Multi-band preprocessing machine learning seismic oscillation data denoising method

Abstract

The invention discloses a multi-frequency-band preprocessing machine learning seismic oscillation data denoising method, relates to the technical field of seismic signal processing, and solves the problems that noise and effective signals are aliasing, accurate recognition is difficult and noise remains due to the fact that seismic data cannot be effectively denoised in an existing method. The adopted seismic data de-noising network has a good de-noising effect on noise in a full frequency range, the noise can be effectively separated from effective signals, the effective signals and interference noise can be automatically distinguished through the network, the situation that the effective signals are mistakenly recognized as noise and filtered out is avoided, the two sub-modules are combined together, and the noise reduction efficiency is improved. According to the method, more details can be extracted, the denoising effect is improved, and a good denoising effect is achieved in actually measured seismic data.

Description

Technical Field

[0001] The present invention relates to the field of seismic signal processing, and in particular to a machine learning seismic ground motion data denoising method with multi-band preprocessing. Background Technique

[0002] In seismological research, noise reduction technology, as the core support for key links such as seismic imaging, inversion, and interpretation, can not only significantly improve the success rate of natural gas and oil exploration, but also provide a reliable data basis for underground resource exploration and development. However, the data collected during seismic exploration is often severely contaminated by environmental noise, equipment noise, and multiple wave interference under complex geological conditions. These noises not only reduce the data quality but also pose potential threats to geological interpretation and engineering decisions. Therefore, seismic data denoising is the primary task in seismic signal processing.

[0003] Traditional seismic data denoising techniques usually build models based on assumptions about the characteristics of noise and achieve noise removal through empirical threshold settings. For example, wavelet transform, short-time Fourier transform (STFT), S transform, empirical mode decomposition, etc. Although the signal-to-noise ratio of seismic signals has been improved to a certain extent, these methods have significant limitations. First, with the increasing complexity of the exploration environment and the influence of equipment itself, the diversity and uncertainty of noise in seismic data have increased significantly, resulting in the denoising effect depending on the experience of researchers, lacking stability and universality. Second, the denoising effect of the model depends on the adjustment of the threshold, and the calculation process is complex. Manual parameter tuning for denoising is very inefficient and difficult to meet the dual requirements of high precision and high efficiency in actual engineering.

[0004] In recent years, deep learning technology has been widely applied in fields such as image data processing and speech data processing and has achieved remarkable results. The core difference between deep learning technology and traditional methods is that it does not need to rely on manual experience to set parameters, but automatically learns the inherent characteristics and laws of data from large-scale datasets, enabling the model to have analytical and processing capabilities similar to humans. In addition, the network model trained based on deep learning algorithms has high computational performance and can meet the rapid processing requirements of massive data in three-dimensional seismic exploration. However, when seismic signals and noise are in the same frequency band, existing methods show obvious limitations in signal and noise separation and are difficult to achieve ideal denoising effects. Traditional seismic data processing methods usually rely on complex parameter selection, and these parameters need to be customized and adjusted for different types of signals to achieve the optimal noise reduction effect, but the parameter selection process lacks objectivity, resulting in limited applicability. Therefore, it has important research significance and practical value to develop a machine learning seismic ground motion data denoising method with multi-band preprocessing for complex noise and signal aliasing scenarios. Summary of the Invention

[0005] Aiming at the defects in the prior art, the technical problem to be solved by the present invention is that the existing methods cannot effectively denoise seismic data, resulting in problems such as noise aliasing with effective signals, difficulty in accurate identification, and noise residue.

[0006] It is realized by constructing a machine learning seismic denoising network based on adaptive multi - band pre - processing, including a multi - band pre - processing module and a seismic denoising module, and comprises the following steps:

[0007] S1: Add Gaussian white noise to the collected clean ground motion data as noisy ground motion data;

[0008] S2: The noisy ground motion data is transformed from the time domain to the frequency domain through the short - time Fourier transform (STFT), and adaptive frequency band division is performed for different seismic signals through the hippopotamus optimization algorithm;

[0009] S3: Dynamically adjust the frequency band division points to minimize frequency band overlap, maximize signal consistency, and reduce reconstruction error;

[0010] S4: Quickly find the globally optimal division scheme through iterative optimization, and divide the signal into three frequency bands: low - frequency, medium - frequency, and high - frequency;

[0011] S5: The divided data are respectively subjected to the inverse short - time Fourier transform (ISTFT) to generate corresponding time - domain signals, which are used as the input of the next denoising module;

[0012] S6: The seismic denoising module takes the obtained time - domain ground motion data of different frequencies as input and performs denoising processing through a targeted neural network.

[0013] Beneficial effects

[0014] The seismic data denoising network adopted by the present invention has a good denoising effect on noise in the full - frequency range, and can effectively separate noise from effective signals;

[0015] Through this network, effective signals and interference noise can be automatically distinguished, avoiding the situation where effective signals are misidentified as noise and filtered out;

[0016] In the machine - learning seismic data denoising network structure of the present invention, two sub - modules are combined. Compared with the existing methods, the network of the present invention can extract more details, improve the denoising effect, and in subsequent simulation experiments, the method of the present invention also achieved a good denoising effect in measured seismic data;

[0017] In the denoising method of the present invention, the urban seismic data denoising network does not need to adjust parameters, and can automatically adjust parameters to complete the task of seismic data denoising, greatly saving manpower and material resources. Description of the drawings

[0018] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non - limiting embodiments read in conjunction with the accompanying drawings:

[0019] Figure 1 This is a technical roadmap of a machine - learning seismic ground motion data denoising method with multi - band pre - processing for the present invention;

[0020] Figure 2 This is a schematic diagram of the multi - band pre - processing module;

[0021] Figure 3 This is a schematic diagram of the seismic denoising module;

[0022] Figure 4 This is an effect diagram of the noisy seismic ground motion data; among them, Figure 4 a is the effect diagram of the clean signal, Figure 4 b is the effect diagram of Gaussian white noise, Figure 4 c is the effect diagram of the noisy signal data, Figure 4 d is the effect diagram after denoising the seismic ground motion data by machine learning with multi - band pre - processing;

[0023] Figure 5 This is the frequency - spectrum effect diagram of the seismic data; among them, Figure 5 a is the frequency - spectrum effect diagram of the pure signal data, Figure 5 b is the frequency - spectrum effect diagram of Gaussian white noise, Figure 5 c is the frequency - spectrum effect diagram of the noisy signal data, Figure 5 d is the frequency - spectrum effect diagram after denoising the seismic ground motion data by machine learning with multi - band pre - processing. Detailed implementation manners

[0024] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made. These all belong to the protection scope of the present invention.

[0025] Embodiment 1: A machine - learning seismic ground motion data denoising method with multi - band pre - processing, which is realized by constructing a machine - learning seismic denoising network based on adaptive multi - band pre - processing, including a multi - band pre - processing module and a seismic denoising module. Figure 1 The following describes the technical roadmap of a machine - learning seismic ground motion data denoising method with multi - band pre - processing.

[0026] As Figure 2As shown in the figure, in the multi - band pre - processing module, the collected clean seismic vibration data is added with Gaussian white noise to obtain noisy seismic vibration data. After the noisy seismic vibration data is transformed from the time domain to the frequency domain through the short - time Fourier transform (STFT), the Hippopotamus optimization algorithm is used to perform adaptive frequency - band division for different seismic signals; the frequency - band division points are dynamically adjusted to minimize frequency - band overlap, maximize signal consistency, and reduce the reconstruction error. Through iterative optimization, a globally optimal division scheme is quickly found, and the signal is divided into three frequency bands: low - frequency, medium - frequency, and high - frequency. The divided data is respectively passed through the inverse short - time Fourier transform (ISTFT) to generate corresponding time - domain signals, which are used as the input for the next denoising module;

[0027] As Figure 3 shown, the seismic denoising module takes the obtained time - domain seismic vibration data of different frequencies as input and performs denoising processing using a targeted neural network. Utilizing the powerful learning ability of the neural network, efficient and accurate denoising is achieved. The shape of each individual time - domain input data is (100, 2000, 1), and the shape of the output data is (100 2000, 1);

[0028] The first low - frequency signal usually contains the main energy components of the seismic vibration, with strong time - dependence and stationarity. In view of this characteristic, LSTM - S is used for denoising: LSTM - S consists of an LSTM and a fully - connected layer, and the folding technique is introduced on the basis of the traditional LSTM network. The network depth of the time - series modeling module is set to 7 layers, which is composed of 1 input layer, 4 LSTM layers, 3 fully - connected layers, and 1 output layer; the size of the convolution kernel of each layer is 3×3; this technology of the time - series modeling module reduces the redundant calculations in the training process by optimizing the network structure and calculation process, significantly accelerating the convergence speed of the model, and at the same time reducing the consumption of computing resources.

[0029] The intermediate-frequency signal is denoised using the AttResU-Net neural network, which is an improved U-Net network that combines the attention mechanism and residual connections, focusing on the extraction of detailed features and denoising of intermediate-frequency signals. Its network structure consists of 4 encoder layers and 4 decoder layers. Each layer is composed of two 3×3 convolutional layers, batch normalization, and ReLU activation functions. Residual connections are added after each convolutional block to alleviate the vanishing gradient problem and improve the feature transfer efficiency. An attention gate is introduced between the encoder and decoder. By calculating the correlation between the encoder features and decoder features, an attention weight map is dynamically generated to enhance the expression ability of important features while suppressing irrelevant features. The multi-scale feature maps of the encoder and decoder are fused through skip connections to retain the detailed information of the signal. Finally, the network maps the feature map to a single-channel output through a 1×1 convolutional layer to generate the denoised intermediate-frequency signal. AttResU-Net combines the multi-scale feature extraction ability of the attention mechanism and the efficient training characteristics of residual connections, significantly improving the denoising accuracy and detail retention ability of intermediate-frequency signals, and is suitable for intermediate-frequency signal processing tasks with different noise levels and signal characteristics. We selected ReLU as the activation function to introduce non-linear characteristics. The advantage of ReLU is that it can effectively avoid the vanishing gradient problem and overfitting, thus accelerating the convergence process of the network.

[0030] The high-frequency signal, which is the most difficult part to process, is processed by SparseNet, which focuses on noise separation and denoising of high-frequency signals. Its network structure includes a sparse coding layer, 4 convolutional modules, a sparse regularization layer, 4 deconvolutional modules, and an output layer. The sparse coding layer extracts the sparse features of high-frequency signals through the sparse coding algorithm. The convolutional modules and deconvolutional modules are used for feature extraction and signal resolution restoration respectively. The sparse regularization layer constrains the sparsity of the output signal through L1 regularization. Finally, the denoised high-frequency signal is output through a 1×1 convolutional layer. This network combines sparse representation theory and deep learning to achieve efficient noise separation through end-to-end training, and has the characteristics of strong robustness and high interpretability, and is suitable for high-frequency signal denoising in complex noise environments.

[0031] The output results of the three neural networks are fused using a weighted fusion strategy. Weights are assigned according to the importance of signals in each frequency band, and the finally output is the denoised time-domain signal after fusion. During the training process, clean ground motion data is used as the supervision signal, and the network parameters are optimized by comparing the denoised results with the real signal. Through multi-frequency band preprocessing and parallel computing, the computing efficiency of the denoising module is improved, making it suitable for processing large-scale ground motion data.

[0032] Example 2: As Figures 4 - 5As shown in the figure, the application of a machine learning seismic ground motion data denoising method with multi - band pre - processing in Embodiment 1: In this example, the feasibility of this method is verified through seismic ground motion data.

[0033] First, establish a suitable data set: There are 500 known clean seismic signals with a signal length of 2000 and a sampling frequency of 100 Hz. A data set of 500 noisy seismic signals is obtained by adding Gaussian white noise with a signal - to - noise ratio of 4 dB. The data set is divided into a training set, a test set, and a validation set according to the ratio of 8:1:1. As Figure 4 shown, where Figure 4 a is the effect diagram of the clean signal, Figure 4 b is the effect diagram of Gaussian white noise, Figure 4 c is the effect diagram of the noisy signal data, Figure 4 d is the effect diagram of the data after adaptive denoising of seismic ground motion data based on machine learning; As Figure 5 shown, Figure 5 a is the spectrum effect diagram of the pure signal data, Figure 5 b is the spectrum effect diagram of Gaussian white noise, Figure 5 c is the spectrum effect diagram of the noisy signal data, Figure 5 d is the spectrum effect diagram after adaptive denoising of seismic ground motion data based on machine learning.

[0034] Through Figure 4 and Figure 5 shown, it can be seen that the noise and the signal are effectively separated; The present invention has a good inhibitory effect on seismic noise, retains the effective signal as much as possible in a complex noise environment, and provides a solid foundation for subsequent seismic data processing.

[0035] By dividing the seismic signal into frequency bands and combining with the machine learning denoising method, compared with the traditional threshold denoising and traditional CNN denoising methods, the present invention shows significant advantages in both denoising effect and efficiency. First, frequency - band division can decompose the signal into different frequency sub - bands, making it easier to separate noise and effective signals in the frequency domain, thus providing a more refined input for subsequent denoising processing. On this basis, the machine learning method can more accurately identify and remove noise by adaptively learning the characteristics of noise and signals, avoiding the limitations of the fixed threshold in the traditional threshold denoising method, reducing the dependence on manual parameter adjustment, and enhancing the self - adaptability and robustness of denoising. In addition, compared with the traditional CNN denoising method, the strategy of frequency - band division combined with machine learning can significantly reduce the computational complexity, improve the denoising efficiency, and avoid the over - fitting problem that may occur in the CNN method, further enhancing the generalization ability of the model.

Claims

1. A machine learning seismic ground motion data denoising method with multi - band pre - processing, characterized in that: It is realized by constructing a machine learning seismic denoising network based on adaptive multi-band preprocessing, including a multi-band preprocessing module and a seismic denoising module, and includes the following steps: S1: Add Gaussian white noise to the collected clean ground motion data as noisy ground motion data; S2: The noisy ground motion data is transformed from the time domain to the frequency domain through the short-time Fourier transform (STFT), and the adaptive frequency band division is carried out for different seismic signals through the hippopotamus optimization algorithm; S3: Dynamically adjust the frequency band division points to minimize the frequency band overlap, maximize the signal consistency and reduce the reconstruction error; S4: Through iterative optimization, quickly find the globally optimal division scheme, and divide the signal into three frequency bands: low frequency, medium frequency and high frequency; S5: The divided data are respectively passed through the inverse short-time Fourier transform (ISTFT) to generate corresponding time-domain signals, which are used as the input of the next denoising module; S6: The seismic denoising module uses the obtained time-domain seismic ground motion data with different frequencies as input to perform denoising processing on the targeted neural network.

2. A machine learning seismic ground motion data denoising method for multi - band pre - processing according to claim 1, characterized in that: The low-frequency signal is denoised by LSTM-S, which is composed of LSTM and a fully connected layer, and the folding technology is introduced on the basis of the traditional LSTM network.

3. A machine learning seismic ground motion data denoising method for multi-band preprocessing according to claim 2, characterized in that: The network depth of the time series modeling module is set to 7 layers, which consists of 1 input layer, 4 LSTM layers, 3 fully connected layers and 1 output layer; the size of the convolution kernel of each layer is 3×3.

4. A machine learning seismic ground motion data denoising method for multi - band pre - processing according to claim 1, characterized in that: The medium-frequency signal is denoised by the AttResU-Net neural network. Its network structure includes 4 layers of encoders and 4 layers of decoders. Each layer consists of two 3×3 convolutional layers, batch normalization and ReLU activation functions, and residual connections are added after each convolutional block to alleviate the problem of gradient disappearance and improve the feature transfer efficiency.

5. A machine learning seismic ground motion data denoising method with multi - band pre - processing according to claim 4, characterized in that: An attention gate is introduced between the encoder and the decoder. By calculating the correlation between the encoder features and the decoder features, an attention weight map is dynamically generated to enhance the expression ability of important features and suppress irrelevant features at the same time.

6. A machine learning seismic ground motion data denoising method for multi - band pre - processing according to claim 5, characterized in that: The multi-scale feature maps of the encoder and the decoder are fused through skip connections to retain the detailed information of the signal. The network maps the feature map to a single-channel output through a 1×1 convolutional layer to generate the denoised medium-frequency signal.

7. A machine learning seismic ground motion data denoising method with multi - band pre - processing according to claim 1, characterized in that: The high-frequency signal is processed by SparseNet, and its network structure includes a sparse coding layer, 4 convolutional modules, a sparse regularization layer, 4 deconvolutional modules and an output layer.

8. A machine learning seismic ground motion data denoising method for multi-band preprocessing according to claim 7, characterized in that: The sparse coding layer extracts the sparse features of the high-frequency signal through the sparse coding algorithm. The convolutional module and the deconvolutional module are used for feature extraction and signal resolution recovery respectively. The sparse regularization layer constrains the sparsity of the output signal through L1 regularization, and outputs the denoised high-frequency signal through a 1×1 convolutional layer.

Images