CN110568485A - A Neural Network-Based Separation Method of Multi-channel Seismic Continuous Recording - Google Patents

Abstract

The invention relates to a neural network-based multi-channel seismic continuous recording and separating method, which comprises the following steps: step S1: obtaining continuous recording data and discontinuous recording data; step S2: constructing a neural network model, intercepting continuous recorded data to obtain a pseudo single shot record as input data, taking discontinuous recorded data as target data, and inputting the input data and the target data into the neural network model for training to obtain a trained neural network; step S3: and obtaining actual measurement earthquake continuous records to be separated, inputting the actual measurement earthquake continuous records to a trained neural network for separation, wherein the output of the trained neural network is the separation result. The method has the advantages of conveniently, quickly and accurately separating multiple seismic continuous records, effectively removing aliasing energy, obtaining single shot records meeting the requirements and having good separation effect.

Description

A Neural Network-Based Separation Method of Multi-channel Seismic Continuous Recording

technical field

The invention relates to the technical field of marine seismic exploration data processing, in particular to a method for separating multi-channel seismic continuous records based on a neural network.

Background technique

Marine seismic exploration is an important means of marine oil and gas and basic geological surveys. For marine seismic exploration data acquisition, the traditional acquisition method adopts non-continuous recording to acquire seismic data. Generally, the minimum value is determined according to geological targets and resolution requirements. , the maximum offset distance, shot spacing and recording length, the shot spacing is generally fixed (such as 25m), and the ship speed is adjusted to ensure that the excitation interval of the source is greater than the length of the single shot record, and to ensure that there is no aliasing between shots . With the improvement of seismic acquisition technology, more and more attention has been paid to the continuous recording method. Compared with the discontinuous recording method, the continuous recording acquisition can not only greatly improve the acquisition efficiency, but also increase the data density.

When seismic data is acquired by continuous recording, the source is fired at intervals, and the geophones keep recording until the source stops firing. Due to the existence of aliasing energy in continuous recording, separation is required, and the quality of data separation directly determines the effect of subsequent processing of continuous recording data. At present, the separation of continuous records is mainly based on the idea of denoising. Among them, scholars Seher and Clarke proposed a two-step method. First, estimate the low-frequency components in the aliasing energy, and remove the low-frequency components in the aliasing energy from continuous records by adaptive subtraction; then, attenuate the aliasing energy based on statistical methods The high-frequency components in the energy mainly use K-L transform, median filter and other technologies. The estimation of aliasing energy is a key part of this kind of method, which directly determines the effect of separation. Generally, the aliased direct wave is estimated in the common detection point domain, and the aliased reflected wave is estimated in the CDP domain through the Randon transformation. At present, it is difficult to estimate the aliasing energy in the separation method of continuous seismic records, especially under the complex geological conditions of the seabed, it is difficult to convert continuous records with energy aliasing into single-shot records, resulting in poor separation effect of continuous seismic records .

Contents of the invention

Aiming at the deficiencies of the prior art, the object of the present invention is to provide a method for separating multi-channel continuous seismic records based on neural network, which can solve the problem of separating multi-channel continuous seismic records.

The technical scheme that realizes the object of the present invention is: a kind of multi-channel seismic continuous record separation method based on neural network, comprises the steps:

Step S1: Obtain continuous recording data and non-continuous recording single-shot data through numerical simulation;

Step S2: Construct a neural network model, wherein a pseudo-single-shot record is intercepted from the continuous recording data according to the excitation time of the excitation and the preset recording length, and the pseudo-single-shot record represents the recording data corresponding to each excitation in the continuous recording data , the pseudo-single-shot record is used as input data, and the single-shot data of non-continuous recording is used as target data, and the input data and target data are input to the neural network model for training to obtain a trained neural network;

Step S3: Obtain the continuous record of the measured earthquake to be separated, intercept the continuous record of the measured earthquake to obtain a pseudo-single-shot record, input the pseudo-single-shot record to the trained neural network for separation, and the output is the separated Single shot results.

Further, obtaining the continuous recording data through numerical simulation includes the following steps: Preset several seismic sources, the seismic sources excite at preset time intervals, the geophones continue to receive until the seismic sources stop exciting, and the initial wave field of the next excitation is the above Stimulate the wave field at the last two moments at a time;

Obtaining the discontinuous recording data by numerical simulation includes the following steps,

Set the same number of seismic sources as for obtaining continuous recording data. The seismic sources excite at the same preset time interval as described above, and the geophones record. When one excitation ends, the geophones stop receiving until the next excitation, each time The initial wavefield of the excitation is reset to a 0-valued wavefield.

Further, the target data and input data are records generated by the same shot point excitation.

Further, the neural network model is a fully connected deep neural network.

The beneficial effects of the present invention are: the present invention can conveniently, quickly and accurately separate multi-channel seismic continuous records, can effectively remove aliasing energy, and obtain single-shot records meeting requirements, with good separation effect.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is the shot point excitation schematic diagram of collecting continuous recording data and non-continuous recording data;

Fig. 3 is the schematic diagram of continuous recording data and non-continuous recording data (including pseudo-single-shot recording) obtained;

Figure 4 is a schematic diagram of tag data;

Fig. 5 is a schematic diagram of a deep neural network structure;

Figure 6 is a comparison diagram of the separated single-shot record and the ideal single-shot record.

specific implementation plan

Below, in conjunction with accompanying drawing and specific embodiment, the present invention is described further:

As shown in Figures 1 to 4, a neural network-based multi-channel seismic continuous record separation method includes the following steps:

Step S1: generating label data for training the neural network.

The implementation process of this step includes the following sub-steps:

Step S11: Generate continuous recording data and non-continuous recording data through forward modeling, and there is no data of aliasing energy in the non-continuous recording data. Forward modeling is also a numerical simulation method.

With reference to the acquisition method shown in Figure 2, several sources (for example, 3) are set. The source is the excitation shot, and one source is a shot point, and the source is excited at a preset time interval. For the continuous recording method, the geophone continues to receive until all the seismic sources stop exciting; while for the discontinuous recording, the geophone does not continue to receive but only records each excitation of the seismic source. After each excitation, the geophone stops receiving. The detector does not start receiving until the next excitation. Figure 3(a) is the continuous recording data obtained by the continuous recording method, and Figure 3(b) is the single-shot recording obtained by the discontinuous recording method.

Continuously recorded data and non-continuously recorded single-shot data can be obtained respectively through two forward modeling simulations. It is only necessary to set the parameters and numerical models of the two forward modeling simulations to be the same.

In the simulation of discontinuously recorded data, before a new excitation shot starts to simulate, the wave field is reset to a zero-value wave field, and each excitation shot is independent of each other and does not interfere with each other. In the simulation of continuous recording data, after the end of each firing shot, the wave field at the last two moments is saved as the initial wave field for the next firing shot simulation. The difference between the continuous recording simulation and the discontinuous recording simulation lies in whether the initial wave field is reset when starting a new excitation shot simulation.

For the discontinuous recording mode, each excitation corresponds to a single shot recording. As for the continuous recording mode, one recording file contains multiple pseudo-single-shot recordings.

Step S12: Make tag data for the continuous recording data and discontinuous recording data obtained in step S11, and the tag data includes input data and target data. The input data is the pseudo-single-shot record intercepted from the continuous recording data. The pseudo-single-shot record is intercepted from the continuous recording data according to the excitation time of the shot point and the preset recording length, as shown in Figure 3(c). 3(c) is the pseudo-single-shot record intercepted from Fig. 3(a). Correspondingly, the non-continuously recorded single-shot data of the same shot point is used as the target data (Fig. 3b).

Figure 4 is an example of label data, Figure 4(a) is a single shot record without aliasing energy recorded by two different excitations at the shot point, Figure 4(b) is a continuous record with aliasing energy, where, Figure 4 Arrows in (b) indicate aliasing energy. The single-shot record without aliasing energy is the target data, and the single-shot data with continuous recording pseudo-separation is the input data.

Step S2: Construct a neural network model, and input the label data into the neural network model for training to obtain a trained neural network.

The middle layer of the neural network model (Fig. 5) of the present embodiment is set to 3 layers, and the number of nodes in the middle layer is 10, and it is a full connection mode between each node, and the excitation function of the neural network model is an ELU excitation function (exponential linear unit function).

Preferably, the neural network model is set as a fully connected deep neural network.

The objective function J (m) of the neural network model of the present embodiment is the two norms of the difference between the output data of the neural network model and the target data, as shown in formula 1.:

Among them, e _i represents the element of the difference between the output data and the target data, N represents the number of sampling points of the target data, and m represents the parameters of the neural network model.

The initial weights and biases of the neural network model are random numbers conforming to the normal distribution. Due to the large amount of continuous recording data and non-continuous recording data, the neural network model is updated using the stochastic gradient algorithm, and the updating formula is shown in ②:

Among them, α ^k represents the learning rate, which is used to control the step size of learning, k represents the number of learning times, and q represents the step size, Represents the gradient of the objective function, whose elements are the partial derivatives of the objective function with respect to the weights and biases of the neural network model, The calculation of can be realized by the backpropagation of the error.

Determine the maximum number of iterations according to the convergence characteristics of the objective function, so as to determine the final trained neural network.

In this step, it also includes evaluating the trained neural network, and dividing the label data into training data and test data, the training data is used to train the neural network, and the test data is used to evaluate the generalization of the trained neural network model Ability to ensure that there is no overfitting and underfitting to obtain a neural network with excellent performance.

Step S3: Pseudo-separate the seismic continuous record data obtained from the actual measurement, that is, according to the excitation time of the shot point and the preset record length, intercept part of the data from the continuous record data to obtain a pseudo-single-shot record, a pseudo-single-shot record That is to say, the data containing aliasing energy, and the pseudo-single-shot data is input into the trained neural network for processing, and the output data is the conventional single-shot record, which is to separate the multi-channel seismic continuous records.

As shown in Figure 6, it is a comparison diagram between the single shot record and the ideal single shot record obtained by adopting the present embodiment to separate the seismic continuous recording data, and Fig. 6 (a) is the pseudo single shot record obtained from the pseudo separation of the seismic continuous recording data , Figure 6(b) is the single-shot record processed and output by using the neural network of this scheme to process the pseudo-single-shot record in Figure 6(a), and Figure 6(c) is an ideal single-shot record without aliasing, that is It is the single-shot record obtained by the actual discontinuous recording method. It can be seen that Figure 6(b) is close to Figure 6(c), and the slight difference is due to the fact that Figure 6(b) contains random noise, which can be attenuated by conventional denoising methods, and can be separated from the ideal single shot Record closer single-shot records.

The embodiment disclosed in this specification is only an illustration of the unilateral feature of the present invention, and the protection scope of the present invention is not limited to this embodiment, and any other functionally equivalent embodiments fall within the protection scope of the present invention. For those skilled in the art, various other corresponding changes and modifications can be made according to the technical solutions and ideas described above, and all these changes and modifications should fall within the protection scope of the claims of the present invention.

Claims (4)

1. A multi-channel seismic continuous recording separation method based on a neural network is characterized by comprising the following steps:

Step S1: respectively obtaining continuous recording data and single shot data recorded discontinuously through numerical simulation;

Step S2: constructing a neural network model, wherein a pseudo single shot record is obtained by intercepting continuous recording data according to the excitation time of excitation and a preset recording length, the pseudo single shot record represents the recording data corresponding to each excitation in the continuous recording data, the pseudo single shot record serves as input data, the single shot data recorded discontinuously serves as target data, and the input data and the target data are input into the neural network model for training to obtain a trained neural network;

step S3: and acquiring actual measurement earthquake continuous records to be separated, intercepting the actual measurement earthquake continuous records to obtain pseudo single shot records, inputting the pseudo single shot records to a trained neural network for separation, and outputting the results, namely the separated single shot results.

2. The neural network-based multi-channel seismic continuity record separation method of claim 1, wherein numerically simulating obtaining the continuity record data includes the steps of,

Presetting a plurality of seismic sources, exciting the seismic sources according to a preset time interval, continuously receiving by using a detector until the seismic sources stop exciting, wherein the initial wave field of the next excitation is the wave field of the last two moments of the previous excitation;

Obtaining the non-contiguously recorded data by numerical simulation includes the steps of,

and setting the same number of seismic sources as the seismic sources for obtaining the continuous recording data, exciting the seismic sources according to the same preset time interval, recording by using the detector, stopping receiving by using the detector when one excitation is finished until the next excitation is finished, and resetting the initial wave field of each excitation to be a 0-value wave field.

3. The method of claim 1, wherein the target data and the input data are recordings generated by the same shot firing.

4. the method for separating the multi-channel seismic continuous recording based on the neural network as claimed in claim 1, wherein the neural network model is a fully-connected deep neural network.