A method for denoising distributed acoustic wave sensing data based on high-quality sparse priors

Abstract

The present invention relates to the field of distributed fiber optic acoustic wave sensing technology and discloses a method for denoising distributed acoustic wave sensing data based on high-quality sparse priors. Multiple spatial points are selected along the fiber optic laying path, and seismic detectors are deployed at each spatial point. By incorporating a small amount of high-quality signals from traditional seismic detectors as input into the deep learning algorithm modeling process, the present invention can further improve the model's denoising performance based on existing intelligent seismic data denoising algorithms. This method can effectively improve the signal quality of seismic data collected by DAS and is applicable to multiple application scenarios requiring DAS monitoring, such as building structure health assessment, geological disaster warning, and underground structure exploration.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed optical fiber acoustic wave sensing, and in particular to a distributed acoustic wave sensing data denoising method based on high-quality sparse priors. Background Art

[0002] Distributed acoustic sensing (DAS) is a novel sensing technology. Because micro-deformations in an optical fiber cause phase shifts in scattered light at specific locations, a demodulator can be used to collect temperature, strain, and other sensor data from different locations on the fiber (Hartog, 2017). The collected strain data captures the characteristics of the seismic wavefield in the region where the fiber is located. Furthermore, because DAS can achieve denser spatial sampling of seismic data at a lower cost (Madsen et al., 2013), it has been widely used in seismic data acquisition (Fernández-Ruiz et al., 2018). However, due to complex coupling between the fiber and the ground, unstable light source processing by the demodulator, and human activity in the deployment environment, seismic data collected using fiber optic sensors are subject to more severe noise contamination than traditional geophones (Feng and Li, 2019).

[0003] Because DAS data has high spatial resolution but is susceptible to noise, noise suppression for seismic data has become an important research topic. The diverse causes of DAS noise lead to local structural correlations, which poses certain challenges to noise suppression (Feng and Li, 2021). In recent years, with the rise of artificial intelligence technology, noise suppression algorithms based on deep neural networks have demonstrated superior denoising effects compared to traditional seismic data noise suppression methods such as the FX transform (Harris and White, 1997) and empirical mode decomposition (Liu et al., 2013) (Zhao et al., 2018). Denoising algorithms based on deep learning are mainly divided into two categories: supervised learning and unsupervised learning. The main difference between these two methods lies in whether noise-free seismic data is provided as labels for the neural network to learn. However, since noise-free seismic data is unknown, supervised learning methods that require labels usually use near-real seismic data simulated based on physical laws as labels (Zhao et al., 2020; Ma et al., 2023). This makes such methods often unstable when applied to actual seismic data denoising (Zhao et al., 2023).

[0004] Noise suppression algorithms based on self-supervised learning are primarily based on mathematical modeling, which relies on the fact that the effective signal at the current pixel can be predicted by the signals at neighboring pixels, while the noise at the current pixel cannot be predicted by the noise at neighboring pixels. These algorithms are widely used in the field of image noise suppression (Lehtinen et al., 2018; Batson and Royer, 2019). By replacing the concepts of the current pixel and neighboring pixels in the image domain with the waveform sampled at the current spatial location and the waveform sampled at neighboring spatial locations in seismic data acquisition during the modeling process, these noise suppression algorithms based on self-supervised learning can be applied to seismic data acquired by DAS, achieving excellent noise suppression results (van den Ende et al., 2021).

[0005] In order to further improve the denoising performance of the noise suppression algorithm based on self-supervised learning, the present invention provides a distributed acoustic sensor data denoising method based on high-quality sparse prior. By leveraging the advantage of the high signal-to-noise ratio quality of traditional detectors, the waveforms recorded at certain spatial points of the DAS are sparsely constrained through the signals recorded by traditional detectors, thereby further improving the noise suppression effect. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a distributed acoustic wave sensor data denoising method based on high-quality sparse prior. By adding a small amount of high-quality signals of traditional seismic detectors as input during the deep learning algorithm modeling process, the denoising performance of the model can be further improved on the basis of the existing intelligent denoising algorithm for seismic data.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] A distributed acoustic sensor data denoising method based on high-quality sparse priors uses waveforms received by geophones to guide a deep neural network to denoise seismic data based on distributed fiber-optic acoustic sensor signals. Specifically, the method includes:

[0009] Select multiple spatial points along the optical fiber laying path and deploy seismic detectors at each spatial point;

[0010] Based on the J-invariance denoising theory, we define a function g with J-invariance. θ :

[0011] g θ (·)=∑ k∈K′ ∏ k (f θ (Π kc (·)));

[0012] Among them, K′ represents a set of multiple spatial points selected on the optical fiber laying path, k∈K′ represents any spatial point in K′, Π k Indicates the operation of setting all dimensions except the dimension represented by k to zero, f θ represents a deep neural network; represents the spatial points near the spatial point k, where N is the total number of spatial points selected on the optical fiber laying path, and s is the interval between the spatial points;

[0013] The parameters θ of the deep neural network are optimized in the following way to achieve denoising of seismic data:

[0014] argmin θ ‖g θ (x)-x k ‖ 2 ;

[0015] where x k Represents the waveform received by the seismic detector at spatial point k.

[0016] Furthermore, the high-quality waveforms received by the seismic detectors are introduced as prior knowledge into a deep neural network to guide the deep neural network to denoise the seismic data based on distributed fiber optic acoustic sensor signals. The deep neural network adopts a U-Net network.

[0017] Compared with the prior art, the beneficial technical effects of the present invention are:

[0018] The distributed acoustic sensor data denoising method based on high-quality sparse prior proposed in this invention combines the high signal-to-noise ratio of data collected by traditional seismic detectors and the high spatial sampling rate of data collected by DAS, and obtains low-noise seismic data with a high spatial sampling rate through an intelligent denoising algorithm based on deep learning. Figure 4 This method demonstrates that, compared to existing self-supervised learning-based approaches, it can further suppress coherent noise. The proposed method can effectively improve the signal quality of DAS-collected seismic data and is applicable to a variety of applications requiring DAS monitoring, such as building structural health assessment, geological disaster warning, and underground structure exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the implementation of the noise reduction method of the present invention. Figure 1 The black dots in the figure represent equivalent spatial locations during DAS data acquisition. The inverted triangles represent traditional nodal geophones, which overlap with certain points acquired by fiber optics. The black waveform represents the low signal-to-noise ratio, densely sampled signal acquired by DAS, while the red waveform represents the signal acquired by traditional geophones.

[0020] Figure 2 The figure compares the principles of the method of the present invention and existing algorithms. The upper part of the figure is a schematic diagram of the existing unsupervised distributed fiber optic acoustic sensor data noise suppression algorithm, which uses the characteristics that the waveform to be denoised can be restored by the waveform sampled in the nearby space, while the noise cannot be predicted by the surrounding spatial waveform. The lower part is the noise reduction method proposed by the present invention. By converting the target predicted by the deep neural network from noisy distributed fiber optic acoustic sensor data into high signal-to-noise ratio data received by traditional seismic detectors, noise with spatial correlation can be further suppressed.

[0021] Figure 3 Schematic diagram of a distributed fiber optic acoustic wave sensing data noise suppression network model based on a U-Net structure in an embodiment of the present invention.

[0022] Figure 4 The three sub-figures from left to right are the noisy data, the denoising results of the existing unsupervised learning denoising algorithm, and the denoising results of the denoising method of the present invention.

[0023] Figure 5 The three sub-figures from left to right are respectively the synthesized seismic waveform, the noise extracted from the real acquired data, and the noisy data obtained by adding the two.

[0024] Figure 6 Schematic diagram of bandpass filtering; where a is the mean of all seismic data and b is the mean of the noise spectrum.

[0025] Figure 7 This is a comparison chart of noise suppression effects on different networks. From top to bottom, each row represents the effect comparison when the signal-to-noise ratio is -10dB, 0dB, and 10dB, respectively. From left to right, each column represents the noisy data, the denoising results of the existing unsupervised learning denoising algorithm, and the denoising results of the algorithm proposed in this paper.

[0026] Figure 8 This figure shows a comparison of parameter statistics, namely, the mean square error (MSE), signal-to-noise ratio (SNR), and correlation coefficient (PCC) for data denoising at different signal-to-noise ratios. The black line represents the data before denoising, while the green and red lines represent the denoising effects of existing unsupervised learning denoising algorithms and the proposed algorithm, respectively. DETAILED DESCRIPTION

[0027] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] The distributed acoustic wave sensor data denoising method based on high-quality sparse prior proposed in this invention can constrain the denoising results output by the deep neural network by introducing traditional seismic detector signals with a high signal-to-noise ratio into the self-supervised deep neural network training process, thereby further improving the noise suppression performance on the existing self-supervised learning method.

[0029] The core idea of ​​the present invention is to improve the denoising effect of distributed fiber optic acoustic wave sensing data by monitoring waveforms from a small number of traditional node-type seismic detectors with a high signal-to-noise ratio. Figure 1 This is a schematic diagram of the method, which requires the deployment of a small number of sparsely distributed traditional node-type seismic detectors at the location of the optical fiber when laying out the instruments. The waveforms collected by these traditional detectors are of higher quality than those of the DAS waveforms, but each detector can only receive data from a single spatial point. It is very costly to deploy traditional detectors at all spatial points collected by the optical fiber. Therefore, traditional detectors only collect data at some points and have spatial sparsity. The present invention adds these high-quality sparse data as prior information to the training process of the denoising neural network, thereby improving the noise suppression effect of the neural network. Below we will introduce the technical content of the present invention in detail from three aspects: method principle, model training and denoising evaluation indicators.

[0030] 1. Principle of the Method

[0031] The theoretical basis of this invention is the J-invariance denoising theory proposed by Batson and Royer (2019), which is defined as:

[0032] definition are noisy data samples and non-noisy data samples respectively, assuming is a set of dimensions {1,…,m}, and dimension J is A subset of If g(x) is represented by g(x) in J dimension J The value of does not depend on x J If true, then define function g: It has J invariance.

[0033] Assume J c Is J relative to The complement of , for a function g with J invariance, if the noise of y in dimension J and the noise of dimension J c The noises in are independent of each other, then:

[0034]

[0035] The above formula shows that when a function g with J invariance is used to fit the representation x of the noisy sample in dimension J, J When , it is equivalent to removing x JThe remaining dimensions cannot be sampled The predicted noise.

[0036] Based on this, Van den Ende et al. (2021) applied this theory to the noise suppression of seismic data based on distributed fiber optic acoustic sensing signals, and defined a function g with J invariance. θ :

[0037]

[0038] Where K is the set of all spatial points collected by DAS, k∈K represents any spatial point collected by DAS, Π k Indicates the operation of setting all dimensions except the dimension represented by k to zero, f θ Represents a deep neural network. Represents the spatial points near the spatial point k. Where N is the number of spatial points taken (usually an odd number), and s is the interval between the spatial points taken. The following optimizes the parameters θ of the deep neural network to achieve seismic data denoising:

[0039] argmin θ ‖g θ (x)-x‖ 2 (3)

[0040] The present invention further innovates on the basis of the above unsupervised deep learning denoising theory. The denoising process of the above theory can be Figure 2 As summarized in the upper part of , the gray waveform is the waveform that needs to be denoised. The deep neural network attempts to predict the current waveform by sampling waveforms near the waveform. Since the seismic signal can be predicted by these waveforms, while the noise is unpredictable, the output of the deep neural network is theoretically only the signal part. However, not all noise is unpredictable. Some noise shows spatial correlation and can be predicted by the surrounding waveforms. Therefore, the present invention further guides the denoising process of the deep neural network through the signal received by the traditional seismic detector. Define g θ for:

[0041]

[0042] The present invention replaces all the spatial points collected by DAS in formula (2) with the set of spatial points arranged by traditional seismic detectors. In this case, g θ It still has J invariance, and the optimization objective becomes:

[0043] argmin θ ‖g θ (x)-x k ‖ 2 ;(5)

[0044] where x k Represents the waveform received by the seismic detector at spatial point k.

[0045] The noise reduction method of the present invention is shown as follows: Figure 2 The deep neural network no longer predicts all noisy data, but only the waveform received by the traditional seismometer. Because the waveform received by the traditional seismometer lacks noise components, it can avoid the interference of spatially coherent noise on the denoising result.

[0046] 2. Model Training

[0047] According to the above introduction to the principle of the method, the present invention does not specify a specific deep neural network structure. Theoretically, all Figure 2 Any deep neural network structure that meets the input and output data shape requirements can be adopted in the present invention. In a preferred embodiment, the present invention adopts the U-Net network structure (Ronneberger et al., 2015), and the specific structure is as follows Figure 3 As shown in the figure, this structure can extract multi-scale features from the input data, effectively convey high-frequency information, improve convergence during model training, and avoid gradient explosion and vanishing problems. The Adam algorithm is used for U-Net network optimization, with an initial learning rate of 0.001 and 100 training cycles, with the learning rate reduced by half every 20 cycles.

[0048] 3. Denoising evaluation indicators

[0049] In order to compare the denoising effect of the method proposed in this invention and the existing deep learning self-supervised algorithm, three comparison indicators are introduced: Mean Square Error (MSE), Signal-to-Noise Ratio (SNR) and Pearson Correlation Coefficient (PCC). Mean Square Error (MSE) is a commonly used indicator to measure the difference between the predicted value and the actual value. It is the average value of the residual (square of the difference) between the actual value and the model predicted value. Assuming that the total number of sampling points of DAS acquisition data is N, the noisy seismic data and the noise-free seismic data are represented by K and I respectively, the mean square error MSE is defined as:

[0050]

[0051] The signal-to-noise ratio (SNR) is a measure of signal quality, describing the ratio of signal strength to background noise. A higher SNR indicates a greater ratio of useful information to noise, and therefore better signal quality. The SNR is typically expressed in decibels (dB). The formula for calculating SNR is as follows:

[0052]

[0053] Among them, P signal and P noise is the power of the signal and noise. For a discrete digital signal x(n), assuming its length is N, its power P can be expressed as:

[0054]

[0055] The Pearson correlation coefficient (PCC) is an indicator that measures the similarity between two signals. It is a value between -1 and 1 that is used to measure the strength of the linear association between two continuous variables. The larger the absolute value, the stronger the linear relationship. The calculation formula of PCC is:

[0056]

[0057] in are the mean of the noisy seismic data and the mean of the noise-free seismic data, respectively.

[0058] The distributed acoustic sensor data denoising method proposed in this paper, based on high-quality sparse priors, is applied to noise suppression in actual DAS recordings. Due to the complex DAS monitoring environment and the inherent noise of the instruments, acquiring noise-free seismic data is difficult in actual data acquisition. However, the evaluation metrics for noise suppression algorithms, such as mean square error, signal-to-noise ratio, and correlation coefficient, all require noise-free seismic data. Therefore, the present invention uses physical simulation to obtain noise-free seismic data and then adds this data to the actual noisy data collected by the DAS to obtain the noisy seismic data to be denoised. Figure 5 This is an example of data set synthesis. The leftmost sub-figure is the seismic waveform obtained by physical simulation through the spectral element method, the middle sub-figure is the noise in the real DAS record, and the right one is the noisy seismic data of the sum of the two, which is the target to be denoised.

[0059] Before the noise reduction method proposed in the present invention is applied to noise suppression, the input data is preprocessed in an appropriate manner, including band-pass filtering, resampling and data normalization. Figure 6 The frequency range of the seismic data and noise is displayed. A filtering range of 5-40 Hz was selected based on the dominant frequency range of the seismic data. The data was then resampled to 200 Hz and normalized by subtracting the mean and dividing by the standard deviation. This results in waveform data with a mean of 0 and a standard deviation of 1 as input to the network. This operation simplifies the data features and increases the network's convergence speed.

[0060] The noise suppression algorithm for distributed fiber-optic acoustic sensing data based on unsupervised learning, proposed by van den Ende et al. (2021), was selected as a comparison target for the noise reduction method proposed in this paper. To ensure that the unsupervised denoising algorithm compared with the noise reduction method proposed in this paper achieved the best performance, we conducted comparative tests on different parameters of the algorithm (the number of waveforms N input to the deep neural network and the interval s between waveforms), which are the main influencing parameters of the algorithm. The comparison results are shown in Table 1. Based on the comparison results, the best-performing parameters (N = 11, s = 3) were selected for comparison with the noise reduction method proposed in this paper.

[0061] Table 1

[0062]

[0063] Figure 7 This figure shows a comparison of noise suppression effects at different signal-to-noise ratios. The three rows in the figure represent denoising experiments at different signal-to-noise ratios (-10dB, 0dB, and 10dB). The leftmost column shows the noisy data, the middle column shows the noise suppression results of existing unsupervised learning, and the right column shows the noise suppression results of the noise reduction method proposed in this paper. The results in the figure clearly show that at different signal-to-noise ratios, the noise reduction method proposed in this paper can further suppress noise and improve signal quality. Figure 8 The noise suppression results were quantitatively analyzed using three statistical indicators: mean square error, signal-to-noise ratio, and correlation coefficient. The results showed that the method proposed in the present invention was superior in different signal-to-noise ratios and different evaluation indicators.

[0064] A comparative analysis of application results demonstrates that the proposed distributed acoustic sensor data denoising method based on high-quality sparse priors can effectively improve the noise suppression of distributed fiber-optic acoustic sensor data. Given that DAS is widely used in diverse applications such as traffic monitoring, resource exploration, and underground structure exploration, the proposed denoising method has significant reference value for practical engineering applications of DAS.

[0065] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0066] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for denoising distributed acoustic wave sensor data based on high-quality sparse priors, characterized in that: The waveforms received by the seismic detectors are used to guide the deep neural network to denoise the seismic data based on distributed fiber optic acoustic sensor signals. Specifically, Select multiple spatial points along the optical fiber laying path and deploy seismic detectors at each spatial point; Based on the J-invariance denoising theory, we define a function g with J-invariance. θ : Among them, K′ represents a set of multiple spatial points selected on the optical fiber laying path, k∈K′ represents any spatial point in K′, Π k Indicates the operation of setting all dimensions except the dimension represented by k to zero, f θ represents a deep neural network; represents the spatial points near the spatial point k, where N is the total number of spatial points selected on the optical fiber laying path, and s is the interval between the spatial points; The parameters θ of the deep neural network are optimized in the following way to achieve denoising of seismic data: argmin θ ‖g θ (x)-x k ‖ 2 ; where x k Represents the waveform received by the seismic detector at spatial point k.

2. The method for denoising distributed acoustic wave sensing data based on high-quality sparse prior according to claim 1, characterized in that: The waveform received by the seismic detector is introduced into the deep neural network as prior knowledge to guide the deep neural network to denoise the seismic data based on distributed optical fiber acoustic sensor signals. The deep neural network adopts the U-Net network.

Images