Uplink and downlink wave deep learning amplitude preservation separation method based on 3D-VSP imaging

Abstract

The invention discloses an uplink and downlink wave deep learning amplitude-preserved separation method based on 3D-VSP imaging, and the method comprises the following steps in sequence: S1, making a data set: extracting part of shot gather data in a target work area, extracting uplink waves and downlink waves through Radon transformation, carrying out the recombination, and building a training data set; s2, training: slicing the training data set, putting the sliced training data set into a Unet network, training a related network model, and stopping training when the signal-to-noise ratio in the training process meets the requirement or does not rise any more; and S3, application: after shot gather data of a target work area is sliced, putting the shot gather data into the trained Unet network model for prediction, after prediction is finished, recovering prediction result slices, and averaging overlapped parts to obtain separated uplink waves or downlink waves. According to the method, full-automatic processing of 3D-VSP data can be realized, and a high-amplitude-preservation separation result can be realized without manual parameter adjustment.

Description

Technical Field

[0001] The present invention relates to seismic data processing in oil exploration, and specifically to a deep learning DAS-VSP high-fidelity traveling wave separation method during the preparation of 3D-VSP imaging data. Background Art

[0002] High-resolution imaging is of great significance for oil and gas exploration.

[0003] Vertical Seismic Profile (VSP) can obtain high-precision underground information beside the well, providing accurate interlayer velocity and attenuation. By using Fiber Optic Distributed Acoustic Sensing (DAS) technology combined with VSP for full-well measurement, more accurate underground rock formation information can be further obtained, and high-resolution VSP data can be obtained. Currently, the application of VSP data mainly relies on the separated up-going wave or down-going wave. However, in the original data, the up-going and down-going waves are mixed together. Therefore, a high-precision and high-fidelity wave field separation technology is the basis for subsequent data migration imaging.

[0004] There are currently many methods for separating the up-going and down-going waves of VSP data. For example: ① f-k filtering, which uses the slowness difference between the up-going and down-going waves to achieve wave field separation. However, spatial aliasing usually occurs after filtering, and it has high requirements for spatial sampling and is extremely dependent on the uniformity of geophone distribution; ② Radon transform waveform. In the τ-p domain, the negative half-plane and the positive half-plane correspond to the up-going wave and the down-going wave respectively. By performing the inverse transform, the separated wave field can be obtained. However, it is difficult to maintain the AVO amplitude information by this method. Therefore, when processing actual data, the down-going wave is extracted by τ-p, and then the up-going wave is obtained by subtracting the down-going wave from the original data to retain the information of the up-going wave as much as possible; ③ Median filtering, whose advantage is amplitude preservation, but its effect is affected by factors such as the accuracy of the first arrival time and the waveform consistency; ④ Adaptive subtraction, which extracts the wave field that conforms to a given time-distance curve according to the waveform and energy similarity between channels within a given time window of VSP data, and then obtains the remaining wave field through adaptive subtraction. The limitation of this method is that it requires accurate picking of the time-distance curve of the wave field to be extracted.

[0005] Among these methods, although the Radon transform waveform method separates relatively cleanly, it does not have amplitude preservation, and like f-k filtering, it will produce spatial aliasing. The effects of median filtering and adaptive subtraction to a large extent depend on the waveform consistency of the data and the accuracy of first arrival picking. Summary of the Invention

[0006] To address the above deficiencies in the existing technology, the present invention aims to provide a deep learning amplitude-preserving separation method for up- and down-going waves based on 3D-VSP imaging, enabling fully automated processing and achieving high-amplitude-preserving separation results without the need for manual parameter adjustment.

[0007] To achieve the above objective, the technical solution adopted by the present invention is as follows: A deep learning amplitude-preserving separation method for up- and down-going waves based on 3D-VSP imaging, including the following steps carried out sequentially: S1. Produce a dataset: Extract a portion of the shot gather data in the target work area, extract the up-going wave and the down-going wave through Radon transform, and recombine them to establish a training dataset; S2. Train: After slicing the training dataset, put it into the Unet network to train the relevant network model. When the signal-to-noise ratio during the training process reaches the requirement or no longer increases, stop the training; S3. Apply: After slicing the shot gather data of the target work area, put it into the trained Unet network model for prediction. After the prediction is completed, restore the sliced prediction results and take the average value of the overlapping parts to obtain the separated up-going wave or down-going wave.

[0008] As a limitation of the present invention, in step S1, when extracting a portion of the shot gather data, it is carried out at a ratio of 1 / 100, and at least 50 shots are extracted.

[0009] As a further limitation of the present invention, in step S1, the way of recombining the extracted up-going wave and down-going wave is: Add the up-going wave and the down-going wave, and name the obtained data as the synthetic input data.

[0010] As a further limitation of the present invention, in step S1, the established training dataset includes the synthetic input data, as well as the up-going wave data and the down-going wave data; If the target result is to obtain the separated up-going wave, then in step S2, the synthetic input data needs to be used as the input data of the Unet network, and the up-going wave data as the label; If the target result is to obtain the separated down-going wave, then in step S2, the synthetic input data needs to be used as the input data of the Unet network, and the down-going wave data as the label.

[0011] As another limitation of the present invention, in step S2, when slicing the training dataset, the training dataset is cut according to the specifications of 64*64 or 128*128, and the sliding step of the slice is half of the slice.

[0012] As a further limitation of the present invention, in step S3, if the separated up-going wave is obtained, the separated down-going wave can be obtained by subtracting the separated up-going wave from the original shot gather data; If the obtained is the downgoing wave after separation, subtracting the separated downgoing wave from the original shot gather data can obtain the separated upgoing wave.

[0013] Due to the adoption of the above technical solution, compared with the prior art, the beneficial effects achieved by the present invention are as follows: Based on the Radon transform, the present invention makes a training data set and trains a related network model based on the Unet network through deep learning. After being applied to actual data, the trained network model can accurately perform wavefield separation, greatly improving the amplitude preservation of the separation result. Among them, during the training process of the network model, the present invention uses the signal-to-noise ratio to quantify the amplitude preservation. The higher the signal-to-noise ratio, the higher the amplitude preservation of the data.

[0014] Figure 2 It is a comparison chart of the effects of wavefield separation using different methods, showing the signal-to-noise ratios of the data after wavefield separation using different methods. It can be observed from this that the signal-to-noise ratio of the result of the present invention far exceeds that of the Radon transform and the f-k filtering method, thus proving that the present invention can greatly improve the signal-to-noise ratio of wavefield separation and increase the amplitude preservation.

[0015] Figure 3 It is a comparison chart of the actual application effects. Figure 3 (2) is the upgoing wave extracted by the Radon transform method, Figure 3 (3) is the downgoing wave obtained by subtracting Figure 3 (2) from the original data. In Figure 3 (3), the residue of the upgoing wave can still be observed, which is caused by the non-amplitude preservation of the Radon transform method; Figure 3 (4) is the upgoing wave extracted by the present invention, Figure 3 (5) is the downgoing wave obtained by subtracting Figure 3 (4) from the original data. In Figure 3 (5), the residue of the upgoing wave can hardly be seen, which proves the advantage of the present invention - it can greatly improve the amplitude preservation of wavefield separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0017] Figure 1 It is the network structure diagram for extracting the upgoing wave in Embodiment 1 of the present invention; Figure 2 It is a comparison chart of the effects of wavefield separation using different methods; Among them, Figure 2 (1) is the signal-to-noise ratio of the synthetic input data before separation; Figure 2 (2) is the signal-to-noise ratio of the upgoing wave extracted by the Radon transform method, SNR = 2.95;Figure 2 (3) is the signal-to-noise ratio of the downgoing wave obtained by subtracting the upgoing wave extracted by the Radon transform method from the synthetic input data, SNR = 3.77; Figure 2 (4) is the signal-to-noise ratio of the upgoing wave extracted by the f-k filtering method, SNR = 8.60; Figure 2 (5) is the signal-to-noise ratio of the upgoing wave extracted by the embodiment of the present invention, SNR = 18.28; Figure 3 It is a comparison chart of the actual application effects; Among them, Figure 3 (1) is the actual shot gather data of a certain work area; Figure 3 (2) is the upgoing wave extracted by the Radon transform method; Figure 3 (3) is obtained by subtracting Figure 3 (2) from the original data to get the downgoing wave; Figure 3 (4) is the upgoing wave extracted by the embodiment of the present invention; Figure 3 (5) is obtained by subtracting Figure 3 (4) from the original data to get the downgoing wave. Specific implementation mode

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and understanding the present invention, and are not used to limit the present invention.

[0019] Embodiment 1 A method for amplitude-preserving separation of up and down waves based on 3D-VSP imaging by deep learning In this embodiment, based on the traditional Radon transform waveform, deep learning is used to achieve fast high-amplitude-preserving wave separation. The specific steps are as follows: S1. Making a data set In the entire data set of the target work area, some shot gather data are extracted at equal intervals as a subset to establish a training data set. Generally, at least 50 shots of data are extracted. If the base number of the data set is relatively large, it can be extracted according to a ratio of 1 / 100.

[0020] The subset is subjected to Radon transform to directly extract the upgoing wave and the downgoing wave, and then recombined: the extracted upgoing wave and downgoing wave are added together, and the obtained data is named synthetic input data. The purpose of recombining after extraction is that the amplitude preservation of the upgoing wave or downgoing wave directly extracted by the Radon transform is poor. Therefore, after adding the extracted upgoing wave and downgoing wave, they are combined into a new unseparated wave field.

[0021] The synthetic input data, the directly extracted upgoing wave data, and the downgoing wave data are integrated to obtain the required training data set.

[0022] S2. Training After slicing the training dataset, it is put into the Unet network for training. When the signal-to-noise ratio during training reaches the requirement or no longer increases, and the training accuracy reaches the ideal state, the training stops, and a Unet network model suitable for studying the entire dataset of the target work area is obtained.

[0023] In this step, when slicing the training dataset, the training dataset needs to be cut according to the specifications of 64*64 or 128*128, and the sliding step of the slice is half of the slice size.

[0024] As Figure 1 shown, it is a network structure diagram for extracting the up-going wave during the training of a network model. The figure shows that when the signal-to-noise ratio reaches 20, the training accuracy reaches a relatively ideal state, and at this time, the training can be stopped.

[0025] It should be noted that if the ultimate goal is to obtain the separated up-going wave, during model training, the synthetic input data in the training dataset needs to be used as the input data of the Unet network, and the up-going wave data as the label; Similarly, if the ultimate goal is to obtain the separated down-going wave, during model training, the synthetic input data in the training dataset needs to be used as the input data of the Unet network, and the down-going wave data as the label.

[0026] S3. Application Apply the trained Unet network model to the data to be processed. Specifically, after slicing the entire dataset of the target work area, it is put into the trained Unet network model for prediction. After the prediction is completed, the predicted result slices are restored, and the average value is taken for the overlapping part, and then the separated up-going wave or down-going wave can be obtained.

[0027] In this step, the following principles are followed when slicing the entire dataset: within the range of the device memory limit, the slices should be as large as possible, and there should be a small overlapping area between the slices.

[0028] Example 2 A method for amplitude-preserving separation of up-going waves in deep learning based on 3D-VSP imaging S1. Making the dataset This step is exactly the same as step S1 in Example 1. The specific steps are as follows: 1.1 Uniformly extract a part of the shot gather data from the entire dataset of the target work area at a ratio of 1 / 100, generally extracting at least 50 shots; 1.2 Perform Radon transform on the extracted subset to directly extract the up-going wave and the down-going wave; 1.3 Add the extracted up-going wave and down-going wave, and name it the synthetic input data.

[0029] S2. Training This embodiment aims to obtain the separated up-going wave. Therefore, the up-going wave data is used as the label during the training of the network model. The specific steps are as follows: 2.1 Slice the synthetic input data and the up-going wave data according to the specification of 128*128, and the sliding step of the slice is half of the slice size; 2.2 Put the prepared slices into the Unet network for training, where the slices of the synthetic input data are used as the input data of the Unet network, and the up-going wave data is used as the label; 2.3 When the signal-to-noise ratio during the training reaches the requirement or no longer increases, stop the training to obtain the Unet network model.

[0030] S3. Application 3.1 Slice the entire dataset of the target work area according to the specification of 128*128 as well; 3.2 Put the prepared slices into the trained Unet network model for prediction to obtain the slices of the up-going wave prediction results; 3.3 Restore the prediction result slices, take the average value for the overlapping parts, and then the separated up-going wave can be obtained.

[0031] If you want to obtain the separated down-going wave, it can be obtained by subtracting the separated up-going wave from the original shot gather data.

[0032] Embodiment 3 A method for amplitude-preserving separation of down-going waves in 3D-VSP imaging based on deep learning This embodiment is basically the same as Embodiment 2, except that in step S2 of this embodiment, the down-going wave data is used as the label for the training of the Unet network model. By applying this Unet network model, the separated down-going wave can be directly obtained in step S3.

[0033] Similarly, on this basis, if you want to obtain the separated up-going wave, it can be obtained by subtracting the separated down-going wave from the original shot gather data.

[0034] It should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the above embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning, characterized in that: it includes the following steps carried out in sequence: S1. Making a data set: Extracting part of the shot gather data in the target work area, extracting the up-going wave and the down-going wave through Radon transform, and recombining them to establish a training data set; S2. Training: After slicing the training data set, putting it into the Unet network to train the relevant network model, and stopping the training when the signal-to-noise ratio during the training process reaches the requirement or no longer increases; S3. Application: After slicing the shot gather data of the target work area, putting it into the trained Unet network model for prediction. After the prediction is completed, restoring the sliced prediction results and taking the average value of the overlapping parts to obtain the separated up-going wave or down-going wave.

2. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning according to claim 1, characterized in that: in step S1, when extracting part of the shot gather data, it is carried out at a ratio of 1 / 100, and at least 50 shots are extracted.

3. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning according to claim 2, characterized in that: in step S1, the way of recombining the extracted up-going wave and down-going wave is: adding the up-going wave and the down-going wave, and naming the obtained data as the synthetic input data.

4. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning according to claim 3, characterized in that: in step S1, the established training data set includes the synthetic input data, the up-going wave data, and the down-going wave data; if the target result is to obtain the separated up-going wave, then in step S2, the synthetic input data needs to be used as the input data of the Unet network, and the up-going wave data as the label; if the target result is to obtain the separated down-going wave, then in step S2, the synthetic input data needs to be used as the input data of the Unet network, and the down-going wave data as the label.

5. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning according to any one of claims 1-4, characterized in that: in step S2, when slicing the training data set, the training data set is cut according to the specifications of 64*64 or 128*128, and the sliding step of the slice is half of the slice.

6. A method for amplitude-preserving separation of up-going and down-going waves based on 3D-VSP imaging using deep learning according to claim 5, characterized in that: in step S3, if the separated up-going wave is obtained, the separated down-going wave can be obtained by subtracting the separated up-going wave from the original shot gather data; if the separated down-going wave is obtained, the separated up-going wave can be obtained by subtracting the separated down-going wave from the original shot gather data.

Images