DAS seismic data denoising processing method based on MFA-PromptIR network

Abstract

The invention discloses a DAS seismic data denoising processing method based on an MFA-PromptIR network, and belongs to the technical field of seismic exploration data denoising. Aiming at the problem of poor denoising effect when a traditional denoising method is used for processing DAS seismic data with low signal-to-noise ratio and various complex noises, a PromptIR network adopts a codec structure, and a multi-scale fusion and attention mechanism is introduced into the network. The multi-scale fusion can prevent detail information from being ignored, and the attention mechanism can make the network more focus on effective information, so that the network can extract useful features more efficiently and accurately, thereby constructing a network model for seismic record denoising. And synthesizing the simulated DAS data as a training set, and adjusting the weight by the network through an error between prediction output and expected output to obtain an optimal network model. And inputting the synthesized noisy signal and a field actual record into a network for de-noising processing, wherein the processed seismic record has a clear waveform and a relatively high signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic exploration data denoising, and particularly relates to a DAS seismic data denoising processing method based on the MFA-PromptIR network. Background Technique

[0002] Seismic exploration is the main means of oil and gas exploration. Generally, in seismic exploration, an artificial excitation source is used to generate elastic waves. By utilizing the elasticity and density differences of different underground media, geophones are used to receive and record the reflected waves, and then professional equipment is used to process the data, and finally seismic data interpretation is carried out. However, due to various interference problems such as exploration environment and instruments, seismic data often contains various noises, which bury the effective signals and cause great obstacles to the interpretation of seismic data. Therefore, noise suppression in seismic data and improvement of signal-to-noise ratio are important contents in seismic data processing work.

[0003] With the continuous increase in the demand for oil and gas resources, traditional exploration methods have been difficult to meet the current high-precision requirements of seismic exploration due to their disadvantages such as low spatial sampling density, high cost, and difficult layout. As a new acquisition technology, Distributed Acoustic Sensing (DAS) has the advantages of low cost, high efficiency, high precision, and can be repeatedly observed multiple times after one layout. However, due to instrument interference, DAS records usually have a large amount of background noise of various types, and the signal-to-noise ratio of seismic records is low. Therefore, the research on suppressing the background noise of DAS data and improving the signal-to-noise ratio is very meaningful.

[0004] Traditional denoising methods such as band-pass filtering (BP), f-k domain filtering, variational mode decomposition, matrix rank reduction, etc., due to their denoising principles, denoising speeds, and parameter adjustment problems, result in unsatisfactory denoising effects when processing DAS seismic data with low signal-to-noise ratio and various complex noises. Summary of the Invention

[0005] Aiming at the problem that traditional denoising methods have poor denoising effects when processing DAS seismic data with low signal-to-noise ratio and various complex noises, the present invention provides a DAS seismic data denoising processing method based on the MFA-PromptIR network.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A DAS seismic data denoising processing method based on the MFA-PromptIR network, comprising the following steps:

[0008] S1: Obtain the clean signal data s through forward simulation. Extract the noise data n from the actual field DAS seismic records, and combine them one-to-one with the clean signal data s to obtain the noisy signal y. Use the clean signal data s and the noisy signal y as a set of training data; Obtain the clean synthetic DAS record through forward simulation, add the actual noise to obtain the noisy record, and use it as the synthetic DAS test data;

[0009] Further, the step S1 is specifically as follows:

[0010] The clean signal data is obtained by intercepting the forward simulation data, with a size of 64*64; The noise data is extracted from the part of the actual DAS seismic records without valid signals, with a size of 64*64.

[0011] S2: On the basis of the PromptIR network, introduce a multi-scale fusion mechanism to construct the neural network model MFA-PromptIR;

[0012] Further, the step S2 is specifically as follows:

[0013] The MFA-PromptIR neural network model uses a structure of 4-layer encoder-decoder and adopts skip connections. The network model includes Transformer blocks and prompt blocks:

[0014] Among them, the Transformer blocks are used to extract features. Each layer in the encoding and decoding stages contains multiple Transformer blocks, and the number of blocks gradually increases from the top layer to the bottom layer to maintain computational efficiency; The prompt blocks supplement the relevant knowledge of the degradation type for the model while restoring the clean image, so as to guide the image restoration;

[0015] Add a feature fusion part of low-resolution features and high-resolution features to each layer in the neural network model to extract extensive information features and avoid ignoring some detailed information during the process of restoring from low resolution to high resolution.

[0016] S3: Add a channel attention mechanism after the network encoding, and also add a channel attention mechanism before each upsampling in the decoding stage; The channel attention mechanism (CA) focuses on useful information, enables the network to learn useful features more efficiently and accurately, and is very helpful for improving the denoising ability of the network.

[0017] S4: Input the training set into the neural network model and train the neural network model;

[0018] Further, the step S4 is specifically as follows:

[0019] In step S4, the training set is input into the neural network model, and the loss function is optimized through the adaptive moment estimation algorithm to obtain the optimal prediction. In this process, there is no need to adjust the parameters too much. The prediction output is the residual output, which is added to the noisy signal to obtain the denoised DAS data;

[0020] The loss function is expressed as:

[0021]

[0022] The denoising process is expressed as:

[0023]

[0024] The neural network model establishes a non - linear mapping between the noisy signal and the reconstructed signal. In the formula, P is the neural network model MFA - PromptIR, and θ are the parameters in the neural network model, including weights and offsets. is the output after denoising of the neural network model, M is the number of training samples. and s i are the prediction output set and the clean data set of the neural network model for M. By minimizing the loss function, the parameters θ of the neural network model are adjusted, so that the gap between the prediction output and the clean signal data is reduced, and a prediction output close to the clean signal data is obtained.

[0025] S5: Process the synthesized DAS test data and the field actual DAS data with the trained neural network model, and analyze the denoising effect through the signal - to - noise ratio.

[0026] Further, the specific step S5 is as follows:

[0027] The network denoising performance is tested through the synthesized DAS test record and the field actual DAS record, and the SNR is selected as a quantitative index to measure the network denoising performance.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] Aiming at the problem that DAS data contains a large amount of complex noise and the signal-to-noise ratio is extremely low, and traditional denoising methods cannot effectively suppress noise and restore clean signals, the present invention proposes a DAS data denoising processing method based on deep learning. Specifically, the network is a multi-scale fusion and attention mechanism-based prompt recovery network (MFA-PromptIR). Multi-scale fusion can avoid the neglect of detailed information, and the attention mechanism can make the network more focused on effective information, enabling the network to extract useful features more efficiently and accurately, greatly improving the denoising performance. The made clean data and noisy data are used as training data and input into the network for training. By minimizing the loss function, the optimal prediction result is obtained. Throughout the training process, no manual parameter adjustment is required. The denoising results of synthetic DAS seismic data and actual DAS data show that the MFA-PromptIR network performs outstandingly in noise suppression and signal restoration, greatly improving the signal-to-noise ratio. Therefore, the present invention can effectively suppress the DAS background noise, retain clear seismic signals, and provide reliable data for subsequent seismic data interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is the network structure diagram of MFA-PromptIR of the present invention;

[0031] Figure 2 It is the clean data and noisy data of the synthetic seismic test data of the present invention. Among them, (a) represents the clean synthetic DAS data, (b) represents the noisy DAS data, (c) represents the denoising processing result, and (d) represents the noise residual after denoising;

[0032] Figure 3 It is from Figure 2 The comparison of single-channel signal waveforms extracted from the synthetic seismic data shown. (a) is the time-domain waveform comparison diagram of the noisy signal, the denoising output of MFA-PromptIR, and the pure signal, and (b) is the frequency spectrum diagram of each time-domain waveform in (a);

[0033] Figure 4 It is the actual noisy seismic test data in the field of the present invention. (a) is the actual DAS noisy record in the field, and (b) is the denoising processing result. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] In order to deeply understand the present invention, we will describe it comprehensively and meticulously. However, the present invention has multiple implementation manners and is not limited to the specific examples listed herein. The presentation of these examples aims to deepen the comprehensive understanding of the disclosed content of the present invention.

[0035] A DAS seismic data denoising processing method based on the MFA-PromptIR network includes the following steps:

[0036] S1: Obtain the clean signal data s through forward simulation. Extract the noise data n from the actual field DAS seismic records, and combine them in one-to-one correspondence with the clean signal data s to obtain the noisy signal y. Use the clean signal data s and the noisy signal y as a set of training data; obtain the clean synthetic DAS record through forward simulation, add the actual noise to obtain the noisy record, and use it as the synthetic DAS test data;

[0037] Further, the specific steps of step S1 are as follows:

[0038] The clean signal data is obtained by intercepting the forward simulation data, with a size of 64*64 and a total of 32727 blocks; the noise data is extracted from the part of the actual DAS seismic records without valid signals, with a size of 64*64, and then the clean noise patches are combined with the noise-free pure patches to obtain the noisy data set.

[0039] S2: Based on the PromptIR network, introduce a multi-scale fusion mechanism to construct the neural network model MFA-PromptIR;

[0040] Further, the specific steps of step S2 are as follows:

[0041] The MFA-PromptIR neural network model uses a 4-layer encoder-decoder structure and adopts skip connections. The network model includes Transformer blocks and prompt blocks:

[0042] Among them, the Transformer blocks are used to extract features. Each layer in the encoding and decoding stages contains multiple Transformer blocks, and the number of blocks gradually increases from the top layer to the bottom layer to maintain computational efficiency; the prompt blocks supplement the relevant knowledge of the degradation type for the model while restoring the clean image, so as to guide the image restoration;

[0043] Add a feature fusion part of low-resolution features and high-resolution features to each layer in the neural network model to extract extensive information features and avoid ignoring some detailed information during the process of restoring from low resolution to high resolution.

[0044] S3: Add a channel attention mechanism after the network encoding and also add a channel attention mechanism before each upsampling in the decoding stage; the channel attention mechanism (CA) focuses on useful information, enables the network to learn useful features more efficiently and accurately, and is very helpful for improving the denoising ability of the network.

[0045] S4: Input the training set into the neural network model and train the neural network model;

[0046] Further, the specific steps of step S4 are as follows:

[0047] The 64*64 small blocks of clean signal data and the small blocks of noisy signal data are used as training data and input into the network. One-tenth of the training data is taken as the validation set, and the remaining part is used as the training set. Moreover, the clean noise data blocks and the noisy blocks correspond one by one. The network extracts features from the input training set and learns. By minimizing the loss function, the gap between the predicted output and the clean data gradually becomes smaller, and finally the optimal result is obtained. In this process, there is no need to adjust the parameters too much. The predicted output is the residual output, which is added to the noisy input to obtain the denoised DAS data. In addition, the network is trained on a computer with an Intel Core i9-7500K CPU and an NVIDIA GeForce GTX 4060 GPU.

[0048] S5: Process the synthesized DAS data and the actual field DAS data with the trained neural network model, and analyze the denoising effect through the signal-to-noise ratio;

[0049] Furthermore, the specific content of step S5 is as follows:

[0050] The denoising performance of the network is tested through the synthesized DAS test records and the actual field DAS records, and the SNR is selected as a quantitative index to measure the denoising performance of the network.

[0051] Next, the method proposed by the present invention is applied to the denoising experiment of the synthesized DAS seismic records and the actual field DAS seismic records:

[0052] Example 1

[0053] The present invention uses a synthesized DAS seismic record for the experiment, as Figure 2 shown. Figure 2 (a) is the clean record, Figure 2 (b) is the noisy record, Figure 2 (c) is the processing result after denoising by the MFA-PromptIR network, Figure 2 (d) is the background noise residual record after processing. It can be seen that after the synthesized record is denoised, almost all the noise is suppressed, and the effective signal is clear and coherent. From the analysis of the quantitative index, the signal-to-noise ratio has been greatly improved from -1 to 23.6.

[0054] Figure 3 shows the time-domain and frequency-domain comparison diagrams of a single-channel data extracted from the Figure 2 record. It can be seen that the noise in the denoising result of MFA-PromptIR is greatly suppressed, and the denoising result is very similar to the pure signal, thus indicating the high performance of the present invention.

[0055] Example 2

[0056] The present invention uses an actual field DAS record for denoising experiments. The actual field record is as shown in Figure 4 (a). It can be seen that there is a large amount of background noise in this actual record, and the effective signal is submerged and cannot be recognized. Figure 4 (b) shows the processing result after denoising by the MFA-PromptIR network. In the processing result, the effective signal is clear, and at the same time, the complex background noise is suppressed.

[0057] The results of both the simulation experiment and the actual field denoising processing show that MFA-PromptIR performs excellently in noise suppression and signal recovery, can recover clear DAS seismic records, and at the same time improve the signal-to-noise ratio of seismic data.

[0058] The above are only the preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto. For any person skilled in the art, any modification, equivalent replacement or change made according to the technical solution and inventive concept of the present invention should be within the protection scope of the present invention.

[0059] The content not detailedly described in the specification of the present invention belongs to the prior art well-known to those skilled in the art. Although the above describes the illustrative specific embodiments of the present invention for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the inventive concept of the present invention are within the scope of protection.

Claims

1. A DAS seismic data denoising method based on MFA-PromptIR network, characterized in that: The following steps are involved: S1: Get clean signal data s by forward modeling, extract noise data n from actual field DAS seismic records, combine it with clean signal data s in one-to-one correspondence to get noisy signal y, and use clean signal data s and noisy signal y as a training set; get clean synthetic DAS records by forward modeling, add actual noise to get noisy records, and use them as synthetic DAS test data; S2: Based on the PromptIR network, a multi-scale fusion mechanism is introduced to construct the neural network model MFA-PromptIR; S3: Add channel attention mechanism after network encoding and before each upsampling in the decoding stage; S4: input the training set into the neural network model and train the neural network model; S5: The trained neural network model is used to process the synthetic DAS test data and actual field DAS data, and the denoising effect is analyzed by the signal-to-noise ratio.

2. A DAS seismic data denoising method based on MFA-PromptIR network according to claim 1, characterized in that: S1: obtaining clean signal data s by forward modeling, extracting noise data n from actual field DAS seismic records, combining the clean signal data s with the noise data n in a one-to-one correspondence to obtain noisy signal y, and using the clean signal data s and the noisy signal y as a training set; obtaining clean synthetic DAS records by forward modeling, adding actual noise to obtain noisy records, and using them as synthetic DAS test data, specifically: The clean signal data is obtained from the forward simulation data and has a size of 64*64; the noise data is extracted from the part of the actual DAS seismic record that does not contain valid signals and has a size of 64*64.

3. A DAS seismic data denoising method based on MFA-PromptIR network according to claim 2, characterized in that: The step S2: based on the PromptIR network, a multi-scale fusion mechanism is introduced to construct a neural network model MFA-PromptIR, specifically: The MFA-PromptIR neural network model uses a 4-layer encoder-decoder structure and adopts skip connections. The network model includes a Transformer block and a prompt block: The Transformer block is used to extract features. Each layer in the encoding and decoding stage contains multiple Transformer blocks, and the number of blocks gradually increases from the top layer to the bottom layer to maintain computational efficiency. The prompt block supplements the model with relevant knowledge of degradation types while restoring the clean image to guide image restoration. A feature fusion part of low-resolution features and high-resolution features is added to each layer of the neural network model to extract extensive information features.

4. A DAS seismic data denoising method based on MFA-PromptIR network according to claim 3, characterized in that: The step S4: inputting the training set into the neural network model and training the neural network model, specifically: In step S4, the training set is input into the neural network model, and the loss function is optimized by the adaptive moment estimation algorithm to obtain the optimal prediction. The prediction output is the residual output, which is added to the noisy signal to obtain the denoised DAS data; The loss function is expressed as: The denoising process is expressed as: The neural network model establishes a nonlinear mapping between the noisy signal and the reconstructed signal. In the formula, P is the neural network model MFA-PromptIR, θ is the parameter in the neural network model, including weights and offsets, is the denoised output of the neural network model, M is the number of training samples, and i For M pairs of predicted output sets of the neural network model and clean signal data sets, the neural network model parameters θ are adjusted by minimizing the loss function, so that the gap between the predicted output and the clean signal data is reduced, and a predicted output close to the clean signal data is obtained.

5. A DAS seismic data denoising method based on MFA-PromptIR network according to claim 4, characterized in that: Step S5: Processing the synthesized DAS test data and the actual field DAS data with the trained neural network model, and analyzing the denoising effect by the signal-to-noise ratio, specifically: The network denoising performance is tested by using synthetic DAS test records and actual field DAS records, and SNR is selected as a quantitative indicator to measure the network denoising performance.

Images