a source domain training data set based on the gather map;

[0147] The model training module 103 is used to perform a primary training on the deep interpolation model using the source domain training data set; and finally migrate the deep interpolation model after the primary training to the target domain data set for secondary training in a self-supervised learning manner;

[0148] The information repair module 104 is used to interpolate virtual measuring point signals based on the second-trained deep interpolation model, thereby completing the repair of the underground vibration field array information.

[0149] The present invention also provides a memory storing a plurality of instructions, wherein the instructions are used to implement the method as in the first embodiment.

[0150] like Fig.12 As shown, the present invention also provides an electronic device, including a processor 301 and a memory 302 connected to the processor 301, the memory 302 stores a plurality of instructions, and the instructions can be loaded and executed by the processor so that the processor can execute the method as in the first embodiment.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent repair of underground vibration field array information, characterized in that: include: S1, arranging sensors and virtual measuring points, using the sensors to collect explosion vibration signals, and then obtaining actual data; at the same time, by simulating three layers of underground media, using the underground explosion propagation model forward simulation to obtain simulation data; S2, combining the actual data and the simulated data into a gather map, and establishing a target domain training data set and a source domain training data set based on the gather map; S3, training the deep interpolation model once using the source domain training data set; finally migrating the deep interpolation model after the once training to the target domain data set for secondary training in a self-supervised learning manner; S4, interpolating virtual measuring point signals based on the second-trained deep interpolation model, thereby completing the restoration of the underground vibration field array information.

2. According to claim 1, a method for intelligent repair of underground vibration field array information is characterized in that: The S1 includes: S11, arranging sensors according to requirements, and using the sensors to collect vibration signals of explosions, thereby obtaining actual data; S12, arranging virtual measuring points according to requirements, simulating three layers of underground media, and obtaining simulation data by forward simulation of an underground explosion propagation model; the underground explosion propagation model is a mathematical model established by an elastic wave equation, and the simulation data obtained by simulation includes obtaining the simulation data by numerical calculation of the wave field using a finite difference method.

3. The method for intelligent repair of underground vibration field array information according to claim 2 is characterized in that: The S12 includes: (1) A structure with one layer of concrete, one layer of compacted gravel, and one layer of semi-infinite compacted clay was used for simulation, and a three-dimensional space velocity matrix was established to simulate the three layers of underground media; (2) The source excitation function used in data simulation is set to Ricker wavelet, expressed as formula (1): Among them, f m represents the main frequency of the wave, which is usually set to 30 Hz, and t represents time. Using the finite difference method, t is replaced by kΔt and brought into the calculation, and the acoustic wave equation is used to perform numerical simulation of transverse and longitudinal waves respectively. The general form of the acoustic wave is shown in formula (2): In equation (2), there are second-order partial derivatives of the fluctuation u and second-order partial derivatives of the spatial coordinates x and y. For the second-order partial derivatives of the time term, the finite difference method is used to express equation (2) as equations (3)-(5) using the three-point difference format (FDM-2): Among them, υ(x,z) represents the speed of a discrete point in space; the subscript of u represents the index of the discrete point in space, the superscript represents the differential index in time, the k moment represents the current moment, k+1 represents the next moment, and k-1 represents the previous moment; in space, since Δx=Δy, it is commonly set to Δh=Δx=Δy, thereby obtaining the fluctuation value for a certain spatial sampling point at a certain moment as formula (6): (3) Set the location and interval of the earthquake source, set the interval of the sensor, and set the time sampling rate and fluctuation time at the same time, so as to obtain the fluctuation value at the sensor point at each moment; for a sensor, combine the fluctuation values ​​at all moments to obtain the signal received by this sensor.

4. The method for intelligently repairing underground vibration field array information according to claim 3 is characterized in that: The S2 includes: S21, combining the actual data and the simulated data into a plurality of gather maps, including: interpolating target points and arranging virtual measurement points, assuming that the signal obtained at the virtual measurement points is 0 through explosion experiments, and combining the virtual measurement points with the actual data obtained to form a gather map; (1) Preprocessing the actual data, analyzing the instantaneous frequency conversion characteristics of the signal, using VMD to select the peak value and number of layers of the threshold to decompose the signal, retaining the useful signal and removing the noise; (2) Taking a row of sensors with a fixed y-axis coordinate as an example, all pre-processed real signals in this row and the signals of the virtual measurement points represented by 0 are arranged in a group according to the x-axis direction sequence, which is represented as a gather map; the vertical direction represents the complete time step, and the horizontal direction represents the sequence of the deployed detectors, and all gather maps are obtained; S22, creating a target domain training dataset; including: (1) For one of the multiple gather maps, set the shape and size, vertical step length and horizontal step length of a rectangular frame, perform sliding extraction on the gather map, and obtain multiple training samples, thereby obtaining all gather integration maps and all training samples; each training sample is recorded as S i , the shape is (1024×64), where represents the signal received by the jth sensor of the ith sample, as a column matrix; (2) At this time, the column matrix of each training sample is a time domain signal received by a certain channel. In order to distinguish the channel of the actual received signal from the channel of the signal that needs to be reconstructed, the actual deployed sensor that can receive the actual vibration signal is set to 1, and the virtual sensor corresponding to the point where the signal needs to be reconstructed is set to 0. A one-dimensional mask matrix is ​​established for each sample, denoted as m i , with a length of 64, and then expanded into a two-dimensional mask matrix with a shape of (1024, 64), where i represents the sample sequence; S23, creating a source domain training data set, including: obtaining source domain training samples from all gather images, all the obtained samples are source domain training samples, and creating a source domain training data set based on the source domain training samples.

5. The method for intelligent repair of underground vibration field array information according to claim 4, characterized in that: The S3 includes: S31, building a deep interpolation model framework, wherein the deep interpolation model framework used for training the source domain training data set and the target domain training data set is the same; the deep interpolation model framework is based on a U-net network and consists of 4 downsampling modules, 3 upsampling modules and three position attention modules, wherein the downsampling modules and the position attention modules are only trained by the source domain training data set; S32, building a downsampling module, including: a downsampling residual block and a downsampling multi-scale convolution block, wherein the downsampling residual block includes multiple two-dimensional convolution layers and a rectified linear unit (ReLU), and multiple parallel convolution paths with kernels of different sizes are added after the downsampling residual block to form a multi-scale convolution module; S33, building an upsampling module, including: a deconvolution layer, an upsampling residual block and an upsampling multi-scale convolution block, wherein the upsampling residual block and the upsampling multi-scale convolution block are designed the same as the downsampling module, except that the step size of the first layer of convolution is set to 1, and the other settings are the same; S34, building a position attention module, including: setting the size of the input to the position attention module to C×H×W, reorganizing and transposing after passing through the first convolution layer to obtain a data shape of N×C, where N=H×W, performing matrix multiplication with the data reorganized after passing through the second convolution layer and obtaining an attention score with a shape of N×N through a softmax layer, performing matrix multiplication with the data reorganized after passing through the third convolution layer and reorganizing back to C×H×W, and finally adding element by element with the input to obtain an output feature; S35, performing model training, including: using a transfer learning method to perform model training using the created simulated data set as a source domain training data set, and then using the real data set as a target domain training data set to perform model training after model adaptive adjustment on the trained interpolation model.

6. The method for intelligently repairing underground vibration field array information according to claim 5, characterized in that: The S35 includes: (1) Use the created simulated dataset as the source domain training dataset for model training, including: Before each round of training, all complete gather samples are randomly created into a one-dimensional mask matrix according to a certain number of seismic traces. "1" means retention and "0" means missing. The missing rate is controlled at 70%. All the created masks are multiplied by the samples using the broadcast mechanism to obtain the corresponding number of missing gather maps as network input. At the same time, the complete gather samples are used as labels. Iterate 100 rounds. During the process, 128 batches of samples are taken for training at a time. The learning rate is set to 0.001 and multiplied by 0.1 every 20 rounds. The loss function is set to MSE, and the optimizer uses the Adam algorithm to obtain the corresponding model weight parameters. (2) Freeze the weight parameters of the four downsampling modules and the three position attention modules of the model trained on the source domain training dataset and migrate them to the target domain training dataset for self-supervised training, including: When training in the target domain, the network parameters of the pre-trained model are introduced, and the weight parameters of the four downsampling modules and the three position attention modules are frozen. The self-supervised training method is used for training, including: For a certain sample, further missing The remaining gather samples are used as the input of the network. During training, 100 rounds are iterated. During the training, 128 batches of samples are taken at a time. The learning rate is set to 0.0001 and multiplied by 0.1 every 20 rounds. The loss function is given by formula (7). They are all different random values; Among them, Out i represents the output value obtained by inputting the i-th missing sample into the network, "·" represents the Hadamard product, and MSE represents the mean square error loss function.

7. The method for intelligently repairing underground vibration field array information according to claim 6, characterized in that: After the training of the deep interpolation model is completed, the samples obtained from one explosion are re-input into the deep interpolation model that has been trained twice, and the same signal in the output results of the horizontally extracted gathers is homogenized; the same signal in the output results of the vertically extracted gathers is spliced ​​in time, and finally, the signal interpolated from the virtual measuring point is fitted into a smooth curve, thereby completing the repair of the array information of the underground vibration field.

8. An underground vibration field array information intelligent repair system, used to implement any of the methods of claims 1-7, characterized in that: include: The actual data and simulation data acquisition module (101) is used to arrange sensors and virtual measuring points, collect explosion vibration signals using the sensors, and obtain actual data; at the same time, simulate the three-layer underground medium and obtain simulation data using the underground explosion propagation model forward simulation; A training data set construction module (102) is used to combine the actual data and the simulated data into a gather map, and to establish a target domain training data set and a source domain training data set based on the gather map; A model training module (103) is used to perform a primary training on the deep interpolation model using the source domain training data set; and finally migrate the deep interpolation model after the primary training to the target domain data set for secondary training in a self-supervised learning manner; The information repair module (104) is used to perform interpolation of virtual measurement point signals based on the depth interpolation model that has undergone the secondary training, thereby completing the repair of the underground vibration field array information.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is used to read the instructions and execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor to execute the method according to any one of claims 1 to 7.

Images