A low signal-to-noise ratio microseismic signal denoising method and system

Abstract

The application belongs to the technical field of microseismic signal processing, and in particular to a low signal-to-noise ratio microseismic signal denoising method and system. The method comprises the following steps: collecting multiple seismic time sequence signals; reconstructing the multiple seismic time sequence signals by using multivariate singular spectrum analysis to obtain reconstructed signals; splicing the reconstructed signals and the original multiple seismic time sequence signals to obtain a two-channel input data set; and inputting the two-channel input data set into a U-Net network for end-to-end recovery. The application significantly improves the recovery capability and phase preservation of weak seismic signals in an ultra-low signal-to-noise ratio environment.

Description

Technical Field

[0001] This application belongs to the field of microseismic signal processing technology, specifically a method and system for denoising low signal-to-noise ratio microseismic signals. Background Technology

[0002] Microseismic monitoring often faces weak seismic waves with extremely low signal-to-noise ratios (SNR), and current microseismic and seismic denoising techniques have significant limitations under ultra-low SNR conditions. While traditional filtering and time-frequency transformation can suppress noise, they often lose phase and subtle waveform details, reducing source inversion accuracy. MSSA (Multivariate Singular Spectrum Analysis) is highly sensitive to embedding window length and truncation rank, and its reconstruction capability decreases in non-stationary or noisy environments. Pure time-domain deep networks are prone to over-smoothing, phase distortion, and neglect of physical priors, resulting in poor generalization. Frequency-domain methods alone struggle to fully utilize inter-channel correlations. In summary, existing methods struggle to simultaneously preserve phase and restore details while maintaining robustness and computational efficiency in extremely low SNR scenarios, thus necessitating new denoising schemes that combine physical priors with frequency domain awareness. Summary of the Invention

[0003] This application provides a method and system for denoising microseismic signals with low signal-to-noise ratio, solving the problems of denoising and phase preservation under ultra-low signal-to-noise ratio conditions in the fields of microseismic and seismic events.

[0004] The first aspect of this application provides a method for denoising low signal-to-noise ratio microseismic signals, the method comprising:

[0005] Acquire multiple seismic time-series signals;

[0006] Multivariate singular spectrum analysis was used to reconstruct multichannel seismic time series signals to obtain reconstructed signals;

[0007] The reconstructed signal is concatenated with the original multichannel seismic time series signal and used as a dual-channel input dataset.

[0008] The dual-channel input dataset is fed into the U-Net network for end-to-end recovery.

[0009] Furthermore, the reconstruction of multichannel seismic time-series signals using multivariate singular spectral analysis includes:

[0010] Construct a Hankel matrix for each channel of the multi-channel seismic time series signal;

[0011] Concatenate the Hankel matrices of all tracks column-wise to form a joint trajectory matrix;

[0012] The joint trajectory matrix is ​​decomposed using singular value decomposition;

[0013] The number of principal components to be retained is determined based on the distribution of the singular values. The top r largest singular values ​​and their corresponding singular vectors are retained, and the remaining singular values ​​are set to zero to obtain the denoised low-rank approximation matrix.

[0014] The low-rank approximation matrix is ​​decomposed into submatrices corresponding to each channel signal, and the diagonal averaging method is applied to each submatrix to inversely map the matrix signal into a one-dimensional time series signal, thus obtaining the reconstructed signals of each channel.

[0015] Furthermore, the U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein:

[0016] The encoder extracts multi-scale temporal features of the input signal through layer-by-layer downsampling operations. Each layer of downsampling operation gradually reduces the feature map resolution and increases the number of channels in the feature map.

[0017] The bottleneck layer processes global semantic features from the encoder's output data while minimizing the dimension of the feature map, establishing dependencies over long time spans.

[0018] The decoder recovers the temporal resolution of the output data from the bottleneck layer by upsampling layer by layer, and uses skip connections to fuse the features of the encoder and the decoder.

[0019] Furthermore, the encoder includes a multi-level cascaded first module, which sequentially includes: a first multi-scale temporal convolution module, a first frequency domain sensing processing module, and an output convolutional layer;

[0020] The first multi-scale temporal convolution module employs multiple parallel convolution branches, with the kernel size of different convolution branches used to capture waveform features at different time scales;

[0021] The first frequency domain sensing processing module performs a fast Fourier transform on the output signal of the first multi-scale time domain convolution module, transforming the signal from the time domain to the frequency domain. The real and imaginary parts in the frequency domain are concatenated together and input into a convolutional layer for processing. The output signal of the convolutional layer is transformed back to the time domain using an inverse Fourier transform, and frequency-related local features are extracted.

[0022] Furthermore, the bottleneck layer includes a second multi-scale temporal convolution module and a second frequency domain sensing processing module. The second multi-scale temporal convolution module employs multiple parallel convolution branches, and the kernel sizes of different convolution branches are used to capture waveform features at different time scales and then integrate them into global features.

[0023] The second frequency domain sensing processing module performs a fast Fourier transform on the data integrated into global features, converting the signal from the time domain to the frequency domain. It then concatenates the real and imaginary parts in the frequency domain and inputs them into a convolutional layer for processing. The output signal of the convolutional layer is converted back to the time domain using an inverse Fourier transform, and all frequency-related features are extracted.

[0024] Furthermore, the decoder includes multiple cascaded second modules, each including: a transposed convolutional layer for upsampling to expand the time dimension and restore the signal resolution; and a third multi-scale temporal convolutional module that employs multiple parallel convolutional branches, with the kernel sizes of different convolutional branches used to capture waveform features at different time scales and then integrate them for feature fusion.

[0025] Furthermore, the upsampled feature map in the decoder is concatenated with the shallow feature map of the corresponding layer in the encoder through skip connections. The concatenated data is then input into the third multi-scale temporal convolution module for convolution processing to fuse deep semantic information and shallow detail information and repair the signal waveform.

[0026] Furthermore, the decoder includes an output convolutional layer that maps high-dimensional feature vectors back to the channel dimension of the original signal.

[0027] A low signal-to-noise ratio microseismic signal denoising system according to a second aspect embodiment of this application includes:

[0028] The preprocessing module is used to reconstruct multichannel seismic time series signals using multivariate singular spectrum analysis to obtain reconstructed signals. The reconstructed signals are then concatenated with the original multichannel seismic time series signals to form a dual-channel input dataset.

[0029] The U-Net network is used for end-to-end recovery of dual-channel input datasets.

[0030] Furthermore, the U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein:

[0031] The encoder extracts multi-scale temporal features of the input signal through layer-by-layer downsampling operations. Each layer of downsampling operation gradually reduces the feature map resolution and increases the number of channels in the feature map.

[0032] The bottleneck layer processes global semantic features from the encoder's output data while minimizing the dimension of the feature map, establishing dependencies over long time spans.

[0033] The decoder recovers the temporal resolution of the output data from the bottleneck layer by upsampling layer by layer, and uses skip connections to fuse the features of the encoder and the decoder.

[0034] Compared with the prior art, the beneficial effects of this application are as follows: This application significantly improves the recovery capability and phase preservation of weak seismic signals in ultra-low signal-to-noise ratio environments, thereby improving the robustness and accuracy of subsequent source inversion.

[0035] This application organically integrates multivariate singular spectral analysis (MSSA) as an explicit physical prior and a learnable module of a deep network. It provides raw and low-rank reconstruction information in parallel through dual channels, preserving the common structure among multiple channels while providing reliable initial values ​​for the network. Under extremely low signal-to-noise ratio conditions, it recovers details and phase better than existing techniques, reduces parameter sensitivity, and improves generalization, providing a more reliable preprocessing scheme for practical microseismic denoising and subsequent inversion work.

[0036] A frequency domain sensing processing module is introduced into the network to perform frequency domain convolution and inverse transformation on intermediate features by splicing real and imaginary parts. The phase deviation is clearly compensated by learnable spectral correction, avoiding the over-smoothing and phase distortion common in pure time-domain networks. At the same time, multi-scale time-domain convolution is used to enhance the perception of features at different time scales, and residual learning is used to focus on small corrections to improve training stability. Attached Figure Description

[0037] Figure 1 A flowchart of a low signal-to-noise ratio microseismic signal denoising method provided in this application embodiment;

[0038] Figure 2 A structural block diagram of a low signal-to-noise ratio microseismic signal denoising system provided in this application embodiment;

[0039] Figure 3 This is a schematic diagram of the U-Net network structure provided in an embodiment of this application;

[0040] Figure 4 The following are the noise reduction results of the measured data provided in the embodiments of this application: (a) is a data comparison diagram of the first detector, (b) is a data comparison diagram of the second detector, (c) is a data comparison diagram of the third detector, (d) is a data comparison diagram of the fourth detector, (e) is a data comparison diagram of the fifth detector, and (f) is a data comparison diagram of the sixth detector.

[0041] Figure 5 The image shows the noise reduction result of the simulated data provided in the embodiments of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] See Figure 1The method shown is a low signal-to-noise ratio microseismic signal denoising method, which includes:

[0044] Acquire multiple seismic time-series signals;

[0045] Multivariate singular spectrum analysis was used to reconstruct multichannel seismic time series signals to obtain reconstructed signals;

[0046] The reconstructed signal is concatenated with the original multichannel seismic time series signal and used as a dual-channel input dataset.

[0047] The dual-channel input dataset is fed into the U-Net network for end-to-end recovery.

[0048] In the data preparation phase, the first step is to collect multiple seismic time-series signals. These signals are generated by... It consists of channels, and the time series length of each channel is [length missing]. , The data dimension of the seismic time series signal is ,in It is the time length of each seismic time sequence signal. This refers to the number of traces in the seismic time series signal. Each trace of the seismic time series signal... It is a length of One-dimensional time series data, representing the data at different time points. The amplitude of the signals needs to be normalized to ensure a consistent amplitude range. Then, a window function is used to reduce boundary effects in the time domain. The window function is typically applied to each seismic time series signal for subsequent multivariate singular spectrum analysis.

[0049] Multivariate singular spectral analysis is used to reconstruct multichannel seismic time series signals to obtain reconstructed signals. In one embodiment, this includes:

[0050] Construct a Hankel matrix for each seismic time series signal;

[0051] Concatenate the Hankel matrices of all tracks column-wise to form a joint trajectory matrix;

[0052] The joint trajectory matrix is ​​decomposed using singular value decomposition;

[0053] The number of principal components to be retained is determined based on the distribution of the singular values. The top r largest singular values ​​and their corresponding singular vectors are retained, and the remaining singular values ​​are set to zero to obtain the denoised low-rank approximation matrix.

[0054] The low-rank approximation matrix is ​​decomposed into submatrices corresponding to each channel signal, and the diagonal averaging method is applied to each submatrix to inversely map the matrix signal into a one-dimensional time series signal, thus obtaining the reconstructed signals of each channel.

[0055] Specifically, a Hankel matrix is ​​constructed for each seismic time series signal. This is done for each preprocessed seismic time series signal. Where g = 1, 2, ..., G, construct the corresponding trajectory matrices. For a single-track time series of length T, select the embedding window length L, usually L = T / 2, and construct the Hankel matrix. ,in , Represents the set of real numbers. Hankel matrix. The form is: ,

[0056] This step maps the one-dimensional time series signal into a two-dimensional matrix, enabling the extraction of local structural features of the seismic signal in phase space.

[0057] Construct the joint trajectory matrix and apply singular value decomposition (SVD). To utilize the spatial correlation between multi-channel signals, the Hankel matrices of all G channels are concatenated column-wise to form a large joint trajectory matrix. : , Let H be the Hankel matrix of the G detectors. Then, the joint trajectory matrix is ​​decomposed using singular value decomposition (SVD) to obtain: ,in , is the left singular vector matrix of the joint trajectory matrix. , is the singular value matrix of the joint trajectory matrix. , is the right singular vector matrix of the joint trajectory matrix. For transpose, It is the truncated rank, representing the number of singular values ​​retained.

[0058] The truncation rank r (i.e., the number of principal components to be retained) is determined based on the distribution of singular values. The top r largest singular values ​​and their corresponding singular vectors are retained, while the smaller, noise-dominant singular values ​​are set to zero, thus obtaining the denoised low-rank approximation matrix, i.e.: ,in, It is the left singular vector matrix of the low-rank approximation matrix obtained by singular value decomposition. It is the singular value matrix of the low-rank approximation matrix obtained by singular value decomposition. It is the transpose of the right singular vector matrix of the low-rank approximation matrix obtained by singular value decomposition. It is a low-rank approximation matrix.

[0059] Diagonal Averaging Reconstruction: The low-rank approximation matrix is ​​decomposed into submatrices corresponding to each signal path. Then, the diagonal averaging method is applied to each submatrix to inversely map the matrix signal into a one-dimensional time series signal, thereby obtaining the reconstructed signals of each channel. After low-rank reconstruction by multivariate singular spectral analysis, the reconstructed signal is obtained. and the original signal They are concatenated into a dual-channel input tensor. Indicates the first Reconstructing signals Indicates the first The original signal. Specifically, the original signal... With MSSA reconstructed signal By concatenating the data, a two-channel input dataset is obtained. , This dual-channel input dataset provides two types of information for subsequent deep learning models: the original signal and the reconstructed signal based on physical priors, enabling the network to utilize both types of information simultaneously for more effective denoising and restoration.

[0060] In one embodiment, see Figure 3 As shown, the U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein:

[0061] The encoder extracts multi-scale temporal features of the input signal through layer-by-layer downsampling operations. Each layer of downsampling operation gradually reduces the feature map resolution and increases the number of channels in the feature map.

[0062] The bottleneck layer processes global semantic features from the encoder's output data while minimizing the dimension of the feature map, establishing dependencies over long time spans.

[0063] The decoder recovers the temporal resolution of the output data from the bottleneck layer by upsampling layer by layer, and uses skip connections to fuse the features of the encoder and the decoder.

[0064] The encoder includes a first module with multiple cascaded layers, which in turn includes: a first multi-scale temporal convolution module, a first frequency domain sensing processing module, and an output convolutional layer.

[0065] The first multi-scale temporal convolution module employs multiple parallel convolution branches, with the kernel size of different convolution branches used to capture waveform features at different time scales;

[0066] The first frequency domain sensing processing module performs a fast Fourier transform on the output signal of the first multi-scale time domain convolution module, transforming the signal from the time domain to the frequency domain. The real and imaginary parts in the frequency domain are concatenated together and input into a convolutional layer for processing. The output signal of the convolutional layer is transformed back to the time domain using an inverse Fourier transform, and frequency-related local features are extracted.

[0067] The bottleneck layer includes a second multi-scale temporal convolution module and a second frequency domain sensing processing module. The second multi-scale temporal convolution module uses multiple parallel convolution branches. The kernel size of different convolution branches is used to capture waveform features at different time scales and then integrate them into global features.

[0068] The second frequency domain sensing processing module performs a fast Fourier transform on the data integrated into global features, converting the signal from the time domain to the frequency domain. It then concatenates the real and imaginary parts in the frequency domain and inputs them into a convolutional layer for processing. The output signal of the convolutional layer is converted back to the time domain using an inverse Fourier transform, and all frequency-related features are extracted.

[0069] The decoder includes multiple cascaded second modules, each including: a transposed convolutional layer for upsampling to expand the time dimension and restore the signal resolution; and a third multi-scale temporal convolutional module that employs multiple parallel convolutional branches, with the kernel sizes of different convolutional branches used to capture waveform features at different time scales and then integrate them for feature fusion.

[0070] The upsampled feature map in the decoder is concatenated with the shallow feature map of the corresponding layer in the encoder through skip connections. The concatenated data is then input into the third multi-scale temporal convolution module for convolution processing to fuse deep semantic information and shallow detail information and repair the signal waveform.

[0071] At the end of the U-Net network, the feature maps processed by the decoder are mapped back to the original signal channels through a 1×1 convolutional layer, outputting the noise residuals after residual learning correction. Then, this residual is added to the original signal D to obtain the final denoised signal. , This residual learning strategy allows the network to focus on correcting small errors in the signal, thereby better recovering the original details of the signal, rather than reconstructing the signal from scratch.

[0072] In this application, the frequency domain sensing processing module does not refer to the time-domain convolution being equal to frequency-domain multiplication in signal processing theory. Instead, it refers to directly extracting features from the frequency domain feature map of the signal using a convolutional neural network (CNN) layer. The specific process is as follows: first, the signal is subjected to a Fast Fourier Transform (FFT) to transform it into the frequency domain; then, the real and imaginary parts of the complex result are concatenated as two independent feature channels; subsequently, a sliding window operation is performed on this frequency domain feature map using a convolutional layer. The purpose of this step is to allow the network to learn the nonlinear relationship between the real and imaginary parts, directly capturing the differences in frequency distribution between the effective signal and noise, thereby achieving explicit modeling and optimization of the spectral features.

[0073] Explanation of the difference between the bottleneck layer and the encoder: Although the bottleneck layer (Bottleneck) is structurally similar to the encoder, containing multi-scale temporal convolutional modules and frequency-domain sensing processing modules, their data states and functions are completely different. In the encoder, the first multi-scale temporal convolutional module and the first frequency-domain sensing processing module are mainly responsible for dimensionality reduction feature extraction, that is, compressing the temporal dimension layer by layer to extract features from local details to abstract semantics. The bottleneck layer, located at the deepest part of the network, has its temporal dimension of the feature map compressed to a minimum, and the number of channels maximized. This is the narrowest point of information flow, converging highly compressed global contextual information of the entire signal. The second multi-scale temporal convolutional module and the second frequency-domain sensing processing module in the bottleneck layer no longer focus on local details but are responsible for processing these global semantic features, establishing dependencies over long time spans, and serving as a crucial bridge connecting feature extraction (encoding) and signal reconstruction (decoding). Therefore, despite the same structure, due to their different layers, their role shifts from feature extraction to global information fusion.

[0074] The decoding process at each layer comprises three key steps: First, upsampling, where data passes through a transposed convolutional layer to expand the temporal dimension and restore signal resolution; second, feature fusion, where the upsampled feature map is concatenated with shallow feature maps from the corresponding layers in the encoder via skip connections. This step aims to recover spatial details such as waveform edges lost during downsampling; and finally, reconstruction, where the concatenated data is input into a third multi-scale temporal convolutional module for convolutional processing to fuse deep semantic information with shallow detail information, thus restoring the signal waveform. The final convolution in the decoder is typically a single convolutional layer. Its function is not to extract spatiotemporal features, but rather to perform channel mapping, mapping the high-dimensional feature vector output from the last layer of the decoder back to the channel dimension of the original signal, thereby outputting the final denoised residual or reconstructed signal.

[0075] During U-Net network training, the pointwise mean squared error (MSE) is used as the loss function to calculate the difference between the network output and the true signal. The formula for calculating the loss function is:

[0076] Each seismic time series signal This represents the true value of the g-th signal at time t. It refers to the time duration, and the reconstructed signals of each channel. This is the denoised signal output by the U-Net network. By minimizing this loss function, the U-Net network learns how to remove noise while preserving important information in the signal.

[0077] During training, the Adam optimizer was used to adjust the learning rate, and an early stopping mechanism was employed to prevent overfitting. To improve the model's generalization ability, data augmentation (such as time shifting and amplitude scaling) was also applied during training.

[0078] After training, the network can process new multichannel seismic time-series data and output denoised signals. During the inference phase, the input data is first reconstructed, and then the original and reconstructed data are concatenated into a dual-channel input dataset, which is then fed into the U-Net network for forward propagation. Finally, the network outputs a denoised signal, and the quality of the denoising results is evaluated using metrics such as SSIM (Structural Similarity Index) and LocalSim (Local Similarity), ensuring that the signal-to-noise ratio and structural similarity of the output are improved while preserving phase and waveform details.

[0079] To further verify the actual performance of the proposed method in on-site data denoising scenarios, the trained model was directly applied to on-site data collected in one embodiment, without the need for transfer learning and retraining. For example... Figure 4 To demonstrate its effectiveness in real-world testing, analysis of six detectors revealed significant noise reduction results. Figure 4 (a) in the figure is a data comparison diagram of the first detector. Figure 4 (b) in the diagram is a data comparison chart of the second detector. Figure 4 (c) in the diagram is a data comparison chart of the third detector. Figure 4 (d) in the graph is a data comparison chart of the fourth detector. Figure 4 (e) in the diagram is a data comparison chart of the fifth detector. Figure 4 (f) in the diagram is a data comparison chart of the sixth detector. By comparing the original data with the data processed by this application (represented using MFU-Net), it can be seen that the noise is significantly suppressed. Figure 5 The results shown are from the simulation data. By adding noise with a signal-to-noise ratio of -5dB, the following can be obtained: Figure 5 Results. Analysis shows that this application significantly expands the feature perception range by virtue of its innovative multi-scale convolutional architecture and dual-channel input strategy. Furthermore, through joint frequency-time domain feature extraction, it successfully eliminates boundary effects, achieving a high degree of overlap between the denoising result and the original signal. This finding demonstrates that this application not only exhibits optimal denoising performance across the entire signal-to-noise ratio range but also overcomes the technical bottlenecks of traditional methods through structural design.

[0080] On the other hand, this application provides a low signal-to-noise ratio microseismic signal denoising system, see [link to relevant documentation]. Figure 2 As shown, it includes:

[0081] The preprocessing module is used to reconstruct multichannel seismic time series signals using multivariate singular spectrum analysis to obtain reconstructed signals. The reconstructed signals are then concatenated with the original multichannel seismic time series signals to form a dual-channel input dataset.

[0082] The U-Net network is used for end-to-end recovery of dual-channel input datasets.

[0083] The preprocessing module employs a physical prior reconstruction followed by learnable time-frequency correction, aiming to recover weak seismic waveforms while preserving phase information as much as possible under extremely low signal-to-noise ratio conditions. First, the original multi-channel seismic time series signals undergo uniform preprocessing and windowing to ensure each sample has the same time length and number of channels. Then, low-rank reconstruction based on multivariate singular spectrum analysis and subsequent deep learning recovery are simultaneously applied to the preprocessed multi-channel seismic time series signals. Multivariate singular spectrum analysis constructs a Hankel matrix embedding from the one-dimensional time series data of each channel and concatenates the sub-matrices of each channel into a large data matrix. Singular value decomposition is then performed on the large data matrix, and several principal components are truncated to obtain a low-rank approximation. Subsequently, diagonal averaging is used to map the reconstructed sub-blocks back to each channel's time series, thereby extracting the correlation and low-rank structure between multiple channels as explicit physical priors to suppress noise while preserving the structural information of common components.

[0084] In one embodiment, the U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein:

[0085] The encoder extracts multi-scale temporal features of the input signal through layer-by-layer downsampling operations. Each layer of downsampling operation gradually reduces the feature map resolution and increases the number of channels in the feature map.

[0086] The bottleneck layer processes global semantic features from the encoder's output data while minimizing the dimension of the feature map, establishing dependencies over long time spans.

[0087] The decoder recovers the temporal resolution of the output data from the bottleneck layer by upsampling layer by layer, and uses skip connections to fuse the features of the encoder and the decoder.

[0088] The main network employs a U-Net structure to achieve multi-scale temporal feature extraction and layer-by-layer reconstruction. Skip connections preserve local details during downsampling and restore temporal resolution during upsampling. To enhance the perception of waveform features at different time scales, the U-Net network introduces parallel multi-scale temporal convolutional modules at several key locations to simultaneously capture short-term abrupt changes and long-term periodic components, thereby helping the U-Net network to reconstruct the morphology and texture of weak waveforms more meticulously in the temporal domain.

[0089] To compensate for the shortcomings of pure time-domain processing in modeling the spectrum and phase, the U-Net network incorporates a frequency-aware processing module (FDP). The FDP module first performs a Fast Fourier Transform (FFT) on the intermediate time-domain features along the time axis. After concatenating the real and imaginary parts, it learns spectral corrections in the frequency domain through a series of convolutional blocks. The corrected features are then reconstructed into a complex spectrum, which is then inversely transformed back to the time domain and added to the original time-domain features as a residual. Through this jointly time-frequency learnable correction, the U-Net network can both rely on the multi-channel low-rank priors provided by the multi-scale time-domain convolutional modules and compensate for phase deviations and frequency band distortions at the spectral level, thereby achieving more robust phase preservation and waveform restoration.

[0090] Specifically, the insertion positions of the frequency domain sensing processing module and the multi-scale temporal convolution module are as follows, with the first, second, and third descriptions used to distinguish the different positions:

[0091] The encoder includes a multi-level cascaded first module, which sequentially includes: a first multi-scale temporal convolution module, a first frequency domain sensing processing module, and an output convolutional layer;

[0092] The first multi-scale temporal convolution module employs multiple parallel convolution branches, with the kernel size of different convolution branches used to capture waveform features at different time scales;

[0093] The first frequency domain sensing processing module performs a fast Fourier transform on the output signal of the first multi-scale time domain convolution module, transforming the signal from the time domain to the frequency domain. The real and imaginary parts in the frequency domain are concatenated together and input into a convolutional layer for processing. The output signal of the convolutional layer is transformed back to the time domain using an inverse Fourier transform, and frequency-related local features are extracted.

[0094] The bottleneck layer includes a second multi-scale temporal convolution module and a second frequency domain sensing processing module. The second multi-scale temporal convolution module uses multiple parallel convolution branches. The kernel size of different convolution branches is used to capture waveform features at different time scales and then integrate them into global features.

[0095] The second frequency domain sensing processing module performs a fast Fourier transform on the data integrated into global features, transforming the signal from the time domain to the frequency domain. It then concatenates the real and imaginary parts in the frequency domain and inputs them into a convolutional layer for processing. The output signal of the convolutional layer is transformed back to the time domain using an inverse Fourier transform, and all frequency-related features are extracted.

[0096] The decoder includes multiple cascaded second modules, which include: a transposed convolutional layer for upsampling to expand the time dimension and restore the resolution of the signal; and a third multi-scale temporal convolutional module that employs multiple parallel convolutional branches, with the kernel sizes of different convolutional branches used to capture waveform features at different time scales and then integrate them for feature fusion.

[0097] This application introduces a frequency-domain sensing processing module within the U-Net network. This module performs frequency-domain convolution and inverse transform on intermediate features, splicing real and imaginary parts, to explicitly compensate for phase deviations with learnable spectral corrections, avoiding the oversmoothing and phase distortion common in pure time-domain networks. Simultaneously, a multi-scale time-domain convolution module is employed to sense features at different time scales, and residual learning is used to focus on minute corrections, improving training stability. Combining these designs, this application can recover details and phase even under extremely low signal-to-noise ratio conditions, reducing parameter sensitivity and improving generalization, providing a more reliable preprocessing solution for practical microseismic denoising and subsequent inversion work.

[0098] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A low signal-to-noise ratio microseismic signal denoising method, characterized in that, The method includes: Acquire multiple seismic time-series signals; Multivariate singular spectrum analysis was used to reconstruct multichannel seismic time series signals to obtain reconstructed signals; include: Construct a Hankel matrix for each channel of the multi-channel seismic time series signal; Concatenate the Hankel matrices of all tracks column-wise to form a joint trajectory matrix; The joint trajectory matrix is ​​decomposed using singular value decomposition; The number of principal components to be retained is determined based on the distribution of the singular values. The top r largest singular values ​​and their corresponding singular vectors are retained, and the remaining singular values ​​are set to zero to obtain the denoised low-rank approximation matrix. The low-rank approximation matrix is ​​decomposed into sub-matrices corresponding to each signal, and the diagonal averaging method is applied to each sub-matrix to inversely map the matrix signal into a one-dimensional time series signal, thus obtaining the reconstructed signals of each channel. The reconstructed signal is concatenated with the original multichannel seismic time series signal and used as a dual-channel input dataset. The dual-channel input dataset is fed into the U-Net network for end-to-end recovery. The U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein: The encoder extracts multi-scale temporal features of the input signal through layer-by-layer downsampling operations. Each layer of downsampling operation gradually reduces the feature map resolution and increases the number of channels in the feature map. The bottleneck layer processes global semantic features from the encoder's output data while minimizing the dimension of the feature map, establishing dependencies over long time spans. The decoder recovers the temporal resolution from the output data of the bottleneck layer by upsampling layer by layer, and uses skip connections to fuse the features of the encoder and the features of the decoder. The encoder includes a multi-level cascaded first module, which sequentially includes: a first multi-scale temporal convolution module, a first frequency domain sensing processing module, and an output convolutional layer. The first multi-scale temporal convolution module employs multiple parallel convolution branches, with the kernel size of different convolution branches used to capture waveform features at different time scales; The first frequency domain sensing processing module performs a fast Fourier transform on the output signal of the first multi-scale time domain convolution module, transforming the signal from the time domain to the frequency domain, concatenating the real and imaginary parts in the frequency domain, and inputting them into a convolutional layer for processing; the output signal of the convolutional layer is transformed back to the time domain using an inverse Fourier transform, and frequency-related local features are extracted. The bottleneck layer includes a second multi-scale temporal convolution module and a second frequency domain sensing processing module. The second multi-scale temporal convolution module employs multiple parallel convolution branches. The kernel sizes of different convolution branches are used to capture waveform features at different time scales and then integrate them into global features. The second frequency domain sensing processing module performs a fast Fourier transform on the data integrated into global features, converting the signal from the time domain to the frequency domain. It then concatenates the real and imaginary parts in the frequency domain and inputs them into a convolutional layer for processing. The output signal of the convolutional layer is converted back to the time domain using an inverse Fourier transform, and all frequency-related features are extracted.

2. The method for denoising low signal-to-noise ratio microseismic signals according to claim 1, characterized in that, The decoder includes multiple cascaded second modules, each including: a transposed convolutional layer for upsampling to expand the time dimension and restore the signal resolution; and a third multi-scale temporal convolutional module that employs multiple parallel convolutional branches, with the kernel sizes of different convolutional branches used to capture waveform features at different time scales and then integrate them for feature fusion.

3. The method for denoising low signal-to-noise ratio microseismic signals according to claim 2, characterized in that, The upsampled feature map in the decoder is concatenated with the shallow feature map of the corresponding layer in the encoder through skip connections. The concatenated data is then input into the third multi-scale temporal convolution module for convolution processing to fuse deep semantic information and shallow detail information and repair the signal waveform.

4. The method for denoising low signal-to-noise ratio microseismic signals according to claim 3, characterized in that, The decoder includes an output convolutional layer that maps high-dimensional feature vectors back to the channel dimension of the original signal.

5. A low signal-to-noise ratio microseismic signal denoising system, used to implement the method described in any one of claims 1-4, characterized in that, include: The preprocessing module is used to reconstruct multichannel seismic time series signals using multivariate singular spectrum analysis to obtain reconstructed signals. The reconstructed signals are then concatenated with the original multichannel seismic time series signals to form a dual-channel input dataset. The U-Net network is used for end-to-end recovery of dual-channel input datasets.

Images