CN113191321A - Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network - Google Patents
Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network Download PDFInfo
- Publication number
- CN113191321A CN113191321A CN202110562514.6A CN202110562514A CN113191321A CN 113191321 A CN113191321 A CN 113191321A CN 202110562514 A CN202110562514 A CN 202110562514A CN 113191321 A CN113191321 A CN 113191321A
- Authority
- CN
- China
- Prior art keywords
- network
- attention
- seismic wave
- das
- noise
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses an optical fiber distributed seismic wave signal noise reduction method based on a generated countermeasure network, which effectively removes random noise and cable wave noise of optical fiber distributed seismic wave signals by using the generated countermeasure network. Constructing an attention cyclic neural network by using an attention mechanism, and generating a spatial attention weight matrix aiming at random noise and cable wave noise in seismic wave image signals; then, generating DAS seismic wave signals subjected to noise reduction by using a context convolution automatic encoder; and when the attention discrimination network cannot judge whether the input denoised DAS seismic signal is real data from a database or data generated by a generation network, training is finished based on an optical fiber distributed seismic signal denoising algorithm for generating a countermeasure network.
Description
Technical Field
The invention belongs to the field of optical fiber distributed seismic wave signal processing, and relates to an optical fiber distributed seismic wave signal noise reduction method based on a generation countermeasure network.
Background
A distributed optical fiber acoustic wave sensing (DAS for short) technology based on a phase-sensitive optical time domain reflectometer is a novel sensing technology for realizing continuous distributed detection of acoustic signals by utilizing an optical fiber backward Rayleigh scattering interference effect. The optical fiber sensing device has the advantages of common optical fiber sensing technology, such as electromagnetic interference resistance, good concealment, corrosion resistance, insulation and the like, can realize long-distance, distributed and real-time quantitative detection of dynamic strain (vibration and sound wave) along an optical fiber line, and has wide application prospect in the fields of security monitoring of important places and major infrastructure, health monitoring of large-scale structures, oil and gas resource exploration and the like. Particularly, when the system is applied to the technical field of vertical seismic profiling (VSP for short), the DAS-VSP technology can replace the former electronic detection seismic data acquisition system through rapid development in recent years.
Due to the characteristics of the DAS, seismic data with higher resolution and better accuracy can be obtained by utilizing the DAS-VSP technology. However, the optical cable and the borehole wall are difficult to be well coupled at a certain depth from the wellhead to the borehole and at irregular positions such as logging routes, so that very strong optical cable resonance interference noise (cable wave for short) can be generated. Although the VSP technology can avoid the noise from the earth surface during the ground observation, the VSP technology itself also has various noises, such as cable waves, borehole waves, poor coupling of downhole instruments, noise of casing waves, and the like, so that effective signals in the seismic data become fuzzy, effective analysis cannot be performed, subsequent seismic interpretation is affected, such as horizon tracking, fault recognition, and the like, and therefore, the seismic interpretation of a corresponding work area cannot be correctly given. Therefore, various noises in the acquired DAS data are removed, the signal to noise ratio of the DAS data is improved, the performance of the whole DAS system is improved, and reasonable seismic explanation is provided better. Therefore, DAS data noise reduction method research is crucial.
The existing seismic wave signal noise removing method can be divided into random noise and cable wave noise according to the type of the acting noise. According to the kind of noise, the noise reduction method can be divided into: a seismic wave signal noise reduction method aiming at cable wave noise mainly comprises 4 steps: (1) the cutting method is to cut off the area with cable wave noise, but the method directly sets the data value in the specified time window to zero, cuts off the effective wave while cutting off the noise, and may cause the distortion of the spectrum estimation value. (2) And (3) a linear coherent wave field removing technology, which is mainly used for eliminating refraction ringing and scattered waves in data. The linear coherent wave field is removed from the seismic trace by using the property of spatial trace correlation. However, the effect of the method depends on the good coherence of the linear wavelength, and the poorer the coherence, the worse the effect. (3) The wavelet transform incorporates an f-k filtering method that applies the f-k filter operator and wavelet transform in series, but loses the effective signal when the noise is removed. (4) Based on the direction filtering technology of fuzzy discrimination, linear homophase axes in the seismic records are detected firstly, cable wave noise is identified from the linear homophase axes, and finally the identified cable wave noise is subjected to direction filtering processing. But if the fuzzy criterion and the model adopted are not perfect enough, the misjudgment can occur. (II) the seismic wave signal noise reduction method aiming at random noise mainly comprises 3 types: (1) the polynomial fitting method adopts a multi-channel correlation method to complete the fitting of two aspects of the time and the amplitude of the effective wave according to the similarity of the effective signals on the space. However, this method may cause pseudo-homomorphic axis and earthworms. (2) And KL transformation, wherein coherent information is extracted and random noise is eliminated on the basis of the idea that effective components of reflected waves of adjacent channels have strong correlation on waveform and energy and random noise does not have correlation. (3) Singular Value Decomposition (SVD), which is based on the idea that effective signals are concentrated on eigenvectors corresponding to larger eigenvalues, selects eigenvectors corresponding to the larger eigenvalues to reconstruct signals and removes random noise. However, the method has a good denoising effect when the in-phase axis is horizontal, and has a poor effect when the in-phase axis is inclined or curved.
In summary, although there are a series of seismic wave signal noise reduction methods and good noise reduction effect is obtained in a specific scene, the following difficulties still exist in practical application:
1. the specific noise type is single, the method is suitable for random noise, the generalization capability is not strong, and seismic wave data with complex noise types cannot be solved, for example, the deterministic cable wave noise removing effect is not good;
2. a large amount of priori knowledge related to noise data is needed during use, and the development period of a denoising algorithm is long;
3. when the signal-to-noise ratio of the seismic wave signal is low, the loss of the effective signal is serious.
Disclosure of Invention
The invention aims to: the method for reducing the noise of the optical fiber distributed seismic wave signals based on the generation countermeasure network solves the problems that the method for reducing the noise of the seismic wave signals of the distributed optical fiber acoustic wave sensing system in the prior art is complex and tedious, poor in generalization capability, single in noise type, poor in high-noise signal processing capability and the like.
The technical scheme adopted by the invention is as follows:
an optical fiber distributed seismic wave signal noise reduction method based on a generated countermeasure network comprises the following steps: and constructing a signal database, and constructing a training set and a testing set. The DAS seismic wave signals are collected by utilizing a DAS system and combining a vertical seismic profile technology. The DAS seismic wave signals are collected, a mask matrix corresponding to a DAS seismic wave signal section diagram before denoising is used, and a signal database is composed of three parts, namely the DAS seismic wave signal section diagram before denoising, the DAS seismic wave signal section diagram after denoising corresponding to the DAS seismic wave signal section diagram before denoising and the mask matrix corresponding to the DAS seismic wave signal section diagram before denoising, and training data and testing data are divided according to a certain proportion.
Step 2: and constructing an attention generation antagonistic network based on the training set. The attention generating countermeasure network is composed of two networks, an attention generating network and an attention discriminating network. Wherein the attention generating network is composed of an attention circulation neural network and a context convolution automatic encoder;
step 2.1: and constructing an attention circulation neural network in the attention generating network. And (3) constructing a residual error network ResNet for extracting the characteristics of the input DAS seismic wave signals before denoising and the characteristics of the corresponding mask matrix in the step 1. And further generating characteristics of random noise and cable wave noise of the DAS seismic wave signals before denoising by using the long-short term memory network LSTM unit. Then, converting the extracted feature map into a 2D space attention weight matrix aiming at the two types of noise through a convolution layer;
step 2.2: a context convolution auto-encoder in an attention-generating network is constructed. Connecting the input DAS seismic wave signals before denoising and the 2D space attention weight matrix finally generated in the step 2.1 together and inputting the DAS seismic wave signals and the 2D space attention weight matrix into a context convolution automatic encoder to generate effectively denoised DAS seismic wave signals;
step 2.3: and constructing an attention discriminating network module, wherein the purpose of the attention discriminating network module is to accurately judge whether the denoised DAS seismic wave signals come from real input data or are generated by an attention generating network. When the model is converged and the attention discrimination network cannot judge the source of the denoised DAS seismic wave signals, the attention for denoising the DAS seismic wave signals generates the training of the countermeasure network, and the parameter weight of the network is obtained;
and step 3: and (5) carrying out online denoising test. And (3) directly utilizing the parameter weight generated in the step (2.3) to initialize the network during online testing, inputting the DAS seismic wave signals before denoising into the attention circulation neural network in the attention generation network constructed in the step (2.1) and the context convolution automatic encoder in the attention generation network constructed in the step (2.2), and generating the DAS seismic wave signals which are effectively denoised.
Specifically, in step 1, a DAS system is used in combination with a vertical seismic profile technology to acquire DAS seismic signals. The DAS seismic wave signals are collected, a DAS seismic wave signal section diagram before de-noising and the corresponding de-noised DAS seismic wave signal section diagram are respectively placed in two folders and are respectively named as noise _ data and clean _ data for distinguishing. Subtracting the denoised DAS seismic wave signal corresponding to the clean _ data folder from the denoised DAS seismic wave signal before denoising in the noise _ data folder to obtain a mask matrix corresponding to the denoised DAS seismic wave signal, wherein a signal database consists of three parts, namely a DAS seismic wave signal before denoising cross-sectional view, a DAS seismic wave signal after denoising corresponding to the DAS seismic wave signal cross-sectional view and a mask matrix corresponding to the DAS seismic wave signal cross-sectional view, and is according to the following steps of 7: the scale of 2 is divided into training data and test data.
The noise removed by the network comprises resonance interference noise (cable wave noise for short) with partially or completely known attributes such as positions, excitation modes and energy ranges of noise sources such as cable waves, borehole waves, underground instrument coupling failure and casing waves, and random noise with uncertain conditions such as positions, excitation modes and energy ranges of noise sources such as wind blowing, grass moving, sea waves, water flowing and rotating motion of surface soil particles.
Specifically, in step 2, based on the training set, an attention generation countermeasure network is constructed. The attention generating countermeasure network is composed of two networks, an attention generating network and an attention discriminating network. The attention discrimination network judges whether the denoised DAS seismic wave signals come from real denoising data or data generated by the attention generation network, and the attention generation network is used for generating denoised DAS seismic wave signals which enable the attention discrimination network to be incapable of distinguishing sources. The training process for generating the antagonistic network with attention is specifically as follows: during each iteration, the attention discrimination network updates K times and the attention generation network updates once, wherein K is a hyper-parameter, and the value depends on specific requirements. The two networks are continuously alternately trained, when the two networks are finally converged, the attention discrimination network cannot judge the source of the input noise-free DAS seismic wave signal, namely the attention generation network can generate de-noising data which are close to the distribution of the real DAS seismic wave signal, and the attention generation network finishes training to obtain the parameter weight of the network.
Specifically, in step 2.1, an attention-cycling neural network in the attention-generating network is constructed. And (3) using 4 residual modules, wherein each residual module comprises two convolution layers with convolution kernels of 3 x 3 and channels of 32, and stacking the convolution layers to form a residual network ResNet to extract the characteristics of the input DAS seismic signals containing noise and the corresponding mask matrix generated in the step 1. The mask matrix guides the network to focus on random noise and cable wave noise areas of the DAS seismic wave signals. And further generating the characteristic description of the random noise and the cable wave noise of the DAS seismic wave signals by taking the learned characteristics of the random noise and the cable wave noise in the DAS seismic wave signals as the feed-in of the long-term and short-term memory network LSTM unit. The extracted feature map is then transformed into a 2D spatial attention weight matrix a1 by a convolution kernel of 3 x 3 and a number of channels of 32 convolution layers.
Specifically, the operation in step 2.1 needs to be repeated 4 times, and iterates from the spatial attention weight matrix a1 to the spatial attention weight matrix a4 which can well characterize the random noise and cable wave areas of the DAS seismic signals.
Specifically, in step 2.2, the context convolutional automatic encoder in the constructed attention generating network is composed of 16 convolutional layers, and the convolutional kernel size and the number of channels of each convolutional layer are set. Connecting the input DAS seismic wave signals containing noise with the 2D space attention weight matrix finally generated in the step 2.1, performing layer-by-layer spatial domain downsampling on the input DAS seismic wave signals by using an encoder of a context convolution automatic encoder, obtaining a feature map of the DAS seismic wave signals effectively denoised by rich semantic representation while preserving the spatial relationship in data, gradually recovering spatial information by using a decoder of the context convolution automatic encoder, and finally generating the DAS seismic wave signals effectively denoised by using a feature map code.
Specifically, in step 2.3, the constructed attention discriminating network module comprises 7 convolutional layers, and the size and the number of channels of the convolutional core of each convolutional layer are set, so as to accurately judge whether the denoised DAS seismic wave signals are from real denoised data or denoised data generated by an attention generating network. When the model is converged and the attention discrimination network cannot judge the source of the input noise-free DAS seismic wave signals, namely the attention generation network can generate de-noising data of DAS seismic wave signals which are close to real distribution, the attention generation network for DAS seismic wave signal de-noising finishes training the anti-network, and the parameter weight of the network is obtained.
Specifically, in step 3, during online denoising, the attention cycle neural network and the context convolution automatic encoder of the network are initialized by using the parameter weight generated in step 2.3, and after the DAS seismic wave signals before denoising are input, the DAS seismic wave signals which are effectively denoised are generated through the attention cycle neural network and the context automatic encoder.
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
1. various noises including random noise and cable wave noise can be effectively removed at one time, and other traditional methods can only specially process one of the noises at one time;
2. the noise can be effectively reduced aiming at DAS seismic wave signals with extremely low signal-to-noise ratio;
3. the average test time of the invention is 3.5 seconds, which meets the requirement of industrial environment on real-time property;
4. the countermeasure network is generated, high-dimensional complex data distribution of input data can be learned under the precondition of less prior assumption and less data volume, and the characteristic of less prior knowledge of the DAS seismic wave signal data volume is adapted.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are required to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and that for those skilled in the art, other relevant drawings can be obtained according to the drawings without inventive effort, wherein:
FIG. 1 is a flowchart of a method according to a first embodiment of the present invention;
FIG. 2 is a diagram of a DAS seismic wave acquisition system according to a first embodiment of the present invention;
FIG. 3 is a block diagram of a second embodiment of the present invention;
FIG. 4 is a diagram of a residual error module according to a fourth embodiment of the present invention;
FIG. 5 is a diagram of random noise and cable wave regions of DAS seismic signals according to a fourth embodiment of the present invention;
FIG. 6 is a parameter diagram of an attention-cycling neural network according to a fourth embodiment of the present invention;
FIG. 7 is a parameter diagram of a convolutional layer in a context convolutional automatic encoder according to a fifth embodiment of the present invention;
FIG. 8 is a parameter diagram of an attention discriminating network according to a sixth embodiment of the present invention;
FIGS. 9(a), (c), (e) are input noisy DAS seismic signals according to a seventh embodiment of the present invention;
FIGS. 9(b), (d) and (f) are DAS seismic signals after noise reduction generated according to a seventh embodiment of the present invention;
fig. 10(a) and (b) are the snr curve and the structural similarity curve in the training phase according to the seventh embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the detailed description and specific examples, while indicating the preferred embodiment of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention. The components of embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
It is noted that relational terms such as "first" and "second," and the like, may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
The features and properties of the present invention are described in further detail below with reference to examples.
This embodiment A
As shown in a solid-line frame part of fig. 1, the invention is a DAS seismic signal denoising method based on attention generation countermeasure network, which performs denoising processing on DAS seismic signals in an end-to-end network manner. The method specifically comprises the following steps:
step 1: and (4) preparing data. And constructing a signal database, and constructing a training set and a testing set. As shown in fig. 2, a DAS system is used as a data acquisition tool for a vertical seismic profile technology to obtain DAS seismic signals, two directions of the seismic data are a depth direction and a time axis, respectively, and DAS seismic signal profiles with appropriate proportions and sampling intervals are selected to form a database. The acquired DAS seismic wave signals are divided into two folders of training data and test data according to a certain proportion. For training data in a training data folder, DAS seismic signals containing noise and DAS seismic signals without noise are respectively placed in two folders and named noise _ data and clean _ data for distinguishing. And subtracting the DAS seismic wave signals which do not contain the noise and correspond to the clean _ data folder from the DAS seismic wave signals which contain the noise in the noise _ data folder to obtain a mask matrix corresponding to the DAS seismic wave signals which contain the noise.
Step 2: and constructing an attention generation antagonistic network based on the training set. The attention generating countermeasure network is composed of two networks, an attention generating network and an attention discriminating network. Wherein the attention generating network is composed of an attention circulation neural network and a context convolution automatic encoder.
2.1: and constructing an attention circulation neural network in the attention generating network. And stacking 4 residual modules to form a residual network ResNet, extracting the input DAS seismic wave signals containing noise and the characteristics of a corresponding mask matrix generated in the previous block, and intensively learning the characteristics of random noise and cable waves of the DAS seismic wave signals by using the mask matrix. The description of the random noise and cable wave characteristics of the DAS seismic signals is further generated using long-short term memory network LSTM units. The extracted feature map is then transformed into a 2D spatial attention weight matrix by a convolutional layer.
2.2: a context convolution auto-encoder in an attention-generating network is constructed. The input noisy DAS seismic signals and the 2D spatial attention weight matrix finally generated in step 2.1 are concatenated together as input to the context convolution auto-encoder to generate noiseless DAS seismic signals. The method comprises the steps of utilizing an encoder of a context convolution automatic encoder to carry out layer-by-layer spatial domain down-sampling on input DAS seismic wave signals, finally obtaining a feature diagram of noise-free DAS seismic wave signals expressed by rich semantics, utilizing a decoder of the context convolution automatic encoder to gradually recover spatial information, and finally generating noise-free DAS seismic wave signals through feature diagram codes.
2.3: and constructing an attention discrimination network module, wherein the aim of the attention discrimination network module is to accurately judge whether the noise-free DAS seismic wave signals are from real input data or generated by an attention generation network. When the model is finally converged, the attention discrimination network cannot judge the source of the input noiseless DAS seismic wave signal, namely, the attention generation network can generate a sample of the noiseless DAS seismic wave signal distribution which accords with the real distribution, and the attention generation network for DAS seismic wave signal noise reduction finishes training the opposing network to obtain the parameter weight of the network.
And step 3: and (5) carrying out online denoising test. In the online test, an attention discriminating network module in the step 2.3 is not needed, the network is initialized by directly utilizing the parameter weight generated in the step 2.3, the DAS seismic wave signal containing noise is input into the attention circulation neural network in the attention generating network constructed in the step 2.1 and the context convolution automatic encoder in the attention generating network constructed in the step 2.2, and the DAS seismic wave signal without noise is generated.
Example two
On the basis of the first embodiment, in a preferred embodiment of the present invention, a structural block diagram of the DAS seismic wave signal data acquisition system serving as the vertical seismic profiling technology in step 1 is shown in fig. 3, and the data acquisition system includes a light source, a modulation unit, an amplification unit, an optical fiber sensing unit, a data processing unit, and a photoelectric conversion unit. Specifically, a high-coherence narrow-linewidth laser is used as a system light source to output laser, a signal driver drives a modulator to modulate continuous light input by the light source into pulsed light, and the pulsed light is low in power and needs to be amplified by an EDFA erbium-doped fiber amplifier to be used as detection pulsed light of the system. The detection pulse light is input from an a port of the circulator and is injected into the sensing optical fiber through a b port of the circulator, backward Rayleigh scattering light is generated in the sensing optical fiber by the detection pulse light, the generated backward Rayleigh scattering light is transmitted to the detector through a c port of the circulator, the detector converts a received light signal into an electric signal, the signal driver drives the acquisition card to acquire a receiving end signal, and finally DAS seismic wave signals are output. And (3) selecting DAS seismic wave signals with the section size ratio of 823-531 and the sampling interval of a time axis of 1ms to construct a database. And (3) acquiring DAS seismic wave signals according to the following steps of: the scale of 2 is divided into two folders of training data and test data. For training data in a training data folder, DAS seismic signals containing noise and DAS seismic signals without noise are respectively placed in two folders and named noise _ data and clean _ data for distinguishing. And subtracting the DAS seismic wave signals which do not contain the noise and correspond to the clean _ data folder from the DAS seismic wave signals which contain the noise in the noise _ data folder to obtain a mask matrix corresponding to the DAS seismic wave signals which contain the noise.
EXAMPLE III
On the basis of the above embodiment, in a preferred embodiment of the present invention, in step 2, based on the training set, the attention generating countermeasure network is constructed. The attention generating countermeasure network is composed of two networks, an attention generating network and an attention discriminating network. The attention discrimination network is used for judging whether the noiseless DAS seismic signals are from real input data or generated data of the attention generation network, and the attention generation network is used for generating the noiseless DAS seismic signals which enable the attention discrimination network to be incapable of distinguishing sources. For an attention discriminating network, the discrimination power at the beginning cannot be too strong, otherwise it is difficult to improve the power of attention generating network, but also cannot be too weak, otherwise the final generation of the attention generating network trained for it is generally poor. The training process for generating the antagonistic network with attention is specifically as follows: during each iteration, the attention discrimination network updates K times and the attention generation network updates once, wherein K is a hyper-parameter, and the value depends on specific requirements. The two networks are continuously alternately trained, when the two networks are finally converged, the attention discrimination network cannot judge the source of the input noise-free DAS seismic wave signal, namely, the attention generation network can generate a sample which accords with the distribution of the real noise-free DAS seismic wave signal, the attention generation network for DAS seismic wave signal noise reduction is used for finishing the training of the opposition network, and the parameter weight of the network is obtained.
Example four
Based on the above embodiment, in a preferred embodiment of the present invention, the structure of the residual module in the attention-cycling neural network in the attention generating network constructed in step 2.1 is shown in fig. 4, each residual block includes two convolutional layers with 3 × 3 convolutional kernels and a channel number of 32, and the activation function is ReLU. And 4 residual modules are stacked to form a residual network ResNet for extracting the characteristics of the input DAS seismic wave signals containing noise and the corresponding mask matrix generated by the previous block. The mask matrix will direct the network to focus on the random noise and cable waves of the DAS seismic signals as shown in fig. 5. And splicing the extracted feature matrixes of the random noise and the cable wave in the DAS seismic wave signals together to serve as feed-in of the long-term and short-term memory network LSTM unit, and further generating feature descriptions of the random noise and the cable wave of the DAS seismic wave signals. The extracted signature is then transformed into a 2D spatial attention weight matrix a1 by a convolutional layer with 3 x 3 convolutional kernel and channel number 32, with the activation function ReLU. As shown in FIG. 6, the complete operation in step 2.1 needs to be repeated 4 times, and iterated from the spatial attention weight matrix A1 to the spatial attention weight matrix A4 which can well characterize the random noise and cable wave regions of the DAS seismic signals.
EXAMPLE five
On the basis of the above embodiment, in a preferred embodiment of the present invention, the context convolutional automatic encoder in the attention generating network constructed in step 2.2 is composed of 16 convolutional layers, as shown in fig. 7. The encoder part of the context convolution automatic encoder comprises a convolution layer with 64 convolution kernels and 5 x 5 channels, a convolution layer with 3 x 3 convolution kernels and 128 convolution kernels and four convolution layers with 3 x 3 convolution kernels and 256 convolution kernels, and all the activation functions are ReLU. The decoder part of the context convolution automatic encoder comprises a convolution layer with 32 convolution kernels and 3 x 3 channels, a convolution layer with 64 convolution kernels and 4 x 4 channels, a convolution layer with 128 convolution kernels and 3 x 3 channels, a convolution layer with 128 convolution kernels and 4 x 4 channels and a convolution layer with 256 convolution kernels and 3 x 3 channels, and all the activation functions are ReLU. Connecting the input DAS seismic wave signals containing noise with the 2D space attention weight matrix finally generated in the step 2.1, performing layer-by-layer spatial domain down-sampling on the input DAS seismic wave signals by using an encoder of a context convolution automatic encoder in a mild mode, obtaining a feature map of the DAS seismic wave signals without noise represented by rich semantics while preserving the spatial relationship in data, gradually recovering spatial information by using a decoder of the context convolution automatic encoder, and finally generating the DAS seismic wave signals without noise by using a feature map code.
EXAMPLE six
Based on the above embodiment, in a preferred embodiment of the present invention, the attention discriminating network module constructed in step 2.3 includes 7 convolutional layers, as shown in fig. 8, including one convolutional layer with 8 convolutional kernels and 5 × 5, one convolutional layer with 16 convolutional kernels and 5 × 5, two convolutional layers with 32 convolutional kernels and 5 × 5, two convolutional layers with 128 convolutional kernels and 5 × 5, one convolutional layer with 64 convolutional kernels and 5 × 5, and one fully-connected layer with 1024 neurons, where the activation functions are all ReLU. The objective is to accurately determine whether the noiseless DAS seismic signals are derived from real input data or are generated by an attention-generating network. When the model is finally converged, the attention discrimination network cannot judge the source of the input noiseless DAS seismic wave signal, namely, the attention generation network can generate a sample of the noiseless DAS seismic wave signal distribution which accords with the real distribution, and the attention generation network for DAS seismic wave signal noise reduction finishes training the opposing network to obtain the parameter weight of the network.
EXAMPLE seven
On the basis of the above embodiment, in a preferred embodiment of the present invention, in the step 3, a schematic block diagram of the online denoising test is shown as a dashed-line frame part in fig. 1. And (3) initializing an attention cycle neural network and a context convolution automatic encoder of the network by using the parameter weight generated in the step 2.3, inputting the DAS seismic wave signals containing noise as shown in the figure 9(a), and generating the DAS seismic wave signals subjected to noise reduction by the context convolution automatic encoder through the attention cycle neural network and the context convolution automatic encoder as shown in the figure 9 (b). The average duration of the online test period was 3.5 seconds, the average PSNR was 25.1296, and the average SSIM was 0.79061.
The technical solution of the present embodiment is essentially or partially contributed to by the prior art, and is/are embodied in the form of a software product.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and should not be taken as limiting the scope of the present invention, and any modifications, equivalents and improvements made by those skilled in the art within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (8)
1. The optical fiber distributed seismic wave signal noise reduction method based on the generation countermeasure network is characterized by comprising the following steps: the method for denoising DAS seismic wave signals by adopting an end-to-end network mode comprises the following steps:
step 1: constructing a signal database, and establishing a training set and a test set:
the DAS seismic wave signals are collected, corresponding DAS seismic wave signal profiles before denoising are subtracted by DAS seismic wave signal profiles before denoising to obtain mask matrixes corresponding to the DAS seismic wave signal profiles before denoising, a signal database is composed of the DAS seismic wave signal profiles before denoising, the DAS seismic wave signal profiles after denoising corresponding to the DAS seismic wave signal profiles before denoising and the mask matrixes corresponding to the DAS seismic wave signal profiles before denoising, and is divided into training data and testing data to obtain a training set and a testing set;
step 2: constructing an attention generation countermeasure network based on a training set:
the attention generation countermeasure network is composed of two networks, namely an attention generation network and an attention discrimination network, wherein the attention generation network is composed of an attention circulation neural network and a context convolution automatic encoder;
step 2.1: constructing an attention circulation neural network in an attention generation network, and constructing a residual error network ResNet for extracting the characteristics of the input DAS seismic wave signals before denoising and the characteristics of the corresponding mask matrix in the step 1; further generating features of random noise and cable wave noise of the DAS seismic wave signals before denoising by using a long-short term memory network LSTM unit, and then converting the extracted feature map into a 2D space attention weight matrix aiming at the two types of noise through a convolution layer;
step 2.2: constructing a context convolution automatic encoder in an attention generating network, connecting the input DAS seismic wave signals before denoising and the 2D space attention weight matrix finally generated in the step 2.1 together, and inputting the DAS seismic wave signals after denoising effectively;
step 2.3: constructing an attention discrimination network module, wherein the purpose of the attention discrimination network module is to accurately judge whether the denoised DAS seismic wave signals come from real input data or are generated by an attention generation network, and when the model is converged and the attention discrimination network cannot judge the source of the denoised DAS seismic wave signals, the attention generation network for DAS seismic wave signal denoising finishes training the resistance network and obtains the parameter weight of the network;
and step 3: and (3) performing online denoising test, namely directly utilizing the parameter weight generated in the step 2.3 to initialize a network during the online test, inputting the DAS seismic wave signals before denoising into the attention circulation neural network in the attention generation network constructed in the step 2.1 and the context convolution automatic encoder in the attention generation network constructed in the step 2.2, and generating the DAS seismic wave signals which are effectively denoised.
2. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 1 comprises the following steps:
collecting DAS seismic wave signals, wherein a DAS seismic wave signal section diagram before de-noising and a corresponding de-noised DAS seismic wave signal section diagram are respectively placed in two folders and respectively named as noise _ data and clean _ data for distinguishing; subtracting the denoised DAS seismic wave signal corresponding to the clean _ data folder from the denoised DAS seismic wave signal before denoising in the noise _ data folder to obtain a mask matrix corresponding to the denoised DAS seismic wave signal, wherein a signal database consists of three parts, namely a DAS seismic wave signal before denoising cross-sectional view, a DAS seismic wave signal after denoising corresponding to the DAS seismic wave signal cross-sectional view and a mask matrix corresponding to the DAS seismic wave signal cross-sectional view, and is according to the following steps of 7: 2, dividing the ratio into training data and testing data to obtain a training set and a testing set;
the noise removed by the network comprises resonance interference noise with partially or completely known attributes such as positions, excitation modes and energy ranges of noise sources such as cable waves, borehole waves, underground instrument coupling failure and casing waves, and random noise with uncertain conditions such as positions, excitation modes and energy ranges of noise sources such as wind blowing, grass moving, sea waves, water flowing and rotating motion of surface soil particles.
3. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 2 comprises the following procedures:
the attention discrimination network judges whether the denoised DAS seismic wave signals come from real denoised data or data generated by an attention generation network, and the attention generation network is used for generating denoised DAS seismic wave signals which enable the attention discrimination network to be incapable of distinguishing sources;
the training process for generating the antagonistic network with attention is specifically as follows:
during each iteration, the attention distinguishing network updates K times and the attention generating network updates once, wherein K is a hyper-parameter, and the value depends on specific requirements; the two networks are continuously alternately trained, when the two networks are finally converged, the attention discrimination network cannot judge the source of the input noise-free DAS seismic wave signal, namely the attention generation network can generate de-noising data which are close to the distribution of the real DAS seismic wave signal, and the attention generation network finishes training to obtain the parameter weight of the network.
4. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 2.1 comprises the following steps:
constructing an attention circulation neural network in an attention generation network, using 4 residual modules, stacking to form a residual network ResNet to extract input DAS seismic wave signals containing noise and the characteristics of corresponding mask matrixes generated in the step 1, wherein each residual module comprises convolution layers with 3 × 3 convolution kernels and 32 channels; the mask matrix guides a network to focus on random noise and cable wave noise areas of the DAS seismic signals, learned characteristics of the random noise and the cable wave noise in the DAS seismic signals are used as feed-in of long and short term memory network LSTM units, characteristic descriptions of the random noise and the cable wave noise of the DAS seismic signals are generated, and then extracted characteristic diagrams are converted into a 2D space attention weight matrix A1 through a convolution layer with a convolution kernel of 3 x 3 and a channel number of 32.
5. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the operation in step 2.1 needs to be repeated 4 times, and iterates from the spatial attention weight matrix a1 to the spatial attention weight matrix a4 which can well characterize the random noise and cable wave areas of the DAS seismic signals.
6. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 2.2 comprises the following steps:
the context convolution automatic encoder in the constructed attention generation network is composed of 16 convolution layers, and the convolution kernel size and the channel number of each convolution layer are set. Connecting the input DAS seismic wave signals containing noise with the 2D space attention weight matrix finally generated in the step 2.1, performing layer-by-layer spatial domain downsampling on the input DAS seismic wave signals by using an encoder of a context convolution automatic encoder, obtaining a feature map of the DAS seismic wave signals effectively denoised by rich semantic representation while preserving the spatial relationship in data, gradually recovering spatial information by using a decoder of the context convolution automatic encoder, and finally generating the DAS seismic wave signals effectively denoised by using a feature map code.
7. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 2.3 comprises the following steps:
the constructed attention discriminating network module comprises 7 layers of convolution layers, and the size and the channel number of convolution kernels of each layer of convolution layer are set, so that the purpose is to accurately judge whether denoised DAS seismic wave signals come from real denoised data or denoised data generated by an attention generating network;
when the model is converged and the attention discrimination network cannot judge the source of the input noise-free DAS seismic wave signals, namely the attention generation network can generate de-noising data of DAS seismic wave signals which are close to real distribution, the attention generation network for DAS seismic wave signal de-noising finishes training the anti-network, and the parameter weight of the network is obtained.
8. The method for noise reduction of fiber optic distributed seismic signals based on a generative countermeasure network of claim 1, wherein: the step 2.3 comprises the following steps:
and during online denoising, initializing an attention cycle neural network and a context convolution automatic encoder of the network by using the parameter weight generated in the step 2.3, and generating the DAS seismic wave signals which are effectively denoised through the attention cycle neural network and the context automatic encoder after the DAS seismic wave signals before denoising are input.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110562514.6A CN113191321B (en) | 2021-05-21 | 2021-05-21 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110562514.6A CN113191321B (en) | 2021-05-21 | 2021-05-21 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113191321A true CN113191321A (en) | 2021-07-30 |
| CN113191321B CN113191321B (en) | 2022-04-22 |
Family
ID=76984827
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110562514.6A Active CN113191321B (en) | 2021-05-21 | 2021-05-21 | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113191321B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115587291A (en) * | 2022-09-26 | 2023-01-10 | 华中科技大学 | Denoising characterization method and system based on crack ultrasonic scattering matrix |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107644400A (en) * | 2016-07-21 | 2018-01-30 | 中国石油化工股份有限公司 | The Texture Segmentation Methods and device of seismic section image |
| CN108828670A (en) * | 2018-08-20 | 2018-11-16 | 成都理工大学 | A kind of seismic data noise-reduction method |
| CN109890043A (en) * | 2019-02-28 | 2019-06-14 | 浙江工业大学 | A kind of wireless signal noise-reduction method based on production confrontation network |
| CN110045419A (en) * | 2019-05-21 | 2019-07-23 | 西南石油大学 | A kind of perceptron residual error autoencoder network seismic data denoising method |
| CN111190227A (en) * | 2020-01-09 | 2020-05-22 | 吉林大学 | Low signal-to-noise ratio seismic data denoising method based on residual convolution generation countermeasure model |
| WO2020174459A1 (en) * | 2019-02-27 | 2020-09-03 | Ramot At Tel-Aviv University Ltd. | A distributed-acoustic-sensing (das) analysis system using a generative-adversarial-network (gan) |
| US20200285916A1 (en) * | 2019-03-06 | 2020-09-10 | Adobe Inc. | Tag-based font recognition by utilizing an implicit font classification attention neural network |
| CN111738940A (en) * | 2020-06-02 | 2020-10-02 | 大连理工大学 | Human face image eye completing method for generating confrontation network based on self-attention mechanism model |
| CN111983681A (en) * | 2020-08-31 | 2020-11-24 | 电子科技大学 | Seismic wave impedance inversion method based on countermeasure learning |
| CN112116906A (en) * | 2020-08-27 | 2020-12-22 | 济南浪潮高新科技投资发展有限公司 | On-site sound mixing method, device, equipment and medium based on GAN network |
-
2021
- 2021-05-21 CN CN202110562514.6A patent/CN113191321B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107644400A (en) * | 2016-07-21 | 2018-01-30 | 中国石油化工股份有限公司 | The Texture Segmentation Methods and device of seismic section image |
| CN108828670A (en) * | 2018-08-20 | 2018-11-16 | 成都理工大学 | A kind of seismic data noise-reduction method |
| WO2020174459A1 (en) * | 2019-02-27 | 2020-09-03 | Ramot At Tel-Aviv University Ltd. | A distributed-acoustic-sensing (das) analysis system using a generative-adversarial-network (gan) |
| CN109890043A (en) * | 2019-02-28 | 2019-06-14 | 浙江工业大学 | A kind of wireless signal noise-reduction method based on production confrontation network |
| US20200285916A1 (en) * | 2019-03-06 | 2020-09-10 | Adobe Inc. | Tag-based font recognition by utilizing an implicit font classification attention neural network |
| CN110045419A (en) * | 2019-05-21 | 2019-07-23 | 西南石油大学 | A kind of perceptron residual error autoencoder network seismic data denoising method |
| CN111190227A (en) * | 2020-01-09 | 2020-05-22 | 吉林大学 | Low signal-to-noise ratio seismic data denoising method based on residual convolution generation countermeasure model |
| CN111738940A (en) * | 2020-06-02 | 2020-10-02 | 大连理工大学 | Human face image eye completing method for generating confrontation network based on self-attention mechanism model |
| CN112116906A (en) * | 2020-08-27 | 2020-12-22 | 济南浪潮高新科技投资发展有限公司 | On-site sound mixing method, device, equipment and medium based on GAN network |
| CN111983681A (en) * | 2020-08-31 | 2020-11-24 | 电子科技大学 | Seismic wave impedance inversion method based on countermeasure learning |
Non-Patent Citations (7)
| Title |
|---|
| H. WU 等: ""A Novel DAS Signal Recognition Method Based on Spatiotemporal Information Extraction With 1DCNNs-BiLSTM Network"", 《 IEEE》 * |
| HASAN ASY’ARI ARIEF 等: ""A Survey on Distributed Fibre Optic Sensor Data Modelling Techniques and Machine Learning Algorithms for Multiphase Fluid Flow Estimation"", 《SENSORS》 * |
| X. DONG 等: ""Denoising the Optical Fiber Seismic Data by Using Convolutional Adversarial Network Based on Loss Balance"", 《 IEEE 》 * |
| 周文辉 等: ""基于残差注意力网络的地震数据超分辨率方法"", 《计算机科学》 * |
| 张旭苹 等: ""相位敏感光时域反射分布式光纤传感技术"", 《光学学报》 * |
| 张晓普: ""分布式地震数据智能感知采集方法研究"", 《中国硕士学位论文全文数据库 基础科技辑》 * |
| 饶云江: ""长距离分布式光纤传感技术研究进展"", 《物理学报》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115587291A (en) * | 2022-09-26 | 2023-01-10 | 华中科技大学 | Denoising characterization method and system based on crack ultrasonic scattering matrix |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113191321B (en) | 2022-04-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111442827B (en) | Optical fiber passive online monitoring system for transformer winding vibration | |
| CN113049084B (en) | Attention mechanism-based Resnet distributed optical fiber sensing signal identification method | |
| US11740372B1 (en) | Method and system for intelligently identifying carbon storage box based on GAN network | |
| CN111104891B (en) | Composite characteristic optical fiber sensing disturbing signal mode identification method based on BiLSTM | |
| CN109818892A (en) | Construct Cyclic Spectrum characteristic parameter extraction model and signal modulation mode recognition methods | |
| CN111948711B (en) | Method and system for extracting low-frequency part of seismic data by deep learning method | |
| CN113687433A (en) | Bi-LSTM-based magnetotelluric signal denoising method and system | |
| CN113191321B (en) | Optical fiber distributed seismic wave signal noise reduction method based on generation countermeasure network | |
| CN115291281A (en) | Real-time micro-earthquake magnitude calculation method and device based on deep learning | |
| CN117056680A (en) | Data noise reduction and signal detection method, device and system and storage medium | |
| CN115905805A (en) | DAS data multi-scale noise reduction method based on global information judgment GAN | |
| Wang et al. | Attribute-guided target data separation network for DAS VSP data | |
| Sui et al. | Complete perception self-attention network for weak seismic signal recovery in distributed acoustic sensing vertical seismic profile data | |
| CN117292494A (en) | Signal identification method, system, computer equipment and medium for sound and vibration fusion | |
| Zhong et al. | Multiscale residual pyramid network for seismic background noise attenuation | |
| Wu et al. | Multi-scale progressive fusion attention network based on small sample training for DAS noise suppression | |
| Zhao et al. | Background noise suppression for DAS-VSP records using GC-AB-unet | |
| CN117214398B (en) | Deep underground water body pollutant detection method and system | |
| CN117688422B (en) | Under-modeling state parameter identification method based on improved sparse component analysis, computer equipment and storage medium | |
| Zhao et al. | Ensemble empirical mode decomposition and stacking model for filtering borehole distributed acoustic sensing records | |
| Li et al. | Removing abnormal environmental noise in nodal land seismic data using deep learning | |
| CN117574062A (en) | Small loop transient electromagnetic signal denoising method based on VMD-DNN model | |
| Cheng et al. | Simultaneous denoising and reconstruction of distributed acoustic sensing seismic data via a multicascade deep-learning method | |
| Zhao et al. | Self-supervised denoising for distributed acoustic sensing vertical seismic profile data via improved blind spot network | |
| CN115236733A (en) | DAS-VSP data background noise suppression method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |