Intelligent repair method for underground seismic field measurement data based on diffusion model

Abstract

The present application relates to a kind of underground seismic field measurement data intelligent repair method based on diffusion model, first using small equivalent ammunition underground shallow explosion experiment scheme to obtain seismic field signal, signal is denoised by VMD algorithm, improve the signal-to-noise ratio of seismic field signal;True sample is obtained by 4 times small equivalent ammunition underground shallow explosion experiment, ensure the effectiveness of sample, reduce training complexity, improve model reliability;The missing common shot gather is used as the input sample of diffusion model, different conditions are simulated by changing random mask to simulate missing condition, increase the diversity of input sample, improve model generalization, ensure that even under the condition of experimental data missing, there is still higher data diversity;Diffusion model used in the present application is all to whole gather in the process of adding noise and denoising, so that the model can better extract global features, and is more stable to different missing conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data repair, and particularly relates to an underground seismic field measurement data intelligent repair method based on a diffusion model. BACKGROUND

[0002] The underground explosion seismic field is a seismic signal obtained by using a limited node to reconstruct the intensity distribution of the explosion seismic load in the test area and the distribution over time in an inversion manner. The parameter is a key technical index for evaluating the whole process of the damage effect of penetration ammunition on the underground target terminal.

[0003] In the underground seismic field reconstruction process, the effectiveness of the test data directly affects the reconstruction accuracy. Researchers at home and abroad mainly improve the effectiveness of the test data through two methods. (1) The data obtained by using a limited node is repaired and compensated in an interpolation manner (such as nonlinear interpolation, radial basis function, and an empirical formula of a shock wave) for the data of invalid measuring points; and (2) a deep learning method (such as an encoder-decoder network and an inflation convolutional neural network, combined with a shock wave sample) is used to repair the partially missing data obtained by a single node.

[0004] With the continuous improvement of large-power ammunition technology, the following problems still exist in the underground seismic field measurement of large-equivalent ammunition at present.

[0005] 1. With the increase of ammunition power, the survival ability of the test instrument is insufficient due to environmental factors such as fragments and high temperature, the data effectiveness is insufficient, and the propagation law is not clear due to the complexity of the underground medium, resulting in a large error in the interpolation algorithm in data repair;

[0006] 2. The instrument has the possibility of being directly scrapped, and the data of single-point measurement will be completely missing, which reduces the experimental success probability and the reliability of the experimental results. SUMMARY

[0007] The application provides an underground seismic field measurement data intelligent repair method based on a diffusion model, which solves the above technical problems.

[0008] In order to solve the above technical problems, the application provides an underground seismic field measurement data intelligent repair method based on a diffusion model, characterized in that the specific steps are as follows:

[0009] S1: Obtain small-equivalent shallow underground explosion experimental signal data;

[0010] S2: Preprocess and integrate the data obtained in S1;

[0011] S3: Generate training samples by using a mask;

[0012] S4: Take the S3 generated training sample as input data, add noise to become pure Gaussian noise, and generate repaired data through the supervised denoising process to obtain a diffusion model;

[0013] S5: Generate large equivalent underground shallow explosion missing trace data by using the diffusion model.

[0014] Beneficial effects: The specific advantages are as follows:

[0015] 1. According to the explosion similarity law and the homologous matching principle, the invention first uses a small equivalent ammunition underground shallow explosion experiment scheme to obtain a vibration field signal, denoises the signal through a VMD algorithm, improves the signal-to-noise ratio of the vibration field signal, obtains real samples through four small equivalent ammunition underground shallow explosion experiments, ensures the effectiveness of the samples, reduces the training complexity, and improves the reliability of the model;

[0016] 2. The invention takes the missing common shot gather as the input sample of the diffusion model, simulates the missing conditions under different conditions by changing the random mask, increases the diversity of the input sample, improves the model generalization, and ensures that even under the condition of missing experimental data, there is still high data diversity;

[0017] The diffusion model used by the invention processes the entire gather in the noise adding and denoising process, so the model can better extract global features and is more stable for different missing conditions (all missing, partial missing). BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 Schematic diagram of underground shallow explosion experiment

[0019] Figure 2 Schematic diagram of diffusion model

[0020] Figure 3 Schematic diagram of attention UNet network model. DETAILED DESCRIPTION

[0021] In order to make the purpose, content and advantages of the invention clearer, the specific embodiments of the invention are described in further detail below.

[0022] The underground vibration field measurement data intelligent repair method based on the diffusion model proposed by the invention has the following specific steps:

[0023] S1: Obtain small equivalent shallow underground explosion experiment signal data;

[0024] For example Figure 1As shown, the number of detectors is N (for example, N is equal to 50), and 1, 2, 3 and 4 underground shallow explosion experiments with equivalent TNT of 5, 10, 30 and 60 kg are designed. Four small equivalent TNTs are arranged at the four corners of the experimental area to reduce the influence of underground medium changes on the collected signals and reduce the engineering workload. The measured large equivalent TNT can be arranged in the experimental area. Experiment 1 is an experiment of 10 kg equivalent TNT explosion at (0, 50, 50), experiment 2 is an experiment of 10 kg equivalent TNT explosion at (0, 0, 50), experiment 3 is an experiment of 30 kg equivalent TNT explosion at (110, 0, 50), and experiment 4 is an experiment of 60 kg equivalent TNT explosion at (110, 60, 50).

[0025] Taking the 5 kg TNT equivalent experiment as an example, the horizontal plane is the xoy plane, the positive direction of the x axis is to the right, the positive direction of the y axis is inward, and the positive direction of the z axis is downward to establish a three-dimensional rectangular coordinate system with the lower left corner of the experimental area as the origin. The length of the area is 110 meters, the width is 60 meters, and the 5 kg TNT explosive is placed at (0, 50, 50). The detector arrangement uses a 5x10 array, and the detectors are labeled. The detector arranged at (10, 10, 0) is No. 1 detector, and the detector label increases by one every 10 meters along the positive direction of the x axis. With No. 1 to No. 10 detectors as a reference, the detector label increases by ten every 10 meters in the positive direction of the y axis. Correspondingly, Figure 1 The leftmost lower corner of the area is No. 1 detector, the rightmost lower corner is No. 10 detector (100, 10, 0), the leftmost upper corner is No. 41 detector (10, 50, 0), and the rightmost upper corner is No. 50 detector (100, 50, 0). The sampling rate of the detector is set to 20 kHz for 5 seconds. The 5 kg TNT is detonated to complete experiment 1.

[0026] S2: Preprocessing the data obtained by S1 and performing trace set integration;

[0027] The underground explosion experiment is complex, and the data obtained by the experiment often has serious background noise. Therefore, the data obtained by the experiment needs to be preprocessed. In order to facilitate subsequent model training and prediction, the preprocessed data needs to be integrated into a trace set. The present application adopts common shot point trace set integration.

[0028] S2.1 Data preprocessing

[0029] First, the data collected by each detector is denoised by VMD to improve the signal-to-noise ratio of the data.

[0030] S2.2 Common shot point trace set integration

[0031] The detector number {1, 2,..., 49, 50} is regarded as a trace number, and the trace data is arranged from left to right according to the trace number from small to large. In experiment 1, the common shot gather is represented as S1, S1 The data size of the common shot gather is 100k*50, and the row and column of S1 represent the sampling of data at different times and different trace numbers.

[0032] After the data processing of experiment 1 is completed, the data preprocessing and common shot gather integration of experiments 2, 3 and 4 are repeated to complete the S2.2 step, and three small equivalent common shot gathers S2, S3 and S4 are obtained.

[0033] S3: generating training samples using a mask;

[0034] According to the existing common shot gathers S1, S2, S3 and S4, the data size of each common shot gather is 100k*50, and at this time, the common shot gathers S1, S2, S3 and S4 are complete data. To simulate the trace missing condition under the condition of large equivalent, the common shot gathers S1, S2, S3 and S4 are missing, the minimum missing is 1 trace, and the maximum missing is 25 traces. When the traces are missing, the missing trace label information needs to be recorded at the same time to facilitate subsequent training of the network model.

[0035] Taking S1 as an example (the same as the rest of the common shot gathers), suppose that S1 is missing k traces, there are k = 1, the survival missing trace common shot gather; when k 1, the present application randomly selects k traces for missing and records the position (detector label or position) of the missing trace as a mask M, M is the same shape as S1, the missing trace position is 0, and the non-missing trace position is 1, and there are 100 times. The common shot gather after missing does not change the trace number compared with S1, only the corresponding data becomes 0. Similarly, repeat S2, S3 and S4 to complete the above missing trace operation, and generate 4*(50+100*24) = 9800 common shot gathers of missing partial trace data, wherein 4 represents the number of complete trace sets; 50 represents the number of missing trace sets generated by missing 1 trace, and 100*24 represents the number of missing trace sets generated by missing 2 to 25 traces, and the missing is 100 times each time.

[0036] S4: training a diffusion model;

[0037] The training sample generated by S3 is used as input data, noise is added to become pure Gaussian noise, and then the repaired data is generated through the process of supervised denoising to obtain a diffusion model; the diffusion model includes a forward diffusion module, a reverse diffusion module and a data synthesis module; the diffusion model is a kind of generative model, which uses the pure noise data obtained by continuously adding noise to denoise, thereby generating samples.

[0038] The process of adding noise to the original data to become pure Gaussian noise is called the forward diffusion process, and the process of generating data from pure Gaussian noise by denoising is called the reverse diffusion process. Figure 2 A general schematic diagram of the diffusion model used by the present application is shown, Figure 2 The model shown mainly includes a forward diffusion module, a reverse diffusion module and a data synthesis module.

[0039] S4.1 Forward diffusion module

[0040] The forward diffusion module is a process of gradually adding Gaussian noise to the input data, so that the input data gradually becomes pure Gaussian noise. Let the total time step be T (usually a known hyperparameter), and the variance of the Gaussian noise added from the first to the Tth step be β1, β2,..., β T-1 ,β T (usually a known hyperparameter), then at the tth step, the added Gaussian noise ε t is as shown in equation (1):

[0041]

[0042] Let the data after adding T-step noise be x T , then x T belongs to the standard Gaussian distribution, and I is the identity matrix, and N() is the normal distribution.

[0043] S4.2 Reverse diffusion module

[0044] The reverse diffusion module is a process of restoring the standard Gaussian noise generated by the forward diffusion module. The reverse diffusion process defines a learnable Gaussian Markov chain starting from p(x T ) = N(x T ; 0, I). Reverse diffusion usually uses a neural network to fit the mapping relationship between the pure Gaussian noise data x T and the generated data, and the neural network infers the original data according to the pure Gaussian noise data generated in the forward diffusion. The neural network uses an Attention UNet model.

[0045] The present application uses an Attention UNet model to fit the probability distribution of the reverse diffusion process, and the network model is as shown in Figure 3 The reverse diffusion model essentially fits the distribution of t to t-1 step data by training the network, and the objective function of the network model can be simplified as equation (2)

[0046]

[0047] Where ε is the noise added to the standard Gaussian, and ε θ is the noise predicted by the model, Alpha t = 1 - Beta t Beta i represents the variance of the Gaussian noise added at the i-th step, E(.) represents the expectation, and the diffusion model generates data by constantly iterating the supervised noise.

[0048] The number of diffusion steps in the application is set to 1000, and the variance of the added noise can also be used as a parameter for training the reverse diffusion model. Related research shows that setting the noise variance as a constant can also achieve good results, so the noise addition scheme uses a linear noise addition scheme from 0.0001 to 0.02 to reduce the training difficulty and make the network more stable. The input data size is 100k x 50, the batch size is set to 1, the network weight is randomly initialized, the loss function is calculated, and the optimizer uses the Adam optimizer. Attention Unet belongs to the network of the encoding and decoding structure, and introduces the attention mechanism on the basis of UNet. Before the corresponding features in the encoder and decoder are spliced, an attention module is used to adjust the output features of the encoder. The attention module is used to control the importance of the features at different spatial positions. The module is combined with ReLU and Sigmoid through 1x1x1 convolution, respectively, to generate a weight map. The weight map is multiplied by the features in the encoder to correct the input of the module.

[0049] The Attention Unet network structure mainly consists of a normal Unet and an attention module. UNet changes the channel number of data features through a series of up-sampling and down-sampling modules. The UNet model mainly uses the input data feature dimension to change from 1 channel to 64 channels, then from 64 channels to 128 channels, and then from 128 channels to 256 channels. Then, the output data channel number is restored to 1 in the order of 256 to 128 to 64 to 1; the UNet input layer uses a double convolution module, and the down-sampling operation is a maximum pooling with a size of 2 followed by a double convolution module; the up-sampling uses a double linear interpolation with a scale factor of 2 followed by a double convolution module; the UNet output layer uses a double convolution module with a convolution kernel size of 1.

[0050] The input of the attention module includes the query signal and the encoder feature. The query signal is composed of the output of the multi-channel convolution module below the current attention mechanism, and the skip link of the attention mechanism constitutes the encoder feature. The attention module passes the query signal and the encoder feature through different 1x1x1 convolution modules, then enters the ReLU activation function, then enters a 1x1x1 convolution module, then enters the Sigmoid activation function, and then resamples to obtain a weight map. Multiplying the weight map with the encoder feature is the output of the attention module.

[0051] S4.3 The data synthesis module obtains the synthesized data by weighting the data acquired by S2 and the generated data output by the inverse diffusion model using the mask M and the inverse mask invM respectively. The synthesized data is input into the inverse diffusion module at time steps T-1 to 1 in the training phase.

[0052] The function of the mask is to record the position of the missing trace, and the detector number or position of the missing data is marked, which is set as the mask M. Taking Experiment 1 as an example, the shape of M is 100k x 50, and the missing trace positions are detector numbers 15 and 20. The 15th and 20th columns of M are all 0, and the rest are all 1. Similarly, taking the above example, let the inverse mask be invM, the 15th and 20th columns of invM are all 1, and the rest are all 0. When generating data, M and invM are used to weight the acquired data and the generated data to synthesize data.

[0053] S5. Generating missing trace data of large equivalent underground shallow explosion by using the diffusion model.

[0054] S5.1 Signal acquisition

[0055] In the area as shown in Figure 1 , a 100 kg large equivalent TNT underground shallow explosion experiment is constructed, and the shot position (75, 25, 50) is set in the area of the detector.

[0056] S5.2 Signal preprocessing

[0057] VMD is used for signal preprocessing to improve the signal-to-noise ratio of the vibration signal.

[0058] S5.3 Automatic detection of missing vibration signals

[0059] The Mahalanobis distance between the vibration signals collected by different detectors is calculated to analyze the consistency of the vibration signals. The calculation formula of the Mahalanobis distance is as follows:

[0060]

[0061] where W is the correlation matrix of p and q, which is a positive definite matrix, w ij is the element in the i-th row and j-th column of the matrix W -1 , p i represents the signal amplitude of the effective signal at time i, q j represents the amplitude of the data collected by the detector to be detected at time j, and T *The number of sampling points is. The value range of the Mahalanobis distance correlation coefficient is [0, 1], the smaller the Mahalanobis distance, the more consistent the two variables are, when the Mahalanobis distance is equal to 1, the two variables are independent of each other, and when the Mahalanobis distance is equal to 0, the two variables are completely correlated. The Mahalanobis distance threshold value is set to 0.8 in the present application, an effective vibration signal data is selected as a reference signal, and the Mahalanobis distance between the to-be-measured signal and the reference signal is calculated. When the Mahalanobis distance is less than 0.8, the signal is judged as an effective signal.

[0062] S5.4 Common shot gather integration

[0063] The common shot gather integration is performed in the same manner as S2.2, and the positions of the effective signals and the invalid signals are recorded to form a mask M and an inverse mask invM.

[0064] S5.5 Inference of missing data using diffusion model

[0065] The gather obtained in S5.4 is subjected to forward diffusion to generate a pure Gaussian noise image x T , M and invM are used to retain or remove the Gaussian noise at the corresponding positions in x T according to the positions of the effective data and the invalid data, a trained neural network model is used to denoise the Gaussian noise, and a reverse iteration process is performed according to the time step t from T in a decreasing iteration manner. When t is 1, the output data after being combined by the data combination module is the final repaired data.

[0066] According to the explosion similarity law, the homologous matching principle, and the advantages of artificial intelligence in data prediction, the present application constructs an underground shallow explosion experiment scheme to obtain a vibration signal, uses a variational mode decomposition (VMD) algorithm to denoise and preprocess the vibration signal, integrates the common shot gather of the vibration signal generated by small equivalent ammunition, uses a random mask to randomly delete the common shot gather, thereby generating training samples of the model, continuously adds noise to these samples until the samples are completely noisy, and then trains an Attention Unet network model. The training of the model is performed by supervising the noise adding process of the samples. After the training is completed, the vibration signal obtained by the large equivalent underground shallow explosion experiment scheme is denoised and preprocessed using VMD, signal consistency analysis is performed according to the similarity measurement index Mahalanobis distance, the missing vibration signal is detected, if there is, the common shot gather of the vibration array signal is integrated, and the gather is input into the network model to predict the missing trace, thereby realizing intelligent repair of the measurement data.

[0067] The above only describes the preferred embodiments of the present application, and it should be noted that those skilled in the art can make several improvements and modifications without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A method for intelligent repair of underground seismic field measurement data based on a diffusion model, characterized in that, The specific steps are as follows: S1: Acquire signal data from low-yield shallow underground explosion experiments; S2: Preprocess the data obtained from S1 and integrate the data sets; S3: Generate training samples using a mask; S4: The training samples generated in S3 are used as input data. Noise is added to them to make them pure Gaussian noise. Then, the repaired data is generated through a supervised denoising process to obtain the diffusion model. S5: Generate missing channel data for high-yield shallow underground explosions using a diffusion model. The specific steps are as follows: S5.1 Signal Acquisition In the experimental area of ​​step S1, a large-yield TNT shallow underground explosion experiment is constructed, and the explosion point is set so that it is located in the area where the detector is located. S5.2 Preprocess the signal; S5.3 Automatic detection of missing vibration signals To perform a consistency analysis of the vibration signals, the Mahalanobis distance between the vibration signals acquired by different detectors is calculated. The formula for calculating the Mahalanobis distance is as follows: (3) in, for The correlation matrix is ​​a positive definite matrix. For matrix The element in the i-th row and j-th column, This represents the amplitude of the effective signal at time i, and is based on empirical theoretical data. This represents the amplitude of the data collected by the detector under test at time j. Set the number of sampling points; set a Mahalanobis distance threshold, select a valid vibration signal data as a reference signal, calculate the Mahalanobis distance between the signal to be measured and the reference signal, and determine that the signal is a valid signal when the Mahalanobis distance is less than the threshold. S5.4 Common Artillery Point Gathering Integration Perform common shot gather integration, and record the positions of valid and invalid signals to form a mask. with inverse mask ; S5.5 Using diffusion models to infer missing data The gather obtained in S5.4 is forward diffused to generate a pure Gaussian noise image. ,use and from The Gaussian noise at the corresponding positions is retained or removed based on the positions of valid and invalid data. The Gaussian noise is denoised using a trained neural network model. The process is reversed by iterating from T in descending order of time step t. When t is 1, the output data is combined into modules to obtain the final repaired data.

2. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 1, characterized in that, In S1, the number of detectors deployed in the experiment is N. A total of 5, 10, 30 and 60 kg TNT equivalent shallow underground explosion experiments 1-4 are designed. The four small-yield TNTs are placed at the four corners of the experimental area, and the large-yield TNT to be tested is placed in the experimental area. Experiment 1 is an experiment of 10 kg TNT equivalent exploding in area (0, 50, 50), Experiment 2 is an experiment of 10 kg TNT equivalent exploding in area (0, 0, 50), Experiment 3 is an experiment of 30 kg TNT equivalent exploding in area (110, 0, 50), and Experiment 4 is an experiment of 60 kg TNT equivalent exploding in area (110, 60, 50).

3. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 1, characterized in that, The specific steps for S2 are as follows: S2.1 Data Preprocessing: Perform VMD noise reduction on the data acquired by each detector; S2.2 Common shot gather integration: The detector serial number is considered as the channel number, and the channel data is arranged from left to right according to the channel number from smallest to largest; let the common shot gather in Experiment 1 be represented as... , , The rows and columns represent data sampling at different times and different channel numbers; After processing the data from Experiment 1, preprocessing and integrating the common shot gathers for Experiments 2, 3, and 4 were completed to obtain three more low-yield common shot gathers. , , .

4. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 3, characterized in that, In S3, the existing common artillery point gathering will be... , , , The number of missing channels is 1 to 25. When missing channels, the label information of the missing channels must be recorded to facilitate subsequent training of the network model.

5. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 4, characterized in that, S3 specifically refers to, let... Missing k channels total In this case, when At that time, survival A missing road, a common artillery point road set; when At that time, k channels are randomly selected for deletion, and the positions of the deleted channels are recorded and set as a mask. , and The shapes are the same, the missing path position is recorded as 0, and the non-missing path position is recorded as 1, for a total of 100 times; the common shot point gather after the missing path is in the same shape as the missing path position. The channel number remains unchanged, only the corresponding data becomes 0; the same applies. , , After completing the above missing trajectories operation, 4×(50+100×24)=9800 common shot point trajectories with missing partial trajectories are generated, where 4 represents the number of complete trajectories; 50 represents the number of missing trajectories generated by missing 1 trajectories; and 100×24 represents the number of missing trajectories generated by missing 2 to 25 trajectories each time by randomly missing 100 times.

6. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 1, characterized in that, In S4, the diffusion model includes a forward diffusion module, a reverse diffusion module, and a data synthesis module; The forward diffusion module gradually adds Gaussian noise to the input data, making the input data gradually become pure Gaussian noise; the reverse diffusion module is the process of restoring the standard Gaussian noise generated by the forward diffusion module. The data synthesis module utilizes a mask. with inverse mask The data obtained by S2 and the generated data output by the backdiffusion model are weighted to obtain synthetic data. This synthetic data is used as the input of the backdiffusion module from time T-1 to 1 during the training phase, following the time step iteration.

7. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 6, characterized in that, For the forward diffusion module, let the total number of time steps be T, and the variance of the Gaussian noise added from step 1 to step T be... Then, at step t, the added Gaussian noise As shown in equation (1): (1) Let the data after adding T-step noise be ,but It belongs to the standard Gaussian distribution, I is the identity matrix, N() is the normal distribution, T, These are known hyperparameters.

8. The intelligent repair method for underground seismic field measurement data based on a diffusion model according to claim 6, characterized in that, For the back diffusion module, a neural network is used to fit the pure Gaussian noise data. The neural network infers the original data based on the pure Gaussian noise data generated during forward diffusion, and uses the Attention UNet model to fit the probability distribution of the backward diffusion process. The objective function of this network model is simplified to equation (2). （2） in To add standard Gaussian noise, It is noise in the model prediction. , , This represents the variance of adding Gaussian noise in step i. The expectation is expressed, and the diffusion model generates data by continuously iterating through supervised noise.

Images