A noise positioning and waveform constraint-based intelligent suppression method for wave signal noise

Abstract

The application discloses a kind of wave signal noise intelligent suppression methods based on noise positioning and waveform constraint, it is related to wave signal coupling noise suppression field.The application first constructs noisy wave signal dataset, and by automatically positioning the depth position of coupled noise existing in wave signal acquisition data, the effective signal feature learning of network to the coupled noise shelter is strengthened, the physical information difference between effective signal and coupled noise is fully mined, the U-Net network trained and constructed by combining the global loss, local loss set to improve the suppression ability of coupled noise.The application can be applicable in the wave signal acquisition data under different geological conditions, different offset, accurately identify and effectively suppress coupled noise, so as to significantly improve the signal-to-noise ratio of effective signal, improve processing efficiency while reducing the degree of artificial participation in the process of suppressing coupled noise, reduce effective signal damage and improve applicability in actual data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wave signal coupling noise suppression, and in particular to a wave signal noise intelligent suppression method based on noise positioning and waveform constraint. BACKGROUND

[0002] Wave signal detection technology has broad development prospects as a high-resolution detection technology. Coupling noise is a typical interference wave field in wave signal data acquisition, and the energy of coupling noise is generally higher than that of the effective signal, and the distribution frequency band and the effective signal frequency band range are highly overlapped, which greatly affects the signal-to-noise ratio of the acquired data.

[0003] At present, the methods suitable for coupling noise suppression mainly include the following: 1. According to the distribution characteristic difference, apparent velocity difference, sparse characteristics and the like of coupling noise and effective signal in the transform domain, a window, a threshold or other methods are set to perform traditional denoising processing, such as band-pass filtering, F-K (frequency-wave number) filtering, Radon transform filtering, singular value decomposition filtering, morphological component analysis method, etc.; 2. Denoising method based on data analysis and machine learning technology, which realizes the purpose of denoising by learning and mining the potential characteristic rules of signal and noise, mainly including dictionary learning, unsupervised learning, supervised learning, etc.

[0004] The above methods can improve the quality of wave signal acquisition data to some extent, but still have certain limitations, such as the denoising effect of traditional methods being excessively dependent on manual parameter adjustment, and the effective signal being easily damaged; in the denoising method based on data analysis and machine learning technology, there is no publicly available high-quality actual wave signal data for training, and the training of artificial synthetic wave signal data does not conform to the actual situation, resulting in unsatisfactory denoising effect, loss of effective signal energy, etc. SUMMARY

[0005] In order to solve the problems of the prior art, the present application provides a wave signal noise intelligent suppression method based on noise positioning and waveform constraint, which is used to construct a rich noisy wave signal data set, and according to the multi-domain characteristic difference between wave signal and coupling noise, to realize automatic positioning of coupling noise and learning of wave signal characteristics in noisy wave signal data, so as to improve the fidelity of denoised wave signal data and reduce the damage to effective signal, thereby improving the accuracy of coupling noise recognition and wave signal data denoising.

[0006] The technical scheme adopted by the present application is as follows:

[0007] A wave signal noise intelligent suppression method based on noise positioning and waveform constraint, comprising the following steps:

[0008] Step 1, constructing a noisy wave signal data set:

[0009] synthetic pure wave signal data; filtering the collected wave signal of the target region to obtain effective wave signal data and coupling noise data; simulating the generation of coupling noise data;

[0010] Based on the synthetic pure wave signal data and the effective wave signal data obtained by filtering, a clean wave signal data set is obtained; based on the coupling noise data obtained by filtering and the simulated coupling noise data, a coupling noise data set is obtained;

[0011] The clean wave signal data set and the coupling noise data set are linearly superimposed, and the intensity ratio of the coupling noise and the clean wave signal data is controlled to obtain synthetic noisy wave signal data;

[0012] Step 2, the noisy wave signal data and its corresponding clean wave signal data are respectively sliced with the same sliding step, and the noisy wave signal data block and the clean wave signal data block at the same slice position after slicing are taken as a pair of sample data block and label data block. The sample data block and the label data block are normalized to the range of [-1, 1], and a training data set is constructed based on the normalized sample data block and the label data block;

[0013] Step 3, a U-Net network for generating clean wave signal is constructed, the input of which is the data block, and the output is the predicted wave signal data block. Based on the constructed training data set, the model parameter training of the U-Net network is carried out, and the trained U-Net network is obtained;

[0014] Wherein, the loss function used in the model parameter training includes: the global loss between the predicted wave signal data block and the label data block, and the local loss of the predicted wave signal data block and the label data block at the spatial position set Q where the coupling noise exists; the global loss includes reconstruction loss, waveform-structure loss and frequency-wave number energy distribution loss; the local loss includes reconstruction loss and waveform-structure loss at the spatial position set Q;

[0015] Wherein, the spatial position set Q is obtained based on the noisy wave signal data set, the signal alignment is carried out based on the waveform cross-correlation coefficient under multiple scales, and the coupling noise positioning is carried out based on the energy change rate and the main frequency change rate under each scale after alignment, to obtain the spatial position set Q;

[0016] Step 4, the sliding slicing and normalization processing of the wave signal data to be suppressed are carried out to match the input of the U-Net network constructed in step 3; then the processed wave signal data is input into the trained U-Net network, and the noise suppression result of the wave signal data to be suppressed is obtained based on the output.

[0017] Further, in step 1, the simulation generates coupling noise data, including:

[0018] Method 1: Collecting strong coupling noise samples before the first arrival of the fluctuation signal acquisition data to obtain coupling noise data with randomly changing waveforms;

[0019] Method 2: Randomly selecting a segment of coupling noise to attenuate the energy to simulate coupling noise data with randomly changing intensities;

[0020] Method 3: Using an attenuated sinusoidal wave function to simulate coupling noise data with high-frequency attenuation after the arrival of the effective signal, and adding it to the position after the arrival of the first arrival wave;

[0021] Randomly combine the above three methods to simulate coupling noise data.

[0022] Further, the normalization processing uses amplitude maximum normalization.

[0023] Further, multiply the predicted fluctuation signal data block output by the U-Net network by the maximum amplitude used in the normalization processing to restore the predicted fluctuation signal data block to the original amplitude.

[0024] Further, the U-Net network includes an encoder and a decoder, as shown in Figure 2 wherein the encoder includes a plurality of encoding units each consisting of two cascaded convolution blocks, the definition N represents the number of encoding units, and a max-pooling layer is cascaded after each of the first to N-1st encoding units, the decoder includes N-1 decoding units each consisting of two cascaded convolution blocks and a convolution layer with a 1x1 convolution kernel, and an up-sampling layer is cascaded before each decoding unit; wherein each convolution block includes a convolution layer, a batch normalization, and a ReLU activation function in sequence; the input of the first up-sampling layer is the output feature map of the last encoding unit of the encoder, the inputs of the 2nd to N-1st up-sampling layers are the output feature maps of the decoding units of the previous levels; and the input feature map of any nth decoding unit is the concatenation and fusion feature map of the output feature map of the nth up-sampling layer and the output feature map of the N-nth encoding unit in the channel dimension; and the number of channels of the output feature maps of the nth decoding unit and the N-nth encoding unit is consistent; that is, starting from the 2nd up-sampling layer, the nth up-sampling layer is used to up-sample the output of the nth-1 decoding unit to the size of the output feature map of the N-nth encoding unit, and then concatenate and fuse them in the channel dimension to serve as the input feature map of the nth decoding unit.

[0025] Further, the acquisition of the spatial position set Q specifically includes:

[0026] The definition z represents a target space sampling position, a plurality of scales are selected, and a plurality of symmetric points with a distance of k unit distances from the target at z are set as reference positions of each scale at each scale, wherein k is a positive integer; and the reference positions of any two scales do not coincide;

[0027] For each scale, the waveform maximum cross-correlation coefficients between each reference position and the target at z are calculated respectively, and the average value is taken as the waveform maximum cross-correlation coefficient of the signal at z under the current scale.

[0028] According to the waveform cross-correlation coefficient, the first existing position of noise (i.e. the position where noise may exist) is recorded: the waveform maximum cross-correlation coefficients are sorted in descending order, and the first occupancy quantile of the signal and noise limit is set (which can be adjusted according to the signal-to-noise ratio), and the signal below the first occupancy quantile is identified as noise, and the existing positions of noise under each scale are recorded.

[0029] Energy and frequency change rate calculation: according to the waveform maximum cross-correlation coefficients under each scale, the signals are aligned, and then the energy change rate and the frequency change rate between each reference position and the target at z under each scale are calculated respectively, and the average value is taken as the energy change rate and the frequency change rate of the signal at z under each scale.

[0030] According to the energy and frequency change rate, the second existing position of noise is recorded: the calculated energy change rates and frequency change rates are sorted in descending order respectively, and the second occupancy quantile of the signal and noise limit is set (which can be adjusted according to the signal-to-noise ratio), and the signal below the second occupancy quantile is identified as noise, and the existing positions of noise under each scale are recorded.

[0031] The intersection position based on the first and second existing positions is obtained as the noise position of each scale.

[0032] Based on the noise position of each scale, a voting mechanism is used to determine the final noise position, and a spatial position set Q is obtained.

[0033] Further, three scales are selected, and the reference positions of each scale are respectively:

[0034] The first scale is at z-1 and z+1;

[0035] The second scale is at z-3, z-2, z+2 and z+3;

[0036] The third scale is at z-5, z-4, z+4 and z+5.

[0037] Based on the noise position of each scale, a voting mechanism is used to determine the final noise position, and a spatial position set Q is obtained.

[0038] Further, the waveform-structure loss includes a waveform similarity coefficient and structural similarity , specifically:

[0039] waveform similarity coefficient :

[0040]

[0041] wherein, is a predicted wave signal data block, is a label data block, is a mean value of the predicted wave signal data block, is a mean value of the label data block, is an expectation according to a distribution level, is an expectation at a batch level;

[0042] structural similarity :

[0043]

[0044] wherein, is a covariance of the predicted wave signal data block and the label data block, is a variance of the predicted wave signal data block, is a variance of the label data block, , is a stabilizing constant to prevent the denominator from being zero.

[0045] The technical solution provided by the present application at least brings the following beneficial effects:

[0046] The method can provide a rich and high-quality noisy wave signal data set for training a network model, and automatically locate the coupled noise, construct a physical constraint condition based on the waveform difference between the effective signal and the coupled noise and the F-K energy distribution form difference, fully learn the physical characteristics of the wave signal data in the time-space domain and the frequency-wave number domain, improve the fidelity of the denoised wave signal data, reduce the damage to the effective signal, and make the coupled noise identification, wave signal data denoising and other work more accurate and efficient. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 A flowchart of a wave signal noise intelligent suppression method based on noise positioning and waveform constraint provided by the embodiment of the present application;

[0049] Figure 2 Synthetic noiseless wave signal data schematic diagram;

[0050] Figure 3 Synthetic coupling noise waveform diagram;

[0051] Figure 4 Synthetic coupling noise schematic diagram;

[0052] Figure 5 Synthetic noise-containing wave signal data schematic diagram;

[0053] Figure 6 Adopted U-Net network structure schematic diagram;

[0054] Figure 7 Denoised wave signal data schematic diagram;

[0055] Figure 8 Removed noise schematic diagram;

[0056] Figure 9 Residual of prediction and label schematic diagram. DETAILED DESCRIPTION

[0057] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will describe the technical solutions in the embodiments of the present application in a more detailed and complete manner with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Generally, the components of the embodiments of the present application described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application.

[0058] The embodiments of the present application propose a wave signal noise intelligent suppression method based on noise positioning and waveform constraint, which is a deep learning method for suppressing coupled noise of wave signals based on noise positioning and waveform apparent velocity constraint. The existing coupled noise suppression method has problems such as difficulty in parameter selection, damage to effective signals and scarcity of high-quality data sets, etc. The method proposed in the embodiments of the present application realizes a method for constructing high-quality noise-containing wave signal data, takes the waveform and F-K energy distribution form as the physical constraint condition, trains the network model to learn the wave signal features, and automatically locates the depth position of the coupled noise existing in the wave signal acquisition data, strengthens the deep learning network to learn the effective signal features in the coupled noise shielding place, further improves the suppression ability of the coupled noise, and effectively overcomes the shortcomings of the traditional method. The denoising of the noise-containing wave signal data is realized, and the method of the present application provides a feasible solution for the coupled noise suppression in the actual work area.

[0059] In one embodiment, referring to Figure 1 The noise positioning and waveform constraint-based intelligent suppression method for wave signals provided by the embodiment of the present application comprises the following steps:

[0060] Step S1, constructing a noisy wave signal data set:

[0061] Synthesizing pure wave signal data by simulation; filtering the collected wave signals of the target region to obtain effective wave signal data and coupled noise data; and simulating coupled noise data;

[0062] Based on the pure wave signal data synthesized by simulation and the effective wave signal data obtained by filtering, a clean wave signal data set is obtained; based on the coupled noise data obtained by filtering and the simulated coupled noise data, a coupled noise data set is obtained;

[0063] Linearly superimposing the clean wave signal data set and the coupled noise data set, and controlling the intensity ratio of the coupled noise and the clean wave signal data to obtain synthesized noisy wave signal data;

[0064] Step S2, constructing a training data set based on sliding slicing and normalization processing;

[0065] The noisy wave signal data and its corresponding clean wave signal data are respectively sliced using the same sliding step, and the noisy wave signal data block and the clean wave signal data block at the same slice position after slicing are taken as a pair of sample data block and label data block. The sample data block and the label data block are normalized to the range of [-1, 1], and the normalized sample data block and the label data block are used to construct a training data set;

[0066] Step 3, constructing a U-Net network for generating clean wave signals;

[0067] The input of the constructed U-Net network is the current data block, and the output is the predicted wave signal data block. The U-Net network is trained based on the constructed training data set to obtain a trained U-Net network;

[0068] Wherein, the loss function used in the model parameter training includes: the global loss between the predicted wave signal data block and the label data block, and the local loss of the predicted wave signal data block and the label data block at the spatial position set Q where the coupled noise exists; the global loss includes reconstruction loss, waveform-structure loss and frequency-wave number energy distribution loss; the local loss includes reconstruction loss and waveform-structure loss at the spatial position set Q;

[0069] Wherein, the spatial position set Q is obtained based on the noisy fluctuation signal data set, the signal alignment is performed based on the waveform cross-correlation coefficient under multiple scales, the coupled noise positioning is performed based on the energy change rate and the main frequency change rate under each scale after alignment, so as to obtain the spatial position set Q.

[0070] Step 4, obtaining the noise suppression result of the fluctuation signal data to be suppressed based on the trained U-Net network;

[0071] The fluctuation signal data to be suppressed is subjected to sliding slicing and normalization processing to match the input of the U-Net network constructed in step 3; then the processed fluctuation signal data is input into the trained U-Net network, and the noise suppression result of the fluctuation signal data to be suppressed is obtained based on the output.

[0072] In one embodiment, in step S1, for the pure fluctuation signal data, the following two ways can be used to obtain: one is to solve the elastic wave fluctuation equation by finite difference forward modeling to generate synthetic clean fluctuation signal data; the other is to obtain by using F-K filtering to process the actual work area collected fluctuation signal data containing coupled noise. For the coupled noise, on the one hand, it can be obtained by using F-K filtering to process the actual work area collected fluctuation signal data containing coupled noise; on the other hand, a large number of strong coupled noise samples existing in the initial arrival of the fluctuation signal acquisition data can be collected to establish a coupled noise sample library; and on the other hand, an attenuation sinusoidal function can be used to simulate the coupled noise after the arrival of the effective signal. Finally, the obtained clean fluctuation signal data set and the coupled noise data set are linearly superimposed, and the intensity ratio of the coupled noise and the clean fluctuation signal data is controlled to obtain the synthetic noisy fluctuation signal data.

[0073] In one embodiment, step S1 specifically includes the following steps:

[0074] S1-1, synthetic fluctuation signal data generation:

[0075] In order to further expand the training data set, in this embodiment, the velocity model is used to synthesize pure fluctuation signal data, i.e. noise-free fluctuation signal data, by forward modeling elastic wave fluctuation equation.

[0076] Firstly, a tilted layered medium containing 6 different velocity layers is constructed, a total of 214 geophones are arranged, the arrangement interval is 15 meters, the arrangement depth range is 400 meters to 3595 meters, a Ricker wavelet with a main frequency of 30 Hz is used as P wave excitation, and the time sampling interval is 2 ms. The elastic wave simulation is carried out in the two-dimensional medium underground to obtain the wave field components along the Z direction (vertically downward) and the X direction (horizontally right). Figure 2 A synthetic Z component data is shown.

[0077] S1-2, actual fluctuation signal data preprocessing:

[0078] Using the actual work area collected coupling noise fluctuation signal data, first using F-K filtering method to suppress coupling noise. In order to solve the problem of dispersion in F-K transform domain filtering, this embodiment uses T-P transform domain filtering to solve the influence caused by dispersion phenomenon, so as to obtain more accurate and effective signal and coupling noise.

[0079] S1-3, coupling noise data set generation:

[0080] In actual data, there are three main characteristics of coupling noise: one is that the waveform of coupling noise changes randomly; two is that the intensity of coupling noise changes randomly; three is that the coupling noise is enhanced after the arrival of effective signal. In addition to the coupling noise that can be obtained in the actual fluctuation signal preprocessing process, in order to expand the coupling noise sample library, this embodiment synthesizes high-quality coupling noise according to the characteristics of coupling noise in actual data. For feature one, this embodiment collects a large number of strong coupling noise samples existing in the first arrival of fluctuation signal acquisition data, obtains the waveform random change coupling noise signal of actual acquisition, part of the coupling noise signal waveform is shown in Figure 3 For feature two, randomly select the energy of the fragment of coupling noise to simulate the intensity random change in actual data; for feature three, use the decay sinusoidal function to simulate the coupling noise with high frequency decay after the arrival of effective signal, which is added to the position after the arrival of first arrival longitudinal wave and first arrival transverse wave:

[0081] (1)

[0082] wherein, is the decay index, is the frequency, is the initial phase, is the time.

[0083] In actual acquisition data, the three main characteristics of coupling noise will appear randomly, and when synthesizing coupling noise, the three methods are randomly combined to obtain coupling noise data that conforms to the actual situation, as shown in Figure 4 .

[0084] S1-4, noisy fluctuation signal data generation:

[0085] The noise-free fluctuation signal data generated in S1-1 and the actual fluctuation signal data obtained in S1-2 are taken as a clean fluctuation signal data set, the actual coupling noise obtained in S1-2 and the coupling noise data generated in S1-3 are taken as a coupling noise data set, the clean fluctuation signal data set is linearly superimposed with the coupling noise data set, and the intensity ratio of the coupling noise and the fluctuation signal data is controlled, so that the synthesized noisy fluctuation signal data is obtained, as shown in Figure 5 .

[0086] In one embodiment, step S2 specifically comprises:

[0087] S2-1, data sliding slicing

[0088] The fluctuation signal data set is a typical spatiotemporal fluctuation signal observation data, and its structure can be characterized as a two-dimensional matrix of , where M is the total amount of time sampling points (long axis dimension), N is the total number of spatial sampling points (short axis dimension), and the dimension characteristics satisfy M >> N, forming a significantly asymmetric narrowband matrix form. In view of the data structure characteristics, the data slicing strategy of a rectangular sliding window is adopted in this embodiment, wherein the sliding window size is set as , H represents the long axis dimension slicing size, and W represents the short axis dimension slicing size. By defining the adjustable sliding step (long axis step) and (short axis step), the training sub-sample set is systematically generated.

[0089] By adjusting the sliding step and parameter combination, the overlap degree between samples can be effectively controlled, so as to maximize the data utilization rate and improve the generalization ability of model training under the premise of ensuring the continuity of features.

[0090] S2-2, data input

[0091] The fluctuation signal data with coupling noise and the corresponding clean fluctuation signal data are sliced with the sliding step , for the time sampling points and the spatial sampling points, and the sliced fluctuation signal data block with coupling noise and the corresponding clean fluctuation signal data block are taken as a pair of training and label data input into the U-Net network constructed in this embodiment, wherein the fluctuation signal data block with coupling noise is training data, and the corresponding clean fluctuation signal data block is label data. When inputting, in order to ensure the randomness of training, all obtained training and label data pairs can be shuffled and then loaded, a batch size corresponding number of data pairs are randomly grabbed, and a four-dimensional data structure is obtained, wherein the dimensions represent the batch size, the channel number, the data height and the width, so that the data can be input into the U-Net network for training and prediction.

[0092] S2-3, amplitude maximum value normalization

[0093] To enhance the model's ability to analyze the dynamics of wave propagation, the embodiment adopts a normalization method based on extreme value constraints. For the amplitude fluctuation characteristics of the wave signal, independent normalization operations are performed on the training data and label data (both have a data dimension of ). The normalization reference of each matrix is obtained by the maximum absolute value operator as shown in equation (2), and a linear scale transformation is performed to the [-1, 1] interval as shown in equation (3).

[0094] (2)

[0095] (3)

[0096] wherein, , represent the sample data block (training data block) and the label data block, respectively, , are the absolute maximum values of the sample data block and the label data block, respectively. The relative amplitude change after normalization is more conducive to the model to capture the wave field dynamics, eliminates the dimensional differences between different data, and makes the gradient update direction more consistent.

[0097] The U-Net network constructed in the embodiment adopts an encoder-decoder architecture and a jump connection structure, which not only preserves data details but also extracts rich features. For the problem of suppressing coupled noise in complex wave signal acquisition data, the sliced wave signal data containing coupled noise (training data) and the corresponding clean wave signal data (label data) are used as the input of the network model. By introducing constraints on the amplitude and waveform of the wave signal data, and combining the energy distribution area and morphological differences of the training data and the label data in the F-K transform domain to establish physical constraint conditions, the deep learning network can effectively learn the multi-domain feature differences between the training data and the label data and establish an accurate mapping relationship. To strengthen the network's ability to suppress coupled noise, the embodiment designs an automatic positioning method for the spatial position of the coupled noise in the wave signal acquisition data, and increases the network's learning of the effective signal features in the energy-shielded place by the coupled noise. Further, the U-Net network constructed in the embodiment realizes precise suppression of coupled noise while preserving the details and structure of the effective signal.

[0098] The encoder part of the U-Net network mainly extracts features and down-samples the input data through a double convolution module and a down-sampling module. The double convolution module is composed of two layers of asymmetric convolution kernels. Due to the asymmetry of the wave signal data, the use of asymmetric convolution kernels can better adapt to the size of the data, realize the improvement of the time direction resolution and the maintenance of the spatial direction feature sensitivity; the down-sampling module cascades a maximum pooling layer and a double convolution module, realizes the compression of the spatial dimension and the expansion of the channel, and enables the encoder to gradually learn more abstract feature expressions from shallow to deep. The encoder contains 3 levels of down-sampling, the feature map size is halved step by step, and the channel number increases exponentially, forming a pyramid feature structure. The decoder mainly maps the deep features extracted by the encoder back to the spatial resolution of the original input to generate the final output. In the up-sampling module, bilinear interpolation is used to realize 2 times of spatial expansion, and cross-layer features are fused through channel splicing. At the same time, in each stage of the decoder, the features of different abstract levels of the encoder are spliced with the up-sampling results through the jump connection to form multi-scale feature fusion. These modules gradually recover the spatial resolution through multi-level up-sampling operations, realize end-to-end mapping through multi-level feature abstraction and cross-layer feature fusion, and provide an efficient computing framework for the denoising task. In the U-Net network, a 1x1 convolution is used to map the high-dimensional feature space obtained by the decoder to the target dimension to obtain the output predicted wave signal data block At this time, the data distribution range is [-1, 1], in order to restore the predicted wave signal data block to its original amplitude range, the original amplitude of the data is restored by multiplying the maximum amplitude used in the normalization process

[0099] (4)

[0100] wherein, is the predicted output, is the absolute maximum amplitude of the corresponding label data used in the normalization process, is the de-normalized wave signal data block (predicted data), i.e. the target predicted output is restored to the original amplitude range consistent with the label.

[0101] In one embodiment, the U-Net network used in the present embodiment is as shown in Figure 6 ​As shown, it comprises an encoder and a decoder, wherein the encoder comprises a plurality of encoding units composed of two cascaded convolution blocks, defining N represents the number of encoding units, and a maximum pooling layer is cascaded after the first to the N-1th level encoding units, and the decoder comprises N-1 decoding units composed of two cascaded convolution blocks and a convolution layer with a 1*1 convolution kernel; an upsampling layer is cascaded before each decoding unit; wherein each convolution block comprises a convolution layer, a batch normalization and a ReLU activation function in turn; the input of the first upsampling layer is the output feature map of the last encoding unit of the encoder, the input of the second to the N-1th upsampling layer is the output feature map of the decoding unit of the previous level; and the input feature map of any nth decoding unit is: the output feature map of the nth upsampling layer and the output feature map of the N-nth encoding unit are spliced and fused feature maps in the channel dimension; and the number of channels of the output feature maps of the nth decoding unit and the N-nth encoding unit is consistent; that is, from the second upsampling layer, the nth upsampling layer is used to upsample the output map of the nth-1 decoding unit to the size of the output feature map of the N-nth encoding unit, and then spliced as the input feature map of the nth decoding unit after splicing in the channel dimension.

[0102] In the embodiment, the size of the input U-Net network is 256*16, the number of channels is 1, the convolution kernel size of each convolution block is set to 5*3, the pooling kernel of the maximum pooling layer is set to 2*2, the encoder comprises 4 encoding units, and the decoder comprises 3 decoding units, wherein the output channel numbers of the 4 encoding units are set to 8, 16, 32 and 32 in turn; and the channel numbers of the output feature maps of the three decoding units are set to 32, 26 and 8 in turn.

[0103] Since the typical feature of the coupling noise distinguishing from the effective signal is that it exists at a fixed depth, there is no necessary correlation between the coupling noises received in different detectors, and the generation of the coupling noise is more random. The existence of such coupling noise destroys the wave field propagation characteristics of the wave signal, including the interference of the coupling signal waveform on the wave signal waveform, the destruction of the continuity of the wave signal phase axis, the sudden energy enhancement at the depth of the coupling noise, and the deviation of the main frequency of the observation signal from the main frequency of the wave signal. Based on the difference between the coupling noise and the wave signal characteristics, an embodiment of the present application designs an automatic coupling noise positioning method. The wave signal at each depth position of the wave signal data containing the coupling noise is calculated in multiple scales, and the specific calculation method is as follows:

[0104] Step one: scale selection. Taking the signal at the zth spatial sampling position as an example, three scales are selected, the first scale is at the z-1, z+1 position, the second scale is at the z-3, z-2, z+2, z+3 position, and the third scale is at the z-5, z-4, z+4, z+5 position.

[0105] For each scale (take scale 1 as an example):

[0106] Step two: Calculate the maximum cross-correlation coefficient of the waveform. Calculate the maximum cross-correlation coefficient of the waveform between z-1 and z+1 and the target z respectively, and take the average as the maximum cross-correlation coefficient of the waveform of the signal at z under this scale.

[0107] Step three: Record the possible location of noise according to the cross-correlation coefficient of the waveform. After calculating the cross-correlation coefficient of the waveform corresponding to each place, sort all the calculated maximum cross-correlation coefficients of the waveform from large to small, set the signal and noise limit occupancy quantile (which can be adjusted according to the signal-to-noise ratio), and identify it as noise if it is below the occupancy quantile. Record the number of places where noise exists under this scale.

[0108] Step four: Calculate the energy and frequency change rate. After aligning the signal according to the maximum cross-correlation coefficient of the waveform, calculate the energy change rate and frequency change rate between z-1 and z+1 and the target z respectively under the same time range, and take the average as the energy and frequency change rate of the signal at z under this scale.

[0109] Step five: Record the possible location of noise according to the energy and frequency change rate. After calculating the energy and frequency change rate corresponding to each place, sort all the calculated energy and frequency change rates from large to small, set the signal and noise limit occupancy quantile (which can be adjusted according to the signal-to-noise ratio), and identify it as noise if it is below the occupancy quantile. Record the spatial position where noise exists under this scale.

[0110] Step six: Noise labeling in each scale. The inverse of the position recorded as noise in steps three and five is labeled as noise under this scale.

[0111] Between multiple scales:

[0112] Step seven: Noise determination. Use a voting mechanism, for example: In three scales, all scales are marked as noise in the spatial position, then determine that this position is noise; Or in two or more scales, the spatial position is marked as noise, then determine that this position is noise.

[0113] Based on the above method, the spatial position of the coupled noise can be automatically and accurately located, saving manual work.

[0114] In one embodiment, the loss function used by the constructed U-Net network during training is set as:

[0115] First, by calculating the waveform similarity coefficient between the clean fluctuation signal data block (label data) and the predicted fluctuation signal data block (output data), the network learns the phase change of the signal waveform in the label data, suppresses the waveform interference of the coupled noise, and the waveform similarity coefficient The calculation formula is as follows:

[0116] (5)

[0117] wherein, is the predicted fluctuation signal data block, is the labeled fluctuation signal data block, is the mean of the predicted fluctuation signal data block, is the mean of the labeled fluctuation signal data block, is the expectation of the data distribution level, is the expectation of the batch level.

[0118] Then, by calculating the structural similarity between the clean fluctuation signal data block (labeled data) and the predicted fluctuation signal data block (output data), the network learns the average energy level, energy fluctuation intensity and spatial structure correlation in the labeled data, and the structural similarity The calculation formula is as follows:

[0119] (6)

[0120] wherein, is the covariance of the predicted fluctuation signal data block and the labeled fluctuation signal data block, is the variance of the predicted fluctuation signal data block, is the variance of the labeled fluctuation signal data block, , is a stabilizing constant to prevent the denominator from being zero. From and The waveform-structure loss is composed of two parts to ensure that the physical characteristics between the output and the labeled signal match, thereby optimizing the performance of the model.

[0121] By converting the clean fluctuation signal data block (labeled data) and the predicted fluctuation signal data block (output data) to the frequency-wavenumber (F-K) domain, the energy distribution difference of the two in the F-K domain is calculated to ensure that the energy intensity of the prediction and the labeled signal matches in each frequency-wavenumber combination. The F-K domain energy distribution loss The purpose is to minimize the distribution cross-entropy of the prediction and the labeled signal, so that the network learns the distribution form of the labeled data in the F-K domain, and suppresses the coupling noise energy. The F-K domain energy distribution loss The calculation formula is as follows:

[0122] (7)

[0123] wherein, is the F-K domain energy distribution of the predicted fluctuation signal data block, is the F-K domain energy distribution of the labeled fluctuation signal data block, KL divergence operator, is the distribution contrast relation. The performance of the model is optimized by ensuring the frequency-wavenumber characteristics between the output and the label consistent through the constraint of F-K energy distribution.

[0124] The main idea of denoising is to consider the differences between the coupled noise and the wave signal based on multiple characteristics, and to comprehensively consider the reconstruction loss , waveform-structure loss and frequency-wavenumber energy distribution loss The loss function between the prediction and the label data is calculated to optimize the model, so that the network can better learn the characteristics of the label data. By defining the weight of each loss, control their contribution to the total loss, so that it can recover the effective signal and suppress the noise on the time and space scale, so as to obtain more accurate denoising effect.

[0125] (8)

[0126] where, is the mean absolute error loss, is the mean square error loss, , respectively represent the number of signal samples and the number of spatial samples.

[0127] In summary, the global loss function between the predicted wave signal data block P and the label wave signal data block S can be obtained as follows:

[0128] (9)

[0129] where, is the reconstruction loss weight, is the waveform-structure loss weight, is the frequency-wavenumber loss weight.

[0130] In order to strengthen the learning of the effective signal characteristics of the coupled noise shielding place by the deep learning network, after automatically identifying the spatial position set Q where the coupled noise exists in the noisy wave signal data block by using the coupled noise automatic positioning method, the local loss function of the predicted wave signal data block P and the label wave signal data block S at the spatial position set Q is calculated as follows:

[0131] (10)

[0132] where, is the reconstruction loss at the spatial position set Q where the coupled noise exists, is the TD Constraint Layer constraint loss at the spatial position set Q where the coupled noise exists. ​​

[0133] To fully and effectively construct the loss function between the physical information difference between the signal and the coupled noise, two parts of the loss are considered as a global loss and a local position loss at the depth of the coupled noise, and the total loss is formed by weighted addition according to the dynamically adjusted weight proportion :

[0134] (11)

[0135] wherein, is a global loss weight, is a local loss weight.

[0136] In the early stage of training, the global loss weight is higher than the local loss weight , and the model pays more attention to the global features of the signal; with the training, the local loss weight is gradually increased, and the model pays more attention to the effective signal features at the position disturbed by the coupled noise. The dynamic adjustment of the weighting mechanism helps the model to balance the optimization of the global and local features at different training stages.

[0137] For the noisy fluctuation signal in Figure 5 , the noise intelligent suppression method based on noise positioning and waveform constraint provided by the embodiment of the present application is used for denoising experiment, and the time is represented by Time, and the depth is represented by Depth, the denoising result is as shown in Figure 7 , the suppressed noise is as shown in Figure 8 , Figure 7 The residual error between the denoising result and the Figure 4 label signal is as shown in Figure 9 . It can be seen that the method provided by the embodiment of the present application can effectively suppress the coupled noise while retaining the effective signal.

[0138] Based on the constructed noisy fluctuation signal data set, the depth position of the coupled noise existing in the fluctuation signal acquisition data is automatically positioned, the learning of the effective signal features at the position shielded by the coupled noise by the constructed U-Net network is strengthened, the physical information difference between the effective signal and the coupled noise is fully mined, the physical constraint conditions are combined with the F-K energy distribution form difference and the waveform difference, and the physical constraint model suitable for various fluctuation signal wave fields is constructed. The method of the present application can be applied to the fluctuation signal acquisition data under different geological conditions and different offset distances, accurately identify and effectively suppress the coupled noise, thereby significantly improve the signal-to-noise ratio of the effective signal, improve the data quality, improve the data processing efficiency, reduce the artificial participation in the process of suppressing the coupled noise, reduce the damage to the effective signal and improve the applicability in actual data.

Claims

1. A method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints, characterized in that, Includes the following steps: Step 1, construct a dataset of noisy fluctuation signals: Pure wave signal data is synthesized through simulation; the wave signal in the acquired target area is filtered to obtain effective wave signal data and coupled noise data; coupled noise data is generated through simulation. Then, a clean fluctuation signal dataset is obtained based on the pure fluctuation signal data and the effective fluctuation signal data; A coupled noise dataset is obtained based on the coupled noise data obtained through filtering and the coupled noise data generated by simulation. The clean wave signal dataset and the coupled noise dataset are linearly superimposed, and the intensity ratio of the coupled noise to the clean wave signal data is controlled to obtain the synthesized noisy wave signal data. Step 2: Slide the noisy fluctuation signal data and the corresponding clean fluctuation signal data with the same sliding step size. After slicing, the noisy fluctuation signal data block and the clean fluctuation signal data block at the same slice position are used as a pair of sample data block and label data block. The sample data block and label data block are normalized to the range of [-1, 1], and then a training dataset is constructed based on the sample data block and label data block. Step 3: Construct a U-Net network for generating clean wave signals. Its input is a block of sample data and its output is a block of predicted wave signals. Train the model parameters of the U-Net network based on the training dataset. The loss functions used during model parameter training include: the global loss between the predicted wave signal data block and the label data block, and the local loss between the predicted wave signal data block and the label data block at the set of spatial locations Q where coupled noise exists; the global loss includes reconstruction loss, waveform-structure loss and frequency-wavenumber energy distribution loss; the local loss includes reconstruction loss and waveform-structure loss at the set of spatial locations Q. The spatial location set Q is obtained based on a noisy wave signal dataset, signal alignment is performed based on waveform cross-correlation coefficients at multiple scales, and coupled noise localization is performed based on the energy change rate and dominant frequency change rate at each scale to obtain the spatial location set Q. Step 4: Perform sliding slicing and normalization on the fluctuation signal data to be suppressed to match the input of the constructed U-Net network; then feed the processed fluctuation signal data into the trained U-Net network, and obtain the noise suppression result of the fluctuation signal data to be suppressed based on its output.

2. The method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints as described in claim 1, characterized in that, In step 1, the simulation of generating coupled noise data includes: Method 1: Collect strong coupling noise samples from the wave signal acquisition data before the first arrival to obtain coupling noise data with randomly changing waveforms; Method 2: Randomly select segments of coupled noise to attenuate energy, and simulate the generation of coupled noise data with randomly varying intensity; Method 3: Use an attenuated sine wave function to simulate the high-frequency attenuation of coupled noise data after the arrival of the effective signal, and add it to the position after the arrival of the first wave; The three methods described above are randomly combined to simulate the generation of coupled noise data.

3. The method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints as described in claim 1, characterized in that, The normalization process uses the maximum amplitude value as the normalization method.

4. The intelligent noise suppression method for fluctuating signals based on noise localization and waveform constraints as described in claim 1, characterized in that, The predicted fluctuation signal data block output by the U-Net network is multiplied by the maximum amplitude used during normalization to restore the predicted fluctuation signal data block to its original amplitude.

5. The method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints as described in claim 1, characterized in that, The U-Net network includes encoders and decoders; The encoder comprises multiple coding units consisting of two cascaded convolutional blocks, with N representing the number of coding units. Each coding unit from level 1 to level N-1 is followed by a max-pooling layer. The decoder comprises N-1 decoding units consisting of two cascaded convolutional blocks and a single convolutional layer with a 1×1 kernel. Each decoding unit is preceded by an upsampling layer. Each convolutional block sequentially includes a convolutional layer, batch normalization, and a ReLU activation function. The input to the first upsampling layer is the output feature map of the last coding unit of the encoder, and the inputs to the second to N-1 upsampling layers are the output feature maps of the previous decoding unit. The input feature map of any nth decoding unit is a fused feature map formed by concatenating the output feature map of the nth upsampling layer and the output feature map of the Nnth coding unit along the channel dimension. The number of channels in the output feature maps of the nth decoding unit and the Nnth coding units remains consistent.

6. The intelligent noise suppression method for fluctuating signals based on noise localization and waveform constraints as described in claim 1, characterized in that, Obtaining the spatial location set Q specifically includes: Define z as the sampling position in the target space. Select several scales, and set several symmetrical points at a distance of k units from the target z at each scale as the reference position of each scale, where k is a positive integer; and the reference positions of any two scales do not coincide. For each scale, the maximum cross-correlation coefficient of the waveform between each reference position and the target position z is calculated and the average value is taken as the maximum cross-correlation coefficient of the waveform at the z-th position at the current scale. Record the first location of noise based on the cross-correlation coefficient of the waveform: sort the waveforms in descending order of the maximum cross-correlation coefficient, set the first placeholder of the signal-noise boundary, and identify the signal below the first placeholder as noise, and record the number of places where noise exists at each scale. Energy and frequency change rate calculation: Align the signal according to the maximum cross-correlation coefficient of the waveform at each scale, and then calculate the energy change rate and frequency change rate between each reference position and the target position z at each scale, and take the average value as the energy change rate and frequency change rate of the signal at the z-th position at each scale. Record the second location of noise based on energy and frequency change rate: Sort all calculated energy change rate and frequency change rate in descending order, set the second placeholder of the signal-noise boundary, identify signals below the second placeholder as noise, and record the spatial location of noise at each scale. The noise location at each scale is obtained based on the intersection of the first and second locations. Based on the noise location at each scale, a voting mechanism is used to determine the final noise location, resulting in a spatial location set Q.

7. The method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints as described in claim 6, characterized in that, When obtaining the spatial location set Q, three scales are selected, and the reference locations for each scale are as follows: The first scale is at the z-1 and z+1 positions; The second scale is at positions z-3, z-2, z+2, and z+3. The third scale is at z-5, z-4, z+4, and z+5.

8. The method for intelligent suppression of fluctuation signal noise based on noise localization and waveform constraints as described in claim 1, characterized in that, Waveform-structure loss includes waveform similarity coefficient. and structural similarity Specifically, it is as follows: Waveform similarity coefficient : ； in, For predicting fluctuation signal data blocks, For the label data block, To predict the mean of the fluctuating signal data block, The mean of the labeled data blocks. According to the expected value of the distribution level, Expectations at the batch level; Structural similarity : ； in, To predict the covariance between the fluctuation signal data block and the tag data block, To predict the variance of fluctuating signal data blocks, The variance of the labeled data blocks. , To prevent the stability constant from having a denominator of zero.

Images