CN-121999794-A - Unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation

CN121999794ACN 121999794 ACN121999794 ACN 121999794ACN-121999794-A

Abstract

The invention relates to the technical field of signal processing, in particular to an unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation. And constructing a cooperative attention network comprising time and frequency dual-attention paths, respectively learning the time dependence and frequency structural relation of signals, and generating a corresponding attention weight matrix. And performing cross-domain interaction fusion on the two weight matrixes to generate a time-frequency cooperative attention mask. And selectively enhancing and suppressing the time-frequency domain features by using the mask to obtain the time-frequency domain features after noise reduction, and finally reconstructing the noise reduction signals through inverse transformation. According to the method, collaborative attention modeling of a signal time-frequency structure is realized through double-channel independent learning and cross-domain fusion, and the accuracy and the robustness of signal noise reduction under a complex noise background are improved.

Inventors

DU SHUAIQUN
XIN JUNSHENG
ZHANG HENG

Assignees

中国雅江集团有限公司
中国科学院青藏高原研究所

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (10)

1. An unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation is characterized by comprising the following steps: Acquiring an original DAS sensing signal sequence, and performing segmentation windowing pretreatment on the original DAS sensing signal sequence to generate a signal frame set; Performing time-frequency joint transformation on the signal frame set to generate a time-frequency domain feature set containing a complex spectrum matrix; Constructing a cooperative attention network comprising a temporal attention path for capturing a time dependency relationship between signal frames and a frequency attention path for capturing a frequency structural relationship between spectral components; Respectively inputting the time-frequency domain feature set into a time attention path and a frequency attention path of the collaborative attention network to generate a time attention weight matrix and a frequency attention weight matrix; Performing cross-domain interaction fusion on the time attention weight matrix and the frequency attention weight matrix to generate a time-frequency cooperative attention mask; selectively enhancing and suppressing the time-frequency domain feature set by using the time-frequency cooperative attention mask to generate a noise-reduced time-frequency domain feature set; and performing inverse time-frequency transformation on the denoised time-frequency domain feature set, and reconstructing to generate a denoised DAS signal sequence.
2. The method for noise reduction of an unsupervised DAS signal based on time-frequency domain attention synergy according to claim 1, wherein said performing a time-frequency joint transform on the set of signal frames to generate a set of time-frequency domain features including a complex spectrum matrix comprises: applying an analysis window function to each signal frame in the signal frame set to obtain a windowed signal frame; Performing a multi-resolution spectral analysis on each windowed signal frame to generate spectral components of a plurality of scales; arranging spectral components of each scale into a matrix in a complex form to form a complex spectral matrix; And organizing the complex spectrum matrixes corresponding to all the signal frames in time sequence to form the time-frequency domain feature set.
3. The method for unsupervised DAS signal noise reduction based on time-frequency domain attention co-ordination as claimed in claim 1, wherein said constructing a co-attention network comprising a time attention path and a frequency attention path comprises: Designing a structure of a time attention path, wherein the time attention path is formed by stacking a plurality of time self-attention layers, and each time self-attention layer is used for calculating the correlation of the same frequency component on different time frames; designing a structure of a frequency attention path, wherein the frequency attention path is formed by stacking a plurality of frequency self-attention layers, and each frequency self-attention layer is used for calculating the correlation between different frequency components in the same time frame; An information exchange bridge between the time attention path and the frequency attention path is established, and the information exchange bridge allows the time attention weight matrix and the frequency attention weight matrix to be mutually referenced and adjusted in the generation process.
4. The method for noise reduction of an unsupervised DAS signal based on time-frequency domain attention coordination according to claim 1, wherein the step of inputting the time-frequency domain feature set into a time attention path and a frequency attention path of the coordinated attention network, respectively, generates a time attention weight matrix and a frequency attention weight matrix, comprises: Slicing the time-frequency domain feature set along a time dimension, generating a feature vector sequence indexed according to frequency, and inputting the feature vector sequence into the time attention path; The time attention path calculates the similarity of any two time point feature vectors in the feature vector sequence, and distributes attention weights based on the similarity to form a preliminary time attention weight matrix; slicing the time-frequency domain feature set along a frequency dimension, generating a feature vector sequence indexed by time, and inputting the feature vector sequence into the frequency attention path; The frequency attention path calculates the similarity of any two frequency point feature vectors in the feature vector sequence indexed according to time, and distributes attention weights based on the similarity to form a preliminary frequency attention weight matrix; And carrying out cooperative correction on the preliminary time attention weight matrix and the preliminary frequency attention weight matrix through the information exchange bridge, and outputting the final time attention weight matrix and the final frequency attention weight matrix.
5. The method for noise reduction of an unsupervised DAS signal based on time-frequency domain attention coordination according to claim 1, wherein the cross-domain interaction fusion of the time attention weight matrix and the frequency attention weight matrix is performed to generate a time-frequency coordination attention mask, which comprises: performing tensor outer product operation on the time attention weight matrix and the frequency attention weight matrix to generate a three-dimensional time-frequency joint attention body; Respectively compressing and polymerizing the time-frequency joint attention body along the time dimension and the frequency dimension to generate a two-dimensional joint attention mapping diagram; And performing nonlinear activation and normalization processing on the joint attention map, mapping the weight value into a preset range, and generating the time-frequency cooperative attention mask.
6. The method for noise reduction of an unsupervised DAS signal based on time-frequency domain attention co-ordination of claim 1, wherein the selectively enhancing and suppressing the time-frequency domain feature set with the time-frequency co-attention mask to generate a noise-reduced time-frequency domain feature set comprises: Performing element level point multiplication on the time-frequency cooperative attention mask and a complex spectrum matrix in the time-frequency domain feature set, wherein a weight value in the time-frequency cooperative attention mask is used for scaling a complex value of a corresponding time-frequency unit; Global energy normalization is carried out on the complex spectrum matrix after the dot multiplication so as to maintain the relative stability of the total energy of the signals; Reorganizing all the processed complex spectrum matrixes to form the time-frequency domain feature set after noise reduction.
7. The method for unsupervised DAS signal noise reduction based on time-frequency domain attention synergy according to claim 1, wherein said performing inverse time-frequency transform on the denoised time-frequency domain feature set, reconstructing to generate denoised DAS signal sequences, comprises: Extracting a complex spectrum matrix after noise reduction corresponding to each signal frame from the time-frequency domain feature set after noise reduction; Performing inverse spectrum analysis on each denoised complex spectrum matrix, and converting the inverse spectrum analysis back to time domain signal components of corresponding scales; Overlapping and synthesizing the time domain signal components of all scales to generate a noise-reduced time domain signal of each signal frame; And performing overlap-add operation on the denoised time domain signals of all the signal frames, eliminating boundary effects introduced by windowing, and finally splicing and reconstructing the complete denoised DAS signal sequence.
8. The time-frequency domain attention co-based unsupervised DAS signal denoising method according to claim 1, further comprising an iterative optimization step performed after generating the denoised DAS signal sequence: Calculating the difference measurement of the DAS signal sequence after noise reduction and the original DAS sensing signal sequence in a preset feature space; Constructing an unsupervised loss function based on the difference metric; Back-propagating to the collaborative attention network by using gradient information of the loss function, and updating parameters of the collaborative attention network; And repeatedly executing the step of converting from time-frequency joint to signal reconstruction by using the cooperative attention network after updating the parameters, and performing a plurality of iterations until a preset stopping criterion is met.
9. The method for unsupervised DAS signal denoising based on time-frequency domain attention synergy according to claim 8, wherein the calculating a difference metric between the denoised DAS signal sequence and the original DAS sensing signal sequence in a preset feature space comprises: Converting the DAS signal sequence after noise reduction and the original DAS sensing signal sequence into the same time-frequency analysis domain respectively; in the time-frequency analysis domain, calculating the amplitude spectrum difference and the phase spectrum difference of the time-frequency units corresponding to the two signal sequences; And carrying out weighted summation on the amplitude spectrum difference and the phase spectrum difference, and carrying out standardization by combining the overall energy distribution of the signals to obtain the difference measurement.
10. The method for unsupervised DAS signal noise reduction based on time-frequency domain attention coordination according to claim 8, wherein the using gradient information of the loss function, back-propagating to the coordinated attention network, updating parameters of the coordinated attention network, comprises: Calculating the gradient of the time-frequency cooperative attention mask according to the loss function; transmitting the gradient to the time attention weight matrix and the frequency attention weight matrix through inverse time-frequency transformation and an inverse process of selective enhancement and suppression operation; Continuing to counter-propagate the gradient along the time attention path and the frequency attention path, and calculating gradients of parameters inside the time attention path and the frequency attention path respectively; And updating all trainable parameters in the collaborative attention network according to the calculated gradient by adopting a gradient descent algorithm.

Description

Unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation Technical Field The invention relates to the technical field of signal processing, in particular to an unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation. Background The existing noise reduction method for the distributed optical fiber acoustic wave sensing signals mostly adopts frequency domain filtering based on a fixed filter, time domain filtering based on a statistical model or a depth network utilizing a single attention mechanism. These solutions usually perform independent processing in the time domain or the frequency domain, or use a simple feature stitching method to introduce time-frequency information. The main disadvantage is that the inherent correlation and interaction effect of the signal between the time domain and the frequency domain cannot be fully considered. Under the background of complex noise, the effective components of the signals and the noise are often aliased in the time domain and the frequency domain, and the independent processing of the time domain or the frequency domain information easily causes the distortion of the effective signals or the residual of the noise, so that the improvement of the noise reduction performance is limited. Existing attention-based noise reduction methods typically employ a single-dimensional attention mechanism or simple time-frequency feature fusion at shallow layers. Such methods have difficulty in accurately modeling the timing dependency of the long range of the signal and the fine frequency structure relationship at the same time. A single attention weight map cannot distinguish between different importance distribution patterns of signals in two dimensions of time and frequency, while simple feature fusion cannot realize depth interaction and collaborative modulation between time-domain and frequency-domain attention information. This results in a limited ability of the generated attention mask to distinguish between signals and noise in the time spectrum, and it is difficult to achieve accurate enhancement and suppression of complex time-frequency structures. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides an unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation. In order to achieve the purpose, the invention adopts the following technical scheme that the unsupervised DAS signal noise reduction method based on time-frequency domain attention cooperation comprises the following steps: Acquiring an original DAS sensing signal sequence, and performing segmentation windowing pretreatment on the original DAS sensing signal sequence to generate a signal frame set; Performing time-frequency joint transformation on the signal frame set to generate a time-frequency domain feature set containing a complex spectrum matrix; Constructing a cooperative attention network comprising a temporal attention path for capturing a time dependency relationship between signal frames and a frequency attention path for capturing a frequency structural relationship between spectral components; Respectively inputting the time-frequency domain feature set into a time attention path and a frequency attention path of the collaborative attention network to generate a time attention weight matrix and a frequency attention weight matrix; Performing cross-domain interaction fusion on the time attention weight matrix and the frequency attention weight matrix to generate a time-frequency cooperative attention mask; selectively enhancing and suppressing the time-frequency domain feature set by using the time-frequency cooperative attention mask to generate a noise-reduced time-frequency domain feature set; and performing inverse time-frequency transformation on the denoised time-frequency domain feature set, and reconstructing to generate a denoised DAS signal sequence. As a further aspect of the present invention, the performing time-frequency joint transformation on the signal frame set to generate a time-frequency domain feature set including a complex spectrum matrix includes: applying an analysis window function to each signal frame in the signal frame set to obtain a windowed signal frame; Performing a multi-resolution spectral analysis on each windowed signal frame to generate spectral components of a plurality of scales; arranging spectral components of each scale into a matrix in a complex form to form a complex spectral matrix; And organizing the complex spectrum matrixes corresponding to all the signal frames in time sequence to form the time-frequency domain feature set. As a further aspect of the present invention, the constructing a cooperative attention network including a time attention path and a frequency attention path includes: Designing a structure of a time attention path, wherein the time attention path is formed b