An artificial intelligence seismic data processing method based on NSST and deep learning

Abstract

The application provides an artificial intelligence seismic data processing method based on NSST and deep learning, and comprises the following steps: step 1, synthesizing simulated pure seismic data, simulating noise in actual seismic data, and simulating actual exploration obtained seismic data; step 2, performing NSST transformation on the noise-containing data and the constructed noise data, slidingly intercepting the transformed Shearlet coefficient matrix by using a window with fixed size and sliding step length, and thereby constructing a training set and a test set; step 3, building a multi-level wavelet convolutional neural network, iteratively training the network by using the training set, updating parameters by using the forward and backward propagation of the network, calculating a loss function after each training, and gradually converging the loss function by continuously correcting hyperparameters; and step 4, using the method to denoise actual seismic data. The scheme solves the problems of great difficulty in suppressing various types of noise in seismic exploration data and poor generalization of traditional denoising methods.

Description

Technical Field

[0001] This invention relates to the field of geophysical technology, and in particular to an artificial intelligence seismic data processing method based on NSST and deep learning. Background Technology

[0002] The demand for fossil fuels such as oil and natural gas is increasing daily, and the exploration scope of fossil fuels is gradually shifting from shallow oil and gas reservoirs to deep and complex ones. The discovery of fossil fuels mainly relies on seismic exploration technology. The process of seismic exploration generally involves first acquiring seismic data, then processing the acquired seismic data, and finally interpreting the processed seismic data.

[0003] One of the main challenges currently facing deep oil and gas exploration is the relatively low signal-to-noise ratio of seismic data. Seismic data acquisition is affected by noise generated by complex geological factors, as well as by noise from the natural environment and human activities near the exploration area. These noises have different characteristics such as randomness, single frequency, high energy, and low velocity.

[0004] Currently, various algorithms exist for suppressing noise in seismic signals to improve their signal-to-noise ratio (SNR). These methods are mainly divided into traditional methods and neural network methods. Traditional methods primarily address noise from the perspectives of transform domain, multi-scale decomposition, and mode decomposition, separating noise from effective signals based on the different characteristics of signal and noise in the transform domain and at different scales. While traditional methods can effectively suppress noise in seismic signals, they also suffer from several problems, such as damaging effective signals with the same frequency as noise, difficulty in determining thresholds, and poor generalization. Due to these issues and the ever-increasing volume of seismic data, artificial intelligence denoising methods based on neural networks have emerged. Neural networks use raw data as algorithm input, extracting features through multi-layer network abstraction to achieve the target output. The intermediate parameters in this process are not affected by individual researchers, effectively overcoming the shortcomings of traditional methods. Furthermore, neural network algorithms can be applied to suppressing various types of noise, possessing strong universality and generalization. Modern artificial neural networks have entered the deep learning stage, capable of performing some complex logical operations and implementing some nonlinear mappings. However, large-scale networks with many layers require a lot of computation, often taking up a significant amount of time and computing power. They also face challenges such as difficulty in acquiring datasets and the fact that the feature extraction capabilities of network models depend on the dataset.

[0005] Therefore, how to construct a faster, more efficient, and more generalizable neural network for suppressing seismic noise has become a research hotspot in the field of seismic data processing. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an artificial intelligence seismic data processing method based on NSST and deep learning, aiming to solve problems such as the difficulty in suppressing various types of noise in seismic exploration data, the poor generalization of traditional denoising methods, and the high computational complexity of general convolutional network denoising methods.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0008] An artificial intelligence seismic data processing method based on NSST and deep learning includes the following steps:

[0009] Step 1: Using the clean seismic data synthesized by the Ricker wavelet, noise in the actual seismic data is simulated and generated. The two are then added together to simulate the seismic data obtained from actual exploration.

[0010] Step 2: Perform NSST transformation on the simulated noisy data and the constructed noisy data. Then, use a window with a fixed size and sliding step to slide and truncate the transformed Shearlet coefficient matrix to generate training samples, thereby constructing the training set and the test set.

[0011] Step 3: Build a multi-level wavelet convolutional neural network, determine the appropriate loss function, learning rate, batch size, and number of training iterations, use the above training set for iterative training, update the parameters using the network's backpropagation, calculate the loss function after each training, and gradually converge the loss function by continuously adjusting the hyperparameters. This process will continuously update the network parameters, and finally retain the model parameters with the optimal loss function.

[0012] Step 4: Use this method to denoise the actual seismic data.

[0013] The present invention is further configured such that in step 1:

[0014] Using the mathematical model of seismic wavelet proposed by Lake to simulate pure seismic data: First, a geological model is designed, and the maximum amplitude, Lake wavelet dominant frequency, propagation velocity, and number of in-phase axes are determined;

[0015] The noise in the actual seismic data is simulated, including Gaussian random noise, single-frequency, multi-frequency, single-frequency high-energy and single-frequency low-energy coherent noise, and surface wave noise. The noise data is added to the synthesized clean seismic data to simulate the actual noisy seismic signal.

[0016] The present invention is further configured such that in step 2:

[0017] The simulated data obtained in step 1 is subjected to NSST transformation to obtain its Shearlet coefficient matrix in different directions. The convolutional network is trained on data in blocks, and the key information of the seismic data is related to the local data. Therefore, it is necessary to slide and truncate the Shearlet coefficient matrix of the simulated data in different directions with a window of fixed size and sliding step size to generate training samples, and then construct the training set and test set.

[0018] The present invention is further configured such that in step 3:

[0019] The network has a symmetrical structure. After input, it first passes through a convolutional layer to extract image features, and then reaches the shrinking subnet stage. Each level of the network consists of a wavelet transform DWT and a convolutional block connected sequentially. In the expanding subnet stage, it is completely symmetrical with the shrinking subnet, only needing to transform the wavelet transform DWT into the inverse wavelet transform IWT.

[0020] This invention offers the following advantages: It presents an artificial intelligence algorithm for suppressing seismic signal noise. This algorithm integrates the traditional NSST method with a neural network method based on multi-level wavelet convolution, and performs multiple iterative training sessions using a simulated earthquake training set. When the loss function converges to a relatively small value, the model parameters are preserved. Finally, the denoising effects of traditional denoising methods and the proposed artificial intelligence denoising method are compared on the same real-world seismic data, further demonstrating the advantages and application prospects of the proposed denoising method.

[0021] To fully utilize the inherent geometric features of seismic signals and address the issues of high computational complexity and low efficiency inherent in neural network methods, this invention innovates from two perspectives: the dataset input to the network and the network model. On one hand, by performing NSST transformation on the noisy seismic signal and the noise, the signal distribution in different directions is obtained, fully utilizing the geometric directional features of the seismic signal. Using the coefficient matrix of the Shearlet domain as input to the network model facilitates better signal feature extraction. On the other hand, the wavelet forward transform process replaces the downsampling pooling layer in the U-net network. While the wavelet transform process also achieves dimensionality reduction, its excellent bioorthogonality allows for the reconstruction of the new signal without information loss through inverse wavelet transform, enabling the network to accurately reconstruct the signal in the expanded subnet. Furthermore, this method increases the receptive field without increasing computational complexity, achieving a more balanced effect between model performance and training efficiency. Experimental results show that this invention can efficiently and effectively suppress noise in seismic data without damaging the effective signal, effectively overcoming the shortcomings of traditional denoising methods and general convolutional neural networks. This method makes a positive contribution to improving the signal-to-noise ratio of seismic data and enhancing signal processing efficiency. Attached Figure Description

[0022] Figure 1 This is the main flowchart of the artificial intelligence method for noise suppression of seismic data in this invention;

[0023] Figure 2 This is pure seismic data for synthetic simulation;

[0024] Figure 3 The simulated seismic data after adding noise;

[0025] Figure 4 Here is the flowchart for the NSST transformation;

[0026] Figure 5 To test the denoising results of simulated earthquake data;

[0027] Figure 6 This is a diagram of a multi-level wavelet convolutional neural network structure.

[0028] Figure 7 This is a diagram of the convolutional block structure.

[0029] Figure 8 Diagram of the channel attention mechanism structure;

[0030] Figure 9 This includes actual pre-stack test seismic data and noise types.

[0031] Figure 10 Image showing the denoising results of intelligent methods for actual pre-stack test seismic data;

[0032] Figure 11 This is a diagram showing the denoising results of actual pre-stack test seismic data using conventional methods. Detailed Implementation

[0033] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] An artificial intelligence seismic data processing method based on NSST and deep learning includes the following steps:

[0035] Step 1. First, use the Ricker wavelet to synthesize and simulate clean seismic data, then simulate and generate common types of noise in actual seismic data, and add the two together to simulate the seismic data obtained from actual exploration.

[0036] The detailed process of this step is as follows:

[0037] Step 11: Simulate clean seismic data using the seismic wavelet mathematical model proposed by Lake. First, design a geological model and determine relevant parameters, such as maximum amplitude, Lake wavelet dominant frequency, propagation velocity, and number of in-phase axes, to obtain... Figure 2 The simulated earthquake data shown is shown.

[0038] Step 12: Simulate common types of noise found in actual seismic data, such as Gaussian random noise, single-frequency, multi-frequency, single-frequency high-energy and single-frequency low-energy coherent noise, and surface wave noise. Add the noise data to the clean seismic data synthesized in Step 11 to simulate the actual noisy seismic signal, obtaining, for example... Figure 3 The simulated earthquake data shown is shown.

[0039] Step 2. Perform NSST transformation on the simulated noisy data and the constructed noisy data. Then, use a window with a fixed size and sliding step to truncate the transformed Shearlet coefficient matrix to generate a large number of training samples, thereby constructing the training set and the test set.

[0040] The detailed process of this step is as follows:

[0041] Step 21: Perform NSST transformation on the simulation data obtained in Step 1 to obtain its Shearlet coefficient matrix in different directions, making full use of the geometric directional characteristics of the seismic signal. Figure 4 This is a flowchart of the NSST transformation.

[0042] Step 22: The convolutional network is trained on data in blocks. Since the key information of the seismic data is related to the local data, it is necessary to slide and truncate the Shearlet coefficient matrix of the simulation data in different directions with a window of fixed size and sliding step size. This method will generate a large number of training samples, thereby constructing the training set and the test set.

[0043] Step 3. Construct a multi-level wavelet convolutional neural network and determine appropriate loss functions and hyperparameters such as learning rate, batch size, and number of training iterations. Then, use the training set mentioned above for iterative training, updating parameters using the network's backpropagation. After each training iteration, calculate the loss function and continuously adjust the hyperparameters to allow the loss function to gradually converge. This process continuously updates the network parameters, ultimately retaining the model parameters with the optimal loss function.

[0044] The detailed process of this step is as follows:

[0045] Step 31: The network of this invention has a symmetrical structure, such as... Figure 6 As shown, after input, the image first passes through a convolutional layer to extract features, then proceeds to the shrinking subnetwork stage. Each network level consists of a wavelet transform (DWT) and a convolutional block connected sequentially. The expanding subnetwork stage is completely symmetrical to the shrinking subnetwork, only requiring the wavelet transform (DWT) to be transformed into the inverse wavelet transform (IWT). To increase network depth and improve training efficiency, each module employs residual learning to train the network. Figure 7 The structure diagram of the convolutional block is given. Figure 8 Provide a structural diagram of the channel attention mechanism.

[0046] Step 32: Build the model, determine a suitable optimization algorithm, and set training hyperparameters such as learning rate, batch size, and number of iterations. The loss function is the root mean square error between the reconstructed output and the noise-free samples. Continuously modify the hyperparameters to make the loss function converge to a small value, and save the model parameters with the minimum loss function value. Finally, use the trained model to compare with... Figure 3 By processing the same type of test data, we obtained the following results: Figure 5 The denoising results.

[0047] Step 4. Compare the denoising effects of this method and conventional methods on actual seismic data.

[0048] The detailed process of this step is as follows:

[0049] Step 41: Use the trained optimal model to... Figure 9 The actual test data shown was denoised to obtain... Figure 10 The results are shown.

[0050] Step 42: To demonstrate the advantages of this invention by comparing the differences in noise reduction effects, conventional methods are used to... Figure 9 The test data shown was denoised, and the result is as follows: Figure 11 As shown.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence seismic data processing method based on NSST and deep learning, characterized in that: Includes the following steps: Step 1: Using the clean seismic data synthesized by the Ricker wavelet, noise in the actual seismic data is simulated and generated. The two are then added together to simulate the seismic data obtained from actual exploration. Step 2: Perform NSST transformation on the simulated noisy data and the constructed noisy data to obtain the Shearlet coefficient matrix in different directions. Utilize the geometric direction characteristics of the seismic signal to slide and truncate the Shearlet coefficient matrix in different directions of the simulated data with a window of fixed size and sliding step size to generate training samples, thereby constructing the training set and test set. Step 3: Build a multi-level wavelet convolutional neural network, determine the appropriate loss function, learning rate, batch size, and number of training iterations, use the above training set for iterative training, update the parameters using the network's backpropagation, calculate the loss function after each training, and gradually converge the loss function by continuously adjusting the hyperparameters. This process will continuously update the network parameters, and finally retain the model parameters with the optimal loss function. Step 4: Use this method to denoise the actual seismic data.

2. The artificial intelligence seismic data processing method based on NSST and deep learning as described in claim 1, characterized in that: In step 1: Using the mathematical model of seismic wavelet proposed by Lake to simulate pure seismic data: First, a geological model is designed, and the maximum amplitude, Lake wavelet dominant frequency, propagation velocity, and number of in-phase axes are determined; The noise in the actual seismic data is simulated, including Gaussian random noise, single-frequency, multi-frequency, single-frequency high-energy and single-frequency low-energy coherent noise, and surface wave noise. The noise data is added to the synthesized clean seismic data to simulate the actual noisy seismic signal.

3. The artificial intelligence seismic data processing method based on NSST and deep learning as described in claim 1, characterized in that: In step 2: The simulated data obtained in step 1 is subjected to NSST transformation to obtain its Shearlet coefficient matrix in different directions. The convolutional network is trained on data in blocks, and the key information of the seismic data is related to the local data. Therefore, it is necessary to slide and truncate the Shearlet coefficient matrix of the simulated data in different directions with a window of fixed size and sliding step size to generate training samples, and then construct the training set and test set.

4. The artificial intelligence seismic data processing method based on NSST and deep learning as described in claim 1, characterized in that: In step 3: The network has a symmetrical structure. After input, it first passes through a convolutional layer to extract image features, and then reaches the shrinking subnet stage. Each level of the network consists of a wavelet transform DWT and a convolutional block connected sequentially. In the expanding subnet stage, it is completely symmetrical with the shrinking subnet, only needing to transform the wavelet transform DWT into the inverse wavelet transform IWT.

Images