CN-122018005-A - Low signal-to-noise ratio micro-seismic signal denoising method and system

Abstract

The application belongs to the technical field of microseism signal processing, in particular to a low signal-to-noise ratio microseism signal denoising method and a system, wherein the method comprises the steps of collecting a plurality of earthquake time sequence signals; reconstructing the multi-channel seismic time sequence signals by adopting multivariate singular spectrum analysis to obtain reconstructed signals, splicing the reconstructed signals and the original multi-channel seismic time sequence signals to be used as a dual-channel input data set, and inputting the dual-channel input data set into a U-Net network for end-to-end recovery. The application obviously improves the recovery capability and phase reservation of weak seismic signals in an ultra-low signal-to-noise ratio environment.

Inventors

• HE RENJIE
• CHEN ZUBIN
• Zhang Qinxu

Assignees

• 吉林大学

Dates

Publication Date

20260512

Application Date

20260413

Claims (10)

1. 1. A method for denoising a microseismic signal with a low signal-to-noise ratio, the method comprising: collecting a plurality of earthquake time sequence signals; Reconstructing the multi-channel seismic time sequence signal by adopting multivariate singular spectrum analysis to obtain a reconstructed signal; Splicing the reconstructed signals with the original multi-channel seismic time sequence signals to serve as a dual-channel input data set; the dual channel input dataset is input to the U-Net network for end-to-end recovery.
2. 2. The method of claim 1, wherein reconstructing the multi-channel seismic timing signal using multivariate singular spectrum analysis comprises: constructing a Hanker matrix for each of the plurality of seismic timing signals; Splicing the Hank matrix of all channels into a joint track matrix according to columns; decomposing the joint track matrix by using singular value decomposition; Determining the number of principal components to be reserved according to the size distribution of the singular values, reserving the singular values with the larger r before and the singular vectors corresponding to the singular values, and setting the rest singular values to zero to obtain a denoised low-rank approximation matrix; And disassembling the low-rank approximation matrix into sub-matrixes corresponding to each channel of signals, applying a diagonal average method to each sub-matrix, and inversely mapping the signals in a matrix form into one-dimensional time sequence signals to obtain each channel of reconstructed signals.
3. 3. A low signal-to-noise ratio microseismic signal denoising method according to claim 1, The U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein: The coder extracts multi-scale time domain features of the input signal through layer-by-layer downsampling operation, and the downsampling operation of each layer gradually reduces the resolution of the feature map and increases the channel number of the feature map; The bottleneck layer is used for processing global semantic features when the dimension of the feature map is minimized for the output data of the encoder, and establishing a long-span dependency relationship; and the decoder restores the time resolution of the output data of the bottle neck layer through up-sampling layer by layer, and fuses the characteristics of the encoder and the characteristics of the decoder by using jump connection.
4. 4. The method for denoising a low signal-to-noise ratio microseism signal according to claim 3, wherein the encoder comprises a first module in multistage cascade connection, and the first module comprises a first multi-scale time domain convolution module, a first frequency domain perception processing module and an output convolution layer in sequence; the first multi-scale time domain convolution module adopts a plurality of parallel convolution branches, and convolution kernel sizes of different convolution branches are used for capturing waveform characteristics of different time scales; The first frequency domain sensing processing module performs fast Fourier transform on the output signal of the first multi-scale time domain convolution module, converts the signal from the time domain to the frequency domain, splices the real part and the imaginary part in the frequency domain together, inputs the spliced real part and the spliced imaginary part into one convolution layer for processing, and converts the output signal of the convolution layer back to the time domain by using inverse Fourier transform to extract local characteristics related to the frequency.
5. 5. A method for denoising a low signal-to-noise ratio microseism signal according to claim 3, wherein the bottleneck layer comprises a second multi-scale time domain convolution module and a second frequency domain perception processing module, the second multi-scale time domain convolution module adopts a plurality of parallel convolution branches, and convolution kernel sizes of different convolution branches are used for capturing waveform features of different time scales and then are integrated into global features; the second frequency domain sensing processing module performs fast Fourier transform on the data integrated into global features, converts signals from a time domain to a frequency domain, splices real parts and imaginary parts in the frequency domain together, inputs the signals into a convolution layer for processing, and converts output signals of the convolution layer back to the time domain by using inverse Fourier transform to extract all features related to the frequency.
6. 6. The method for denoising a microseismic signal with a low signal-to-noise ratio according to claim 3, wherein the decoder comprises a plurality of cascaded second modules, the second modules comprise a transposed convolution layer for upsampling, expanding the time dimension and recovering the resolution of the signal, and a third multi-scale time domain convolution module adopts a plurality of parallel convolution branches, and convolution kernel sizes of different convolution branches are used for capturing waveform features with different time scales and then are integrated into feature fusion.
7. 7. The method for denoising a microseism signal with a low signal to noise ratio according to claim 6, wherein the feature map sampled in the decoder is spliced with the shallow feature map of the corresponding level in the encoder through jump connection, and the spliced data is input into a third multi-scale time domain convolution module for convolution processing so as to fuse deep semantic information and shallow detail information and restore signal waveforms.
8. 8. A method of denoising a low signal-to-noise ratio microseismic signal according to claim 6, wherein the decoder comprises an output convolution layer mapping high-dimensional feature vectors back to the channel dimensions of the original signal.
9. 9. A low signal-to-noise ratio microseismic signal denoising system, comprising: The preprocessing module is used for reconstructing the multi-channel seismic time sequence signals by adopting multivariate singular spectrum analysis to obtain reconstructed signals, and splicing the reconstructed signals with the original multi-channel seismic time sequence signals to be used as a dual-channel input data set; and the U-Net network is used for carrying out end-to-end recovery on the two-channel input data set.
10. 10. The low signal-to-noise ratio microseismic signal denoising system according to claim 9, wherein the U-Net network comprises an encoder, a bottleneck layer, and a decoder, wherein: The coder extracts multi-scale time domain features of the input signal through layer-by-layer downsampling operation, and the downsampling operation of each layer gradually reduces the resolution of the feature map and increases the channel number of the feature map; The bottleneck layer is used for processing global semantic features when the dimension of the feature map is minimized for the output data of the encoder, and establishing a long-span dependency relationship; and the decoder restores the time resolution of the output data of the bottle neck layer through up-sampling layer by layer, and fuses the characteristics of the encoder and the characteristics of the decoder by using jump connection.

Description

Low signal-to-noise ratio micro-seismic signal denoising method and system Technical Field The application belongs to the technical field of microseism signal processing, and particularly relates to a low signal-to-noise ratio microseism signal denoising method and system. Background The microseismic monitoring is often faced with weak seismic waves with extremely low signal to noise ratio, and the current microseismic and seismic denoising technology has obvious defects under the condition of extremely low signal to noise ratio. Although the traditional filtering and time-frequency transformation can inhibit noise, the phase and weak waveform details are often lost, the inversion precision of a seismic source is reduced, MSSA (multiple singular spectrum analysis) is highly sensitive to the length of an embedded window and the cut-off rank, the reconstruction capability is reduced under a non-stable or strong noise environment, a pure time domain depth network is easy to generate excessive smoothness and distortion phase, the physical priori is ignored, the generalization is poor, and the single frequency domain method is difficult to fully utilize the correlation among multiple channels. In summary, in a very low signal-to-noise ratio scene, both phase preservation and detail restoration are required, robustness and calculation efficiency are simultaneously considered, and the existing method is difficult to consider, so that a new denoising scheme combining physical priori and frequency domain sensing is required. Disclosure of Invention The embodiment of the application provides a low signal-to-noise ratio microseism signal denoising method and system, which solve the denoising and phase reservation problems in the fields of microseism and earthquake under the ultra-low signal-to-noise ratio. An embodiment of a first aspect of the present application provides a low signal-to-noise ratio micro-seismic signal denoising method, which includes: collecting a plurality of earthquake time sequence signals; Reconstructing the multi-channel seismic time sequence signal by adopting multivariate singular spectrum analysis to obtain a reconstructed signal; Splicing the reconstructed signals with the original multi-channel seismic time sequence signals to serve as a dual-channel input data set; the dual channel input dataset is input to the U-Net network for end-to-end recovery. Further, the reconstructing the multi-channel seismic timing signal using multivariate singular spectrum analysis comprises: constructing a Hanker matrix for each of the plurality of seismic timing signals; Splicing the Hank matrix of all channels into a joint track matrix according to columns; decomposing the joint track matrix by using singular value decomposition; Determining the number of principal components to be reserved according to the size distribution of the singular values, reserving the singular values with the larger r before and the singular vectors corresponding to the singular values, and setting the rest singular values to zero to obtain a denoised low-rank approximation matrix; And disassembling the low-rank approximation matrix into sub-matrixes corresponding to each channel of signals, applying a diagonal average method to each sub-matrix, and inversely mapping the signals in a matrix form into one-dimensional time sequence signals to obtain each channel of reconstructed signals. Further, the U-Net network includes an encoder, a bottleneck layer, and a decoder, wherein: The coder extracts multi-scale time domain features of the input signal through layer-by-layer downsampling operation, and the downsampling operation of each layer gradually reduces the resolution of the feature map and increases the channel number of the feature map; The bottleneck layer is used for processing global semantic features when the dimension of the feature map is minimized for the output data of the encoder, and establishing a long-span dependency relationship; and the decoder restores the time resolution of the output data of the bottle neck layer through up-sampling layer by layer, and fuses the characteristics of the encoder and the characteristics of the decoder by using jump connection. Further, the encoder comprises a first module of multistage cascade connection, wherein the first module sequentially comprises a first multi-scale time domain convolution module, a first frequency domain perception processing module and an output convolution layer; the first multi-scale time domain convolution module adopts a plurality of parallel convolution branches, and convolution kernel sizes of different convolution branches are used for capturing waveform characteristics of different time scales; The first frequency domain sensing processing module performs fast Fourier transform on the output signal of the first multi-scale time domain convolution module, converts the signal from the time domain to the frequency domain, splices the real part and t