CN113721293B - Multi-wave seismic signal artificial intelligence matching method based on deep learning - Google Patents

Abstract

The invention discloses a multi-wave seismic signal artificial intelligence matching method based on deep learning. The method specifically comprises the following steps: firstly, single-wave layered resampling is utilized, rotating wave transverse waves (PS waves) are extracted in a fractional-multiple mode, and the PS waves are preliminarily compressed to a longitudinal wave (PP wave) time domain; building and training a CNN; selecting all convolution results of 8 layers of convolution after reserving CNN as output; carrying out dimension-raising replication on the PS matrix data into three channels and carrying out normalization; dividing the complete data into a plurality of small data volumes according to the profile information for independent processing; inputting PP and PS data into a network, and outputting respective feature matrixes; copying PS data into an original output matrix, and inputting the original output matrix into a neural network to obtain an output characteristic matrix; defining a characteristic matrix double-weighted loss function, calculating the gradient of the loss function and an output matrix, updating the output matrix by gradient descent, and performing iterative updating; and averaging the data of the three channels of the output matrix to obtain single-channel data, and performing inverse normalization and noise reduction to obtain final matching data.

Description

Multi-wave seismic signal artificial intelligence matching method based on deep learning

Technical Field

The invention belongs to the field of geoscience, and particularly relates to a multi-wave seismic signal artificial intelligence matching method based on deep learning.

Background

In the past decades, the conventional seismic exploration technology plays a key role in discovery, exploitation and other aspects of resources such as petroleum, natural gas and the like, and is rapidly developing towards the direction of refinement, automation, intellectualization and the like. However, the conventional seismic exploration technology has the bottleneck problems of single wave field (only longitudinal waves), insufficient information quantity (lacking transverse waves), insufficient precision, strong multi-solution property and the like. The multi-wave multi-component seismic exploration technology (two-dimensional three-component, 2D3C, three-dimensional three-component, 3D3C, three-dimensional four-component, 3D4C) can acquire wave field information of various types such as longitudinal waves, transverse waves and converted transverse waves (P-SV), has unique advantages such as multi-wave combined imaging, multi-wave combined calibration, multi-wave combined inversion and full-wave attribute fusion, and can effectively solve complex problems such as 'gas cloud' (gas chimney) shielded imaging, lithology discrimination, crack (anisotropy) detection and fluid (oil, gas and water) identification.

However, to fully exploit the unique advantages of multi-wave multi-component seismic exploration, the problem of matching multi-wave seismic signals needs to be solved first. At present, the matching processing of multi-wave seismic signals has the following problems:

1. the matching precision of time, frequency, phase and the like is not high enough, and the reliability of multi-wave matching is poor especially in an area with low signal-to-noise ratio;

2. the matching process requires manual interference, the efficiency is low, and an automatic and intelligent matching method is lacked.

In the field of oil and gas exploration and development, the deep learning technology plays an important role in the aspects of automatic processing, data mining, knowledge learning and representation, intelligent information extraction, cost reduction, reduction of complex manual work, efficiency improvement and the like. However, in multi-wave multi-component seismic exploration, the research and application of deep learning are very few due to the influence of a plurality of factors such as the frontier of the deep learning method, the complexity of a multi-wave theory, the difficulty of data processing interpretation and feature information extraction and the like.

Although deep learning has become a research focus in the field of oil and gas exploration and development; but still in the technology starting stage at present, the industrial application from maturity is still early in the aspects of construction interpretation, fault tracking, oil and gas identification and the like. Especially in the field of multi-wave multi-component seismic exploration with great development potential, the method is limited by factors such as frontier performance, high threshold performance and the like, the research and application of the deep learning method are few, the method has wide development prospect in the aspects of multi-wave calibration, multi-wave matching, multi-wave inversion, full-wave attribute extraction and the like, and plays an important role in automation, intellectualization and high efficiency of the multi-wave seismic exploration technology.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a multi-wave seismic signal artificial intelligence matching method based on deep learning.

In the traditional multi-wave signal matching method, the aim of waveform matching is finally and indirectly achieved by changing the time, amplitude and phase of the PS wave through mathematical calculation, but the method provided by the invention does not need traditional calculation, directly finishes waveform matching by directly extracting image characteristics (waveform characteristics) of PP and PS waves through CNN, finishes phase and amplitude matching while finishing waveform matching, and greatly improves the efficiency of the whole matching process.

The technical scheme of the invention is as follows: a multi-wave seismic signal artificial intelligence matching method based on deep learning comprises the following steps:

step 1, converting the original data of the earthquake PP and PS waves into matrix data.

Step 2, taking a single-channel seismic signal (corresponding to a column in the matrix), utilizing the pre-extracted PP and PS wave position data, dividing the PP and PS waves into j parts in a time domain according to position coordinates, and recording the j parts as PP₁、PP₂、…、PP_jAnd PS₁、PS₂、…、PS_j。

Step 3, time domain compression is carried out by utilizing fractional extraction, and PP is firstly calculated₁And PS₁Length ratio N/M, where N ═ len (PP)₁) Is PP₁A length; m ═ len (PS)₁) Is PS₁A length; for PS₁Performing interpolation by N times, and using gain N and cut-off frequency N

Filtering with elliptic filter (Coule filter), and finally filtering with PS₁Performing M times of extraction to finally obtain compressed PP₁PS in time domain₁Data, PS after completing n-part compression₁，PS₂…PS_jSplicing and restoring to obtain a single PS wave which is preliminarily compressed to a PP time domain;

step 4, executing the operation of the step 2-3 on each wave, and splicing the wave into complete PS data after the operation is finished;

step 5, designing a CNN network framework: the network consists of 10 layers of convolution layers with the length of 1, 4 layers of convolution layers with the length of 2, 2 x 2 layers and the maximum pooling layer with the length of 2, two full-connection layers and one softmax layer, and network training is respectively completed on a training set, a test set and a verification set;

step 6, the network locking parameters pre-trained in the step 5 are not trained, 8 layers of convolutions after the networks are combined are extracted to be used as the output of the neural network, and the weight when the loss is calculated is given to each layer of convolution results;

step 7, preprocessing input data of the neural network: the PP and PS data are normalized, the complete data volume is divided into a plurality of small data volumes according to the section position information, the small data is copied into a three-channel matrix in an ascending-dimension mode, and each small data volume is matched independently;

step 8, copying PS data into an original output matrix OUT, inputting the original output matrix OUT into a network to obtain a characteristic matrix O, inputting PP and PS data into a neural network to obtain characteristic matrices P and S, and enabling O_n、P_n、S_nThe convolution results of the nth layer convolution pair OUT, PP and PS are respectively.

And 9, defining the square distance loss function of the nth layer convolution result as follows:

wherein, ω is^p、ω^SRespectively representing the characteristic weight values of PP and PS waves;

the total loss function is then defined as:

wherein, sum (gamma)_n) Representing the pair matrix Γ_nA value, Q, obtained by summing each element_nThe size (length × width × dimension), ω, of the feature matrix representing the convolution result of the nth layer_nA weight representing the nth layer result;

step 10, solving gradients of gamma and OUT, reducing the gradients by using an Adam optimization algorithm, updating an output matrix OUT, and performing iterative updating;

step 11, averaging the data of the three channels of the matching result to obtain single-channel data, splicing the single-channel data into complete data, and then performing inverse normalization;

and 12, carrying out noise reduction processing on the data subjected to the inverse normalization in the step 11 to obtain final matching output data.

Further, step 12 specifically uses gaussian filtering to perform noise reduction processing.

The invention has the beneficial effects that: aiming at the problems of low matching precision and low efficiency of multi-wave seismic signals in the multi-wave multi-component seismic exploration technology, the method disclosed by the invention is based on deep learning and migration learning ideas, and a Convolutional Neural Network (CNN) is adopted to extract the characteristics of the seismic longitudinal wave signals and the seismic transverse wave signals, so that the high-precision, high-efficiency, automatic and intelligent matching of the signals on time, phase and amplitude is realized, the method is suitable for pre-stack and post-stack seismic longitudinal wave and transverse wave data, and can provide technical support for subsequent applications such as multi-wave calibration, multi-wave inversion, full-wave attribute extraction, deposition and structure interpretation, fault identification, stratum tracking, reservoir prediction, hydrocarbon-containing prediction, trap discovery and the like.

Drawings

Fig. 1 is a schematic diagram of the overall process of performing preliminary matching on single-wavelength hierarchical resampling in steps 2-4.

Fig. 2 is a schematic diagram of the resampling fraction multiple extraction process in step 3.

Fig. 3 is a schematic diagram of the single-wave hierarchical matching effect in steps 2-4.

Fig. 4 is a schematic diagram of a CNN framework of the original building training.

Fig. 5 is a diagram showing the effect of retaining all convolution results of 8 convolutional layers after CNN as output.

FIG. 6 is a schematic diagram of the overall process of matching details of the neural network in steps 5-11.

Fig. 7 is a schematic diagram comparing cross sections of PP wave (left) and original PS wave (right).

FIG. 8 is a cross-sectional comparison of PP wave (left) and matched output wave (right).

FIG. 9 is a diagram showing the comparison of PP, PS and output wave amplitude spectra.

FIG. 10 is a diagram showing the comparison of PP, PS and output wave phase spectra.

Detailed Description

The following further describes the implementation of the present invention with reference to the accompanying drawings.

The method of the invention comprises the following steps:

step 1, sgy format (or other storage format) data of original seismic PP and PS waves are converted into matrix data.

Step 2, taking a single channel seismic signal (corresponding to a row of data in a matrix), as shown in the overall primary matching flow chart of fig. 1, utilizing pre-extracted PP and PS wave layer data, and dividing each channel seismic signal (corresponding to each row of the matrix) in PP and PS waves into j parts in a time domain according to a layer coordinate, wherein the cutting results are respectively recorded as: PP (polypropylene)₁、PP₂、…、PP_jAnd PS₁、PS₂、…、PS_j。

And 3, performing time domain preliminary compression on the single PS wave by utilizing a fractional extraction method in resampling, wherein as shown in fig. 2, the detailed process of the step is as follows:

step 31: first calculate PP₁，PS₁To obtain a fractional compression ratio

Step 32: for PS₁By interpolation by N, i.e. at the initial PS₁Inserting N-1 zero points between adjacent points of the signal sequence;

step 33: the signal is filtered through an elliptic filter (Cour filter) with a filter gain of N and a cut-off frequency of

Step 34: finally for PS₁Performing integer M times extraction, namely extracting 1 point every M-1 points in the signal to obtain PS compressed in time domain₁And (6) data.

Step 35: the remaining group data (e.g. PP)₂And PS₂) The same operation is carried out to complete n partsAfter compression of PS₁，PS₂…PS_jAnd splicing and reducing the PS wave into a complete PS wave.

The fractional extraction has the advantages that interpolation is carried out in the first step, the signal can be ensured not to be lost, extraction is carried out after filtering, the characteristics of the original signal can be kept to the maximum extent, and the image spectrum caused by interpolation and the spectrum aliasing caused by integral multiple extraction can be eliminated by only using an elliptic filter in the middle. The layered position compression is selected instead of one-time compression because the propagation speed ratios of PP and PS signals on different geological structures are different, the compression ratios of two waves in different time periods are different, if the two waves are directly compressed according to the compression ratios, the distance between partial obvious positions needing to be corresponding is far, the difficulty in subsequent neural network processing is high, the perfect correspondence of the obvious positions on a time domain can be ensured by using the layered position compression, the initial matching of detail positions is realized, and the subsequent small change of the detail positions is only needed in a small range. The compression effect is shown in fig. 3, the head and the tail of the arrow in the figure indicate the position of the formation point, the wave is divided into three sections according to the formation point, and the compression of the layering position is completed.

And 4, performing the operation of the step 2-3 on each channel of signal in the whole signal data, and splicing and restoring the signals into the whole PS data after finishing the time domain initial compression of all the channels of signals. The single wave hierarchical matching effect is shown in fig. 3.

Step 5, this embodiment trains a Convolutional Neural Network (CNN) using the ILSVRC-2012 data set, and the network structure is shown in fig. 4. The detailed process of the step is as follows:

step 51: downloading an ILSVRC-2012 data set, the data set comprising 1000 categories of images in total, and the data set is divided into three groups: training set (130 ten thousand), validation set (5 ten thousand), test set (10 ten thousand), raw images were normalized and randomly cropped to 224 x 3 size.

Step 52: the convolution layer is completely convoluted by 3 multiplied by 3, and 3 multiplied by 3 small convolution kernels are used for replacing large-size convolution kernels for multiple times, so that the parameter calculation amount is reduced under the condition that the receptive field is not reduced, the convolution step length is 1, and the same padding is selected to ensure that the resolution of the image input and output by the model is not changed.

Step 53: the spatial pooling is performed by 4 largest pooling layers, with pooling nucleus sizes of 2 × 2 and step length of 2.

Step 54: two fully-connected layers follow, the first layer is fully-connected with 2048 channels, the second layer is fully-connected with 1000-dimensional ILSVRC classification and contains 1000 channels, both fully-connected layers are subjected to random discarding regularization, and the last layer is a soft-max layer. All hidden layers are configured with the ReLU activation function.

Step 55: initialization of network weights is important because poor initialization may hinder learning due to instability of gradients in the deep network, and random initialization procedures of Glorot & Bengio (2010) are also used to initialize network weights.

Step 56: the batch size is set to be 256, the momentum is set to be 0.9, the learning rate is initially set to be 0.01, when the accuracy of the verification set stops improving, the learning rate is reduced by 10 times, and iterative training is carried out.

And 6, locking the network pre-trained in the step 5 to be free from training. Because the convolutional layers with different depths have different emphasis on global and local features when extracting features, in order to ensure that signals can obtain better extraction and mapping matching effects on the global and local features, as shown in fig. 5, 8 layers of convolutional layers after the network is reserved when performing PP and PS wave waveform extraction are selected to extract feature matrices (the reception fields of the first two layers are too low, and the calculation is not considered temporarily), 8 feature matrices in total are combined to be used as the output of a neural network, and the weight when calculating loss is given to the convolution result of each layer.

Step 7, preprocessing the input data of the neural network, as shown in fig. 6: because the original PP and PS data may differ by hundreds of times, the PP and PS data need to be normalized to 0-1, so that the characteristic value of the PS is prevented from being directly ignored due to small loss in matching calculation; then, according to the section position information, namely how many waves are shared by each section in the data body, the complete data is divided into a plurality of small data bodies (single small data is single section data), and each small data body is subjected to independent matching processing, so that the influence of signal data mutation at the section on subsequent feature extraction and matching can be avoided; because the network model is formed by pre-training three-channel pictures, direct processing of seismic data which can be regarded as a single channel is not supported, and the data needs to be copied into a three-channel matrix in an up-to-down mode.

Step 8, copying PS data into an original output matrix OUT, inputting the original output matrix OUT into a network to obtain a characteristic matrix O, inputting PP and PS data into a neural network to obtain characteristic matrices P and S, and enabling O_n、P_n、S_nThe convolution results of the nth layer convolution pair OUT, PP, PS, respectively.

And 9, defining the square distance loss function of the nth layer convolution result as follows:

wherein, ω is^p、ω^SRespectively representing the characteristic weight values of PP and PS waves.

The total loss function is then defined as:

wherein, sum (gamma)_n) Representing the pair matrix Γ_nA value, Q, obtained by summing each element_nThe size (length × width × dimension), ω, of the feature matrix representing the convolution result of the nth layer_nRepresenting the weight of the nth layer result.

Step 10.O_i ⁿCalculated from OUT through a neural network, which can be represented as O_i ⁿG (OUT), the loss function is given with O_i ⁿInstead of the output, Γ ═ f (o ut) may be used to calculate the gradient of the loss function Γ and the output matrix OUT, and the gradient is decreased using Adam optimization algorithm to update the output matrix OUT, and the OUT is updated iteratively.

The detailed process of the step is as follows:

step 10.1: initializing the updating step number t to 1, and if the updating step number t is initialized, updating t to t + 1;

step 10.2: the gradient of Γ versus the parameter OUT is calculated,

step 10.3: calculating the first moment (g) of the gradient_tDesired) m_t＝β₁m_t-1+(1-β₁)g_tI.e. the average of the past gradient and the current gradient, like a smoothing operation, where beta₁Is the first moment attenuation coefficient;

step 10.4: calculating the second moment (g) of the gradient_t ²Is expected) v_t＝β₂v_t-1+(1-β₂)g_t ²I.e. the average of the square of the past gradient and the current gradient, where β₂Is the second moment attenuation coefficient;

step 10.5: for the first moment m_tCorrection is made because m_tThe initial value is 0, so the bias is towards 0, and the bias influence is reduced after the processing, and the specific calculation formula is

β₁ ^tIs represented by beta₁To the t power;

step 10.6: for second moment v_tCorrection is made because v_tThe initial value is 0, so the bias is towards 0, and the bias influence is reduced after the processing, and the specific calculation formula is

β₂ ^tIs represented by beta₂To the t power;

step 10.7: updating parameters

Wherein, alpha is the learning rate, and epsilon represents the random error;

step 10.8: the steps 10.1 to 10.7 are circulated within the preset circulation times

And 11, averaging the data of the three channels of the matching result OUT to obtain single-channel data, splicing the single-channel data into complete data, and then performing inverse normalization.

And step 12, performing noise reduction processing by using Gaussian filtering to obtain final matched output data.

Fig. 7 is a cross section comparison of a PP wave (left) and an original PS wave (right), fig. 8 is a cross section comparison of a PP wave (left) and a matching output wave (right), the matching result not only realizes the whole time domain compression to the PP wave on the basis of the original PS wave, the significant layer can also better correspond to the significant layer of the PP wave, and for the insignificant layer and the detail feature which are difficult to process before, a good matching effect is also achieved after the neural network processing.

Fig. 9 is a comparison of amplitudes of PP, PS, and output waves, and it can be seen that differences between different frequency bands of amplitudes of original PP and PS waves are obvious, and after matching is completed, the output waves have a trend of approaching PP obviously on the basis of original amplitude characteristics of PS waves, so that high-precision matching of signals on the amplitudes is realized.

Fig. 10 is a phase comparison of PP, PS and output waves, and it can be seen that the phase difference between the original PP and PS waves is relatively obvious, the PP wave phase is concentrated at ± 20 degrees, and the PS wave phase is concentrated at ± 25 degrees. After matching is completed, most of the phases of the output waves are concentrated to +/-20 degrees, and the output waves have the trend of obviously approaching PP on the basis of the original phase characteristics of PS waves, so that high-precision matching of signals on the phases is realized.

The method is based on deep learning and migration learning ideas, and the convolutional neural network is adopted to extract the features of the seismic longitudinal wave signal and the seismic transverse wave signal, so that high-precision, high-efficiency, automatic and intelligent matching of the signals in time, phase and amplitude is realized. The matching method is suitable for pre-stack and post-stack seismic longitudinal wave and transverse wave data, can provide technical support for subsequent applications such as multi-wave calibration, multi-wave inversion, full-wave attribute extraction, deposition and structure interpretation, fault identification, stratum tracking, reservoir prediction, oil-gas-containing prediction, trap discovery and the like, and can lay a solid foundation for fully exerting the unique advantages of the multi-wave multi-component seismic exploration technology in the subsequent process.

Claims (2)

1. A multi-wave seismic signal artificial intelligence matching method based on deep learning comprises the following steps:

step 1, converting original data of earthquake PP and PS waves into matrix data;

step 2, taking a single-channel seismic signal, utilizing pre-extracted PP and PS wave position data, dividing PP and PS waves into j parts in a time domain according to position coordinates, and recording the j parts as PP₁、PP₂、…、PP_jAnd PS₁、PS₂、…、PS_j；

Step 3, time domain compression is carried out by utilizing fractional extraction, and PP is firstly calculated₁And PS₁Length ratio N/M, where N ═ len (PP)₁) Is PP₁A length; m ═ len (PS)₁) Is PS₁A length; for PS₁Performing N-fold interpolation, and using gain N and cut-off frequency N

Filtering by the elliptic filter, and finally, filtering by the PS₁Performing M times of extraction to obtain compressed PP₁PS in time domain₁Data; after completing n part compression, PS₁，PS₂…PS_jSplicing and restoring to obtain a single PS wave which is preliminarily compressed to a PP time domain;

step 4, executing the operation of the step 2-3 on each wave, and splicing the wave into complete PS data after the operation is finished;

step 5, designing a CNN network framework: the network consists of 10 layers of convolution layers with the length of 1, 4 layers of convolution layers with the length of 2 and the length of 2, two layers of full connection layers and one layer of softmax layer, and the network training is respectively completed on a training set, a testing set and a verification set;

step 6, the network locking parameters pre-trained in the previous step are not trained, 8 layers of convolutions after the networks are combined are extracted to be used as the output of a neural network, and the weight when loss is calculated is given to each layer of convolution results;

step 7, preprocessing input data of the neural network, normalizing PP and PS data, dividing a complete data body into a plurality of small data bodies according to section position information, copying the small data into a three-channel matrix in an ascending dimension mode, and independently matching each small data body;

step 8, copying PS data into original outputOutputting the matrix OUT, inputting the network to obtain a characteristic matrix O, inputting PP and PS data to a neural network to obtain characteristic matrices P and S, and enabling O_n、P_n、S_nConvolution results of the nth layer of convolution pair OUT, PP and PS are respectively obtained;

and 9, defining the square distance loss function of the nth layer convolution result as follows:

wherein, ω is^p、ω^SRespectively represent the characteristic weight values of PP and PS waves,

the total loss function is then defined as:

wherein, sum (gamma)_n) Representing the pair matrix Γ_nA value, Q, obtained by summing each element_nSize of the feature matrix, ω, representing the result of the n-th layer convolution_nA weight representing the nth layer result;

step 10, solving gradients of gamma and OUT, reducing the gradients by using an Adam optimization algorithm, updating an output matrix OUT, and performing iterative updating;

step 11, averaging the data of the three channels of the matching result to obtain single-channel data, splicing the single-channel data into complete data, and then performing inverse normalization;

and 12, carrying out noise reduction processing on the data subjected to the inverse normalization in the step 11 to obtain final matching output data.

2. The deep learning-based artificial intelligence matching method for multi-wave seismic signals according to claim 1, wherein step 12 specifically uses gaussian filtering for noise reduction.

Images