Wave field separation method based on combination of curvelet transformation and deep learning

Abstract

The invention provides a curvelet transform and deep learning combination-based wave field separation method, which is applied to the field of seismic exploration. The method comprises the following steps: (1) carrying out curvelet transformation, and then carrying out inverse transformation to obtain an inverse transformation diagram of each angle of each scale; (2) curvelet classification is carried out, the average gradient of each inverse transformation diagram is calculated by using a Sobel operator, the inverse transformation diagrams with the positive average gradient are accumulated, the inverse transformation diagrams with the negative average gradient are accumulated, the accumulated diagrams with the positive gradients are input into a downlink wave encoder, and the accumulated diagrams with the negative gradients are input into an uplink wave encoder; (3) constructing a loss function which mainly comprises reconstruction loss and gradient loss and is used for constraining uplink and downlink wave features extracted by the convolutional auto-encoder; according to the invention, through curvelet preprocessing, the characteristic space of the uplink wave encoder and the characteristic space of the downlink wave encoder are compressed, and high-precision separation of the vertical seismic profile wave field is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic exploration, and particularly to a wave field separation method based on the combination of curvelet transform and deep learning. Background Art

[0002] Vertical seismic profile technology is an important seismic exploration method, which plays a key role in the fine imaging of underground structures and reservoir description. However, vertical seismic profile data often contains multiple wave fields such as up and down traveling waves, and the aliasing of these wave fields seriously affects the data quality and the effect of subsequent processing. Therefore, how to efficiently and accurately separate the up and down wave fields and extract the effective wave field is a core challenge in vertical seismic profile technology.

[0003] Curvelet transform is a signal processing method based on multi-scale and multi-directional decomposition. In recent years, it has been widely used in the field of seismic data processing, especially in wave field separation. Curvelet transform can effectively separate wave fields with different propagation directions through its good directionality and sparse representation ability. The prior art uses the generalized Curvelet transform to extract the features of the vertical seismic profile wave field, constructs a feature dataset integrating phase space angle information, and then uses the K-means clustering algorithm to classify the wave field features in the phase space to achieve the adaptive separation of the vertical seismic profile wave field. However, due to the complexity of seismic data, curvelet transform still has deficiencies when dealing with complex underground structures or high wave field overlap situations. It is difficult to overcome the "false axis" caused by the Fourier transform mechanism, so cumbersome data preprocessing or subsequent processing is inevitable.

[0004] With the rapid development of artificial intelligence technology, deep learning has shown great potential in the field of seismic exploration. Deep learning can automatically extract complex features in data and effectively capture non-linear relationships by establishing an end-to-end learning model. Especially in the wave field separation task, deep learning can achieve the efficient separation of complex wave field signals and significantly improve the separation accuracy. The prior art starts from the axial slope difference of the vertical seismic profile wave field and uses a modified Sobel operator to propose a morphological prior-guided traveling wave separation method. However, the operation stability of this method is poor, and a cumbersome parameter tuning process is required to determine the appropriate loss weight, which increases the labor cost of algorithm implementation. Therefore, combining curvelet transform with deep learning can make full use of the sparse representation ability of curvelet transform and the non-linear feature extraction ability of deep learning, so as to construct an efficient wave field separation method to solve the deficiencies in the prior art. Summary of the Invention

[0005] In view of the above problems in the prior art, the present application proposes a wave field separation method based on the combination of curvelet transform and deep learning. The specific steps are as follows:

[0006] S1: Perform curvelet transform on the vertical seismic profile data to be separated to obtain the curvelet coefficients of the vertical seismic profile data to be separated;

[0007] S2: Perform inverse curvelet transform to obtain the inverse transform map for each angle at each scale;

[0008] S3: Perform curvelet classification. Calculate the average gradient of each inverse transform map using the Sobel operator, and accumulate those with positive average gradient and accumulate those with negative average gradient;

[0009] S4: Construct a double convolutional autoencoder network, including a downgoing wave autoencoder and an upgoing wave autoencoder;

[0010] S5: Determine the loss function, which includes reconstruction loss and gradient loss;

[0011] S6: Input the accumulated map with positive average gradient into the downgoing wave encoder, input the accumulated map with negative average gradient into the upgoing wave encoder, and start training;

[0012] S7: Save the model obtained in S6, and use this model to separate the downgoing and upgoing wave fields of the vertical seismic profile data to be separated.

[0013] In one embodiment, in step S1, the angle of the curvelet transform is selected as 5, and the scale of the second coarsest angle is selected as 16.

[0014] In one embodiment, in step S4, each autoencoder includes seven convolutional layers. The first three convolutional layers are used to construct the encoding layer of the neural network, and a max pooling layer is used during the encoding process to expand the receptive field of the neural network. The last four convolutional layers are used to construct the decoding process of the neural network, and an upsampling layer symmetric to the pooling layer is used during the decoding process to simulate the inverse process of pooling.

[0015] In one embodiment, each convolutional layer performs two-dimensional convolutional operations with a 3*3 kernel: max pooling uses a 2*2 kernel: and the activation function uses the Tanh function.

[0016] In one embodiment, in step S2, traverse each angle at each scale, set the remaining angle coefficients of this angle to zero, and perform inverse curvelet transform.

[0017] In one embodiment, step S3 further includes preprocessing the vertical seismic profile data and normalizing it between -1 and 1.

[0018] The above technical features can be combined in various suitable ways or replaced by equivalent technical features as long as the object of the present invention can be achieved.

[0019] A wave field separation method based on the combination of curvelet transform and deep learning provided by the present invention has at least the following beneficial effects compared with the prior art:

[0020] Through curvelet preprocessing, the present invention compresses the feature spaces of the upward wave encoder and the downward wave encoder while achieving high-precision separation of the vertical seismic profile wave field. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be described in more detail below based on embodiments and with reference to the drawings. Among them:

[0022] Figure 1 is a flowchart of the wave field separation method based on the combination of curvelet transform and deep learning provided by the embodiment of the present invention;

[0023] Figure 2 is a schematic diagram of the double convolutional autoencoder network provided by the example of the present invention;

[0024] Figure 3 is a model data separation diagram provided by the embodiment of the present invention;

[0025] Among them Figure 3 (a) of which is the validation set data, Figure 3 (b) of which is Figure 3 (c) of which and Figure 3 (d) of which are added together, Figure 3 (c) of which is the downward wave separated from the validation set data, Figure 3 (d) of which is the upward wave separated from the validation set data.

[0026] Figure 4 is a comparison diagram of the separation results of each method under the model data provided by the present invention;

[0027] Among them Figure 4 (a) of which is the up and down wave fields separated by the FK filtering method, Figure 4 (b) of which is the up and down wave fields separated by the curvelet transform, Figure 4 (c) of which is the up and down wave fields separated based on the present invention.

[0028] Figure 5 is the F-K spectrum of the model data separation diagram provided by the embodiment of the present invention;

[0029] Among them Figure 5 (a) of which is the F-K spectrum of the validation set data, Figure 5 (b) of which is the F-K spectrum of the downward wave separated from the validation set data, Figure 5 (c) of which is the F-K spectrum of the upward wave separated from the validation set data;

[0030] Figure 6It is a diagram of complex model data separation provided by an embodiment of the present invention;

[0031] Among them Figure 6 (a) of which is complex model data, Figure 6 (b) of which is Figure 6 the addition of (c) of Figure 6 and (d) of Figure 6 (c) of which is the down-going wave separated from the complex model data, Figure 6 (d) of which is the up-going wave separated from the complex model data;

[0032] Figure 7 It is a diagram of actual vertical seismic profile data separation from a certain area in Tarim provided by an embodiment of the present invention;

[0033] Among them Figure 7 (a) of which is the actual data, Figure 7 (b) of which is Figure 7 the addition of (c) of Figure 7 and (d) of Figure 7 (c) of which is the down-going wave separated from the actual data, Figure 7 (d) of which is the up-going wave separated from the actual data. Specific implementation manner

[0034] The present invention will be further described below with reference to the accompanying drawings.

[0035] As Figure 1 shown, a wavefield separation method based on the combination of curvelet transform and deep learning provided by the present invention includes the following steps:

[0036] S1: Perform curvelet transform on the vertical seismic profile data to be separated to obtain the curvelet coefficients of the vertical seismic profile data to be separated;

[0037] In one embodiment, in step S1, the angle of the curvelet transform is selected as 5, and the scale of the second coarsest angle is selected as 16;

[0038] S2: Perform inverse curvelet transform to obtain the inverse transform diagram of each angle of each scale;

[0039] In one embodiment, in step S2, traverse each angle of each scale, set the remaining angle coefficients of this angle to zero, and perform inverse curvelet transform;

[0040] S3: Perform curvelet classification, calculate the average gradient of each inverse transform diagram using the Sobel operator, and accumulate those with a positive average gradient and accumulate those with a negative average gradient;

[0041] S4: Preprocess the vertical seismic profile data, normalizing it between -1 and 1.

[0042] S5: Construct a dual convolutional autoencoder network, including a down-wave encoder and an up-wave encoder;

[0043] In one embodiment, in step S5, as Figure 2 shown, the dual convolutional autoencoder network includes two convolutional autoencoders, one is a down-wave autoencoder and the other is an up-wave autoencoder. A convolutional autoencoder includes seven convolutional layers. The first three convolutional layers are used to construct the encoding layer of the neural network, and a max-pooling layer is used during the encoding process to expand the receptive field of the neural network: The latter four convolutional layers are used to construct the decoding process of the neural network, and an upsampling layer symmetric to the pooling layer is used during the decoding process to simulate the inverse process of pooling: All convolutional layers use an activation function one for activation.

[0044] In one embodiment, each convolutional layer performs a two-dimensional convolutional operation with a 3*3 kernel: the max-pooling uses a 2*2 kernel: the activation function uses the Tanh function. In this embodiment, the receptive field of the neural network can be expanded through the max-pooling layer, the information redundancy introduced by the convolutional operation can be reduced, and the performance of the network can be improved. The Tanh activation function can avoid the numerical overflow of the vertical seismic profile data because the vertical seismic profile data of the present invention is normalized between -1 and 1.

[0045]

[0046] We name the two convolutional autoencoders C1 and C2, where C1 is the down-wave encoder and C2 is the up-wave encoder. We input the accumulated map with a positive average gradient after normalization into C1, and the accumulated map with a negative average gradient into C2. The two outputs are named outputs1 and outputs2.

[0047] S6: Construction of the loss function. The loss function mainly includes the reconstruction loss and the gradient loss;

[0048] In one embodiment, for the reconstruction loss, we use the mean squared error loss function MSE. The reconstruction loss is the reconstruction loss of outputs1, outputs2 and the vertical seismic profile data to be separated. For the gradient loss function, we use the improved Sobel operator to constrain outputs1 and outputs2. The total loss function is the reconstruction loss plus 1e -3 of the gradient loss.

[0049] Mean squared error loss function: where y i is the true value, is the predicted value.

[0050] Gradient loss function: where The Sobel operator for predicting the angle, ⊙ represents the Hadamard product, and * represents convolution.

[0051] S7: Perform training and save the trained model.

[0052] In this embodiment, iterative training for 5000 epochs is set, and the total loss function decreases from 1e -2 to 1e -5 , which indicates that the training is effective.

[0053] S8: Use the model saved in S7 to separate the up - and - down - going wavefield data of the vertical seismic profile to be separated;

[0054] The input data is an image after normalization, so it is necessary to denormalize the output data for visual analysis.

[0055] In one embodiment, the filtering process includes the following steps: data pre - processing, normalizing the vertical seismic profile data to be separated; transforming to the frequency - wavenumber domain, first performing a fast Fourier transform on the vertical seismic profile data to be separated to transform from the time domain to the frequency domain, and then performing a Fourier transform to obtain the wavenumber; FK - domain filtering, retaining the regions with positive and negative wavenumbers respectively, and performing an inverse transform.

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0057] Simulate a validation set of data to test the separation performance of the double - convolutional auto - encoder network described in the present invention. The results are as Figure 3 shown. By comparing Figure 3 (a) and (b), it can be seen that the separation result of the present invention basically does not lose valid information. From Figure 3 (c), it can be seen that the down - going wave separated by the present invention basically does not contain the up - going wave. From Figure 3 (d), it can be seen that the up - going wave separated by the present invention basically does not contain the down - going wave.

[0058] To further verify the separation performance of the model proposed in the present invention, the up - and - down - going waves separated by the FK filtering and curvelet transform methods are compared. The results are as Figure 4 shown. By comparison, this method can well separate the up - and - down - going waves of the model data, will not appear the false axis caused by the Fourier transform mechanism, and the separated up - going wave will not contain the residue of the down - going wave.

[0059] Further perform frequency - wavenumber (F - K) spectral analysis on the separation results. The results are as Figure 5As shown, through comparison, it can be found that the separation of the up and down traveling waves achieved by the present invention separates the number of positive waves and negative waves in the validation set data, and there is no obvious energy loss, indicating that the separation method of the present invention has good effects.

[0060] To further verify the separation performance of the model proposed by the present invention, the vertical seismic profile data of a complex model was forward modeled. The results are as Figure 6 shown. Through comparison, this method can still well separate the up and down traveling waves of the complex model data.

[0061] To further verify the separation performance of the model proposed by the present invention, a part of the actual data in a certain area of Tarim was selected. The results are as Figure 7 shown.

[0062] Compared with the model data, the real vertical seismic profile wave field data is more complex and disordered. Through comparison, the up traveling wave does not contain the down traveling wave component, and the down traveling wave does not contain the up traveling wave component, and the up and down traveling waves therein can be well separated.

[0063] In summary, the wave field separation method based on the combination of curvelet transform and deep learning proposed by the present invention can intelligently and accurately separate the up and down traveling waves of the vertical seismic profile wave field data.

[0064] Although the present invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not depart from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A wave field separation method based on the combination of curvelet transform and deep learning, characterized in that: The specific steps are: S1: Performing a curvelet transform on the vertical seismic profile data to be separated to obtain the curvelet coefficients of the vertical seismic profile data to be separated; S2: Perform inverse curvelet transform to obtain the inverse transform map of each angle at each scale; S3: Perform curvelet classification, use the Sobel operator to calculate the average gradient of each inverse transformation image, and accumulate the average gradients with positive values ​​and the average gradients with negative values; S4: Construct a dual convolutional autoencoder network, including a downlink autoencoder and an uplink autoencoder; S5: Determine the loss function, which includes reconstruction loss and gradient loss; S6: Input the accumulated graph with positive average gradient to the downlink encoder, input the accumulated graph with negative average gradient to the uplink encoder, and start training; S7: Save the model obtained in S6, and use the model to separate the uplink and downlink waves of the vertical seismic profile wave field data to be separated.

2. The wave field separation method based on the combination of curvelet transform and deep learning according to claim 1, characterized in that: In step S1, the angle of the curvelet transform is selected as 5, and the scale of the second coarse angle is selected as 16.

3. The wave field separation method based on the combination of curvelet transform and deep learning according to claim 1, characterized in that: In step S4, each autoencoder includes seven convolutional layers, of which the first three convolutional layers are used to construct the encoding layer of the neural network, and the maximum pooling layer is used to expand the receptive field of the neural network during the encoding process: the last four convolutional layers are used to construct the decoding process of the neural network, and an upsampling layer symmetrical to the pooling layer is used in the decoding process to simulate the inverse process of pooling.

4. The wave field separation method based on the combination of curvelet transform and deep learning according to claim 3 is characterized in that: Each convolution layer uses a 2D convolution operation with a 3*3 kernel. The maximum pooling uses a 2*2 kernel. The activation function uses the Tanh function.

5. The wave field separation method based on the combination of curvelet transform and deep learning according to claim 1, characterized in that: In step S2, each angle of each scale is traversed, and the remaining angle coefficients of the angle are set to zero, and an inverse curvelet transform is performed.

6. The wave field separation method based on the combination of curvelet transform and deep learning according to claim 1, characterized in that: Step S3 also includes preprocessing the vertical seismic profile data and normalizing it between -1 and 1.

Images