Multi-scale neural network multi-angle constraint seismic data low-frequency extension method and device

Abstract

The invention discloses a multi-scale neural network multi-angle constraint seismic data low-frequency extension method and device. The method comprises the following steps: constructing a plurality of speed models; constructing broadband wavelets to drive a plurality of speed models to perform forward modeling to obtain synthetic data; the synthetic data comprises high-frequency data and low-frequency data which have an implicit association relationship; constructing a bidirectional circulation multi-scale neural network, and pre-training the bidirectional circulation multi-scale neural network based on the synthetic data; inputting the collected high-frequency operation data into the pre-trained bidirectional circulation multi-scale neural network to predict and obtain low-frequency prediction data, and inputting the low-frequency prediction data into the pre-trained bidirectional circulation multi-scale neural network to predict and obtain high-frequency prediction data; and performing multi-angle loss calculation on the high-frequency prediction data and the high-frequency operation data, constraining the bidirectional circulation multi-scale neural network to perform model optimization processing, and performing low-frequency extension on the seismic data by using the bidirectional circulation multi-scale neural network after the model optimization processing.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of seismic data, and specifically to a method and device for low-frequency extension of seismic data with multi-scale neural network and multi-angle constraints. Background Art

[0002] Low-frequency signals in seismic data are of great significance in marine seismic exploration. Low-frequency seismic signals can improve mid-deep imaging, enhance seismic data resolution and imaging accuracy, improve inversion quality, and even be used for reservoir inversion and fluid detection. Therefore, it is necessary to protect and expand low-frequency signals in seismic data based on deep learning. In the wide-band processing of seismic data, compared with the expansion of the high-frequency end, the effective signal at the low-frequency end is easily affected by factors such as low-frequency noise and interference waves, so it is more difficult to recover. Due to the influence of low-frequency surge noise, ghost waves, formation absorption, velocity model and other factors, low-frequency imaging of offshore towed cable data is extremely difficult.

[0003] The expansion of seismic low-frequency signals plays an important role in resolution enhancement, waveform inversion, impedance inversion, etc. The expansion of seismic signals in the low-frequency part can effectively reduce the amplitude of the sub-wave sidelobe and reduce the interference with adjacent times. The scattering and formation absorption of low-frequency components have little effect, so they penetrate deeper in the actual seismic exploration process and can provide more structural information of the middle and deep layers for seismic data. In the full waveform inversion process, low-frequency data can make the inversion process converge to a more realistic geological model. In the impedance inversion work, the rich low-frequency components can effectively eliminate the frequency band gap during inversion, and have a better signal-to-noise ratio on the low-frequency signal without relying on logging information. Therefore, the lack of low-frequency components in marine streamer seismic data has a huge impact on the inversion process.

[0004] In order to eliminate the adverse effects of the lack of low-frequency components of seismic data on inversion and imaging, domestic and foreign scholars have studied the problem of low-frequency extension of seismic data, using methods such as frequency extension methods based on continuous wavelet transform, frequency extension methods based on frequency domain filtering, low-frequency extension methods based on phase tracking, low-frequency extension methods based on compressed sensing, seismic envelope inversion methods, and convolution methods based on reflectivity coefficients and broadband source wavelets. However, these methods have the following problems: traditional methods are sensitive to signal noise, there is uncertainty in the optimization process, and it is difficult to handle complex geological structures; artificial intelligence methods rely on a large amount of high-quality training data and computing resources, and have limitations in capturing long-distance dependencies. In addition, the complexity of multi-task learning and self-supervised learning is high, and the task weight selection and optimization strategy have a great impact on the model effect. Summary of the invention

[0005] In view of the above problems, an embodiment of the present invention is proposed to provide a method and device for low-frequency extension of seismic data using multi-scale neural network and multi-angle constraints, which overcomes the above problems or at least partially solves the above problems.

[0006] According to one aspect of an embodiment of the present invention, a multi-scale neural network multi-angle constrained seismic data low-frequency extension method is provided, the method comprising:

[0007] Construct multiple velocity models;

[0008] Construct broadband wavelet to drive multiple velocity models for forward modeling to obtain synthetic data; the synthetic data includes high-frequency data and low-frequency data, and there is an implicit correlation between the high-frequency data and the low-frequency data;

[0009] Construct a bidirectional recurrent multi-scale neural network and pre-train it based on synthetic data;

[0010] The collected high-frequency operation data is input into a pre-trained bidirectional recurrent multi-scale neural network to predict low-frequency prediction data, and the low-frequency prediction data is input into a pre-trained bidirectional recurrent multi-scale neural network to predict high-frequency prediction data; multi-angle loss calculation is performed on the high-frequency prediction data and the high-frequency operation data, and the bidirectional recurrent multi-scale neural network is constrained to perform model optimization processing, so as to use the bidirectional recurrent multi-scale neural network after model optimization processing to perform low-frequency extension on the seismic data; wherein, the loss calculation is performed using a multi-angle mixed loss function composed of a time domain loss function, a trace correlation loss function and a frequency domain loss function.

[0011] According to another aspect of an embodiment of the present invention, a multi-scale neural network multi-angle constrained seismic data low-frequency extension device is provided, which includes:

[0012] Velocity model module, suitable for building multiple velocity models;

[0013] The synthetic data module is suitable for constructing broadband wavelet-driven multiple velocity models for forward modeling to obtain synthetic data; the synthetic data contains high-frequency data and low-frequency data, and there is an implicit correlation between the high-frequency data and the low-frequency data;

[0014] Pre-training module, suitable for building bidirectional recurrent multi-scale neural networks, and pre-training bidirectional recurrent multi-scale neural networks based on synthetic data;

[0015] The model optimization module is suitable for inputting the collected high-frequency operation data into the pre-trained bidirectional recurrent multi-scale neural network to predict low-frequency prediction data, and inputting the low-frequency prediction data into the pre-trained bidirectional recurrent multi-scale neural network to predict high-frequency prediction data; performing multi-angle loss calculation on the high-frequency prediction data and the high-frequency operation data, constraining the bidirectional recurrent multi-scale neural network to perform model optimization processing, and utilizing the bidirectional recurrent multi-scale neural network after model optimization processing to perform low-frequency extension on the seismic data; wherein, the loss calculation is performed using a multi-angle mixed loss function composed of a time domain loss function, a trace correlation loss function, and a frequency domain loss function.

[0016] According to another aspect of an embodiment of the present invention, there is provided a computing device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0017] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method mentioned above.

[0018] According to another aspect of an embodiment of the present invention, a computer storage medium is provided, wherein the storage medium stores at least one executable instruction, and the executable instruction enables a processor to perform operations corresponding to the above-mentioned multi-scale neural network multi-angle constrained seismic data low-frequency extension method.

[0019] According to another aspect of an embodiment of the present invention, there is provided a computer program product comprising at least one executable instruction, wherein the executable instruction enables a processor to execute operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method as described above.

[0020] According to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method and device provided by the embodiment of the present invention, based on the implicit correlation between high-frequency data and low-frequency data, a bidirectional cyclic multi-scale neural network is used to learn the implicit correlation to ensure that the reconstruction of low-frequency data is more accurate, and the low-frequency features are prevented from being weakened or lost during the learning process. The adaptability and stability of the bidirectional cyclic multi-scale neural network are enhanced by pre-training of the complete spectral characteristics of synthetic data and supplementary training of actual operating data, which makes up for the lack of low-frequency data in the actual operating data. The loss function adopts multi-angle joint constraints, which not only ensures the morphological similarity between the prediction and the operating data, but also further improves the accuracy of low-frequency compensation, so that the network can generate more realistic seismic data on the basis of alignment, solving the problem of insufficient low-frequency data.

[0021] The above description is only an overview of the technical solution of the embodiment of the present invention. In order to more clearly understand the technical means of the embodiment of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the embodiment of the present invention more obvious and easy to understand, the specific implementation method of the embodiment of the present invention is specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the embodiments of the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0023] Figure 1 A flowchart of a multi-scale neural network multi-angle constrained seismic data low-frequency extension method according to an embodiment of the present invention is shown;

[0024] Figure 2 A schematic diagram of the velocity model is shown;

[0025] Figure 3 A schematic diagram of synthetic data is shown;

[0026] Figure 4 A schematic diagram of bidirectional recurrent multi-scale neural network training is shown;

[0027] Figure 5 A schematic diagram of the comparison of low-frequency data predicted based on synthetic data is shown;

[0028] Figure 6 It is a schematic diagram of the convergence curve of the bidirectional recurrent multi-scale neural network label data and prediction results;

[0029] Figure 7 It shows a schematic structural diagram of a multi-scale neural network multi-angle constrained seismic data low-frequency extension device according to an embodiment of the present invention;

[0030] Figure 8 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0031] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0032] Figure 1FIG. 4 is a flowchart of a multi-scale neural network multi-angle constrained seismic data low-frequency extension method according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0033] Step S101, constructing multiple velocity models.

[0034] This embodiment is based on the actual situation of the velocity model of the actual work area. Through geological deformation, 471 different velocity models are generated. Multiple different velocity models are used to subsequently obtain seismic data covering different frequencies, especially low-frequency parts, such as 0-6Hz, so as to obtain sufficient training data for low-frequency extension.

[0035] For each velocity model, forward simulation can be performed to cover the reproduction of complex geological conditions such as deep and shallow water environments and rugged seabed structures, fault structures, and flat layer structures, so that multiple velocity models can be used to obtain synthetic data covering high and low frequencies. Figure 2 As shown, 6 different speed models are included.

[0036] Step S102: construct a broadband wavelet to drive multiple velocity models for forward modeling to obtain synthetic data.

[0037] During forward modeling, the actual data collection conditions of the operation can be maximized based on the setting of acquisition parameters of the actual operation environment, such as a single-line acquisition length of about 24 kilometers, a water depth of 8 kilometers, and an acquisition time of 8 seconds. Based on free boundary conditions that are closer to the actual operation conditions, a wide-band wavelet is constructed, and multiple velocity models are driven by the wide-band wavelet to perform forward simulations, including synthetic data simulations under complex geological conditions such as deep and shallow water environments, faults, rugged seabeds, and flat structures. This ensures that the synthetic data obtained contains both high-frequency data and significant enhancements in low-frequency components, improves the spectral distribution of synthetic data, and provides effective support for the integrity of low-frequency information, so that the low-frequency features can be fully learned, so that the low-frequency part is not ignored or weakened in the subsequent processing and learning process, and also makes the low-frequency recovery ability in complex structural areas better than existing traditional methods. The synthetic data obtained by forward modeling can be as high as 226,080 shot data, which can provide sufficient data samples for pre-training. Synthetic data such as Figure 3 As shown, the left side is synthetic data.

[0038] The synthetic data contains high-frequency data and low-frequency data. Taking acoustic media as an example, in the frequency domain, x s The pressure wave field generated by a point source satisfies the Helmholtz equation as follows:

[0039]

[0040] in, ω is the angular frequency, v(x) is the velocity model, x is the spatial position vector, describing the location of the wave field p, p represents the pressure wave field, p(x,ω) is the wave intensity at the spatial position x and angular frequency ω, s is the source signal, δ is the Dirac function, δ(x s ) represents the position x of the wave source s , which defines that the wave source of the wave equation is a point source, rather than an extended source distributed in space. The equation considers the Born approximation of the perturbation wave field in a homogeneous medium, and the relationship between the wave field data and the mode spectrum can be expressed as:

[0041] δp(s,g,ω)∝δv(K) (2)

[0042] Among them, wave number K = ω / v(s+g), s and g are unit vectors pointing to the source and the receiver respectively. Since most underground modes are usually dominated by horizontal structures, the vertical wave number in the mode spectrum controls the change of the horizontal structure along the vertical direction. Therefore, the relationship between the disturbance wave field and the wave number velocity model can be further simplified as:

[0043] δp(θ,ω)∝δv(K z ) (3)

[0044] Vertical wave number K z It can be expressed as:

[0045]

[0046] Where θ is the opening angle between s and g, i.e. the offset.

[0047] Based on formula (2), if there is a disturbance wave field δp(θ1,ω1), a vertical wave number K is established based on the disturbance wave field. z (θ1, ω1) velocity model. Further simulating another wave field based on this velocity model, such as δp(θ2, ω2), the simulated wave field and the disturbance wave field have the following relationship:

[0048] δp(θ1,ω1)→δv(K z )→δp(θ2,ω2) (5)

[0049] Based on formula (5), it can be seen that there is a correlation between data of different frequencies ω. Therefore, it can be confirmed that there is an implicit correlation between high-frequency data and low-frequency data. For this implicit correlation, deep learning neural network can be used to build learning, so that low-frequency data such as below 6 Hz in actual seismic data can be compensated according to high-frequency data.

[0050] Step S103: construct a bidirectional recurrent multi-scale neural network, and pre-train the bidirectional recurrent multi-scale neural network based on the synthetic data.

[0051] This embodiment uses a bidirectional recurrent multi-scale neural network, which uses convolution kernels of different sizes, multi-size convolution kernels, dilated convolution and multi-column convolution, as well as a bandpass filter at the output end, so that the bidirectional recurrent multi-scale neural network can capture the frequency change characteristics in the seismic data at multiple scales, and improve the feature extraction ability of the neural network. The addition of the bandpass filter can ensure that the output frequency band range is consistent with the spectrum requirements of the required target data.

[0052] The generated synthetic data has complete spectrum information, including both high-frequency data and low-frequency data. The synthetic data is filtered in different frequency bands to obtain high-frequency data (5-14Hz) and low-frequency data (0-6Hz). Figure 3 As shown, the synthesized data is divided into different frequency bands and filtered using different frequency bands, such as band-pass filtering to obtain high-frequency data (5-14 Hz) and low-pass filtering to obtain low-frequency data (0-6 Hz). Figure 3 Below are the spectrum diagrams of synthetic data, high-frequency data, and low-frequency data, with the horizontal axis being the spectrum, the left vertical axis being the frequency, and the right vertical axis being the amplitude.

[0053] By using high-frequency data and low-frequency data as training sample pairs, the bidirectional recurrent multi-scale neural network can be pre-trained so that the bidirectional recurrent multi-scale neural network can learn the implicit correlation between high-frequency data and low-frequency data. The bidirectional cycle of the bidirectional recurrent multi-scale neural network uses high-frequency data to predict low-frequency data, and uses low-frequency data to predict high-frequency data. Therefore, the bidirectional recurrent multi-scale neural network can learn the implicit correlation between high-frequency data and low-frequency data, and can predict low-frequency data based on high-frequency data to complete the low-frequency data in the seismic data.

[0054] Specifically, after the introduction of convolution kernels of different sizes in the bidirectional recurrent multi-scale neural network, effective separation and fusion of high-frequency and low-frequency data can be achieved. Larger convolution kernels can be used to extract global low-frequency features, enabling the network to identify and reconstruct long-wavelength smooth information. Smaller convolution kernels focus more on capturing high-frequency details, thereby enriching the model's ability to express local details. Compared with the single-scale convolution structure, the multi-scale convolution structure is more adaptable to the complexity of seismic data, enabling the neural network to retain rich multi-scale information during the spectrum conversion process. For the bidirectional recurrent neural network, whether in the prediction from high frequency to low frequency or in the reconstruction from low frequency to high frequency, the spectrum distribution can be adaptively adjusted. Considering that different frequency bands of seismic data have different physical meanings, the bidirectional recurrent multi-scale neural network adds a bandpass filter before the network output layer. The bandpass filter can accurately control the output frequency band range to ensure that the data spectrum generated by the bidirectional recurrent multi-scale neural network is consistent with the frequency distribution of the required target data. In the bidirectional cycle process of high frequency predicting low frequency and low frequency predicting high frequency, the frequency band range of the bandpass filter will be adjusted according to different prediction frequency results, so that the output spectrum content perfectly matches the required frequency components, thereby improving the authenticity of the output data. In addition, when pre-training and model optimization of the bidirectional recurrent multi-scale neural network, more attention can be paid to the frequency band of the required target data, thereby improving the convergence speed and prediction accuracy of the bidirectional recurrent multi-scale neural network.

[0055] Furthermore, the multi-column convolution structure of the bidirectional recurrent multi-scale neural network can enhance the expressive power of the neural network. By processing features in parallel in multiple channels, the neural network can learn feature patterns of different scales in different columns and effectively fuse them in the subsequent stage, thereby improving the feature extraction capability at different frequencies and better preserving the correlation between high- and low-frequency features. Compared with the single feature extraction path of the existing convolutional network, the multi-column convolution structure can more comprehensively integrate high- and low-frequency information and ensure the stability and consistency of spectrum conversion. In order to expand the receptive field and capture long-range dependent information in seismic data, the dilated convolution of the bidirectional recurrent multi-scale neural network can effectively increase the feature extraction range while maintaining fewer parameters, thereby better obtaining global information of geological structures. Compared with standard convolution, dilated convolution enables neural networks to obtain contextual information in a larger range, which is particularly suitable for processing structural features with long-range correlations in seismic data.

[0056] Step S104, input the collected high-frequency operation data into the pre-trained bidirectional cyclic multi-scale neural network to predict low-frequency prediction data, and input the low-frequency prediction data into the pre-trained bidirectional cyclic multi-scale neural network to predict high-frequency prediction data; perform multi-angle loss calculation on the high-frequency prediction data and the high-frequency operation data, constrain the bidirectional cyclic multi-scale neural network to perform model optimization processing, and use the bidirectional cyclic multi-scale neural network after model optimization processing to perform low-frequency extension on the seismic data.

[0057] After the pre-training is completed, the collected high-frequency operation data is input into the pre-trained bidirectional cyclic multi-scale neural network to predict the low-frequency prediction data. The generated low-frequency prediction data is then input into the bidirectional cyclic multi-scale neural network again to predict the high-frequency prediction data. The high-frequency prediction data and the high-frequency operation data are subjected to multi-angle loss calculation. Based on the cycle consistency, the bidirectional cyclic multi-scale neural network performs model optimization processing in the mutual conversion between high and low frequencies, so as to effectively improve the effect of low-frequency completion using actual operation data.

[0058] For multi-angle loss calculation, a multi-angle mixed loss function is used. The multi-angle mixed loss function processes the time domain loss function, channel correlation loss function and frequency domain loss function based on the preset balance coefficient, and is calculated using amplitude weighting, as shown below:

[0059] Loss = W(a×L c +b×L d +c×L f )

[0060] Among them, Loss is a multi-angle mixed loss function, W is the amplitude weighting, L c is the channel correlation loss function, L d is the time domain loss function (such as using the mean absolute value loss function in the time domain), L f is the frequency domain loss function (such as using the average absolute value loss of the amplitude in the frequency domain for calculation), a, b, and c are the preset balance coefficients of the channel correlation loss function, the time domain loss function, and the frequency domain loss function, respectively.

[0061] For the trace correlation loss function, the similarity between the target data and the predicted data can be determined according to the variance between the target data and the predicted data of each seismic data, and the trace correlation loss function can be determined according to the similarity between the target data and the predicted data, as shown below:

[0062]

[0063] Among them, d k is the target data (such as high-frequency data or low-frequency data, etc.), is the predicted data (the data predicted by the bidirectional cyclic multi-scale neural network), and k is the kth seismic data.

[0064]

[0065] The above function is used to approximate the cosine similarity Pearson coefficient, that is, the similarity between the target data and the predicted data. s represents the signal source, and cov is the variance of x and y, as shown below:

[0066]

[0067] Among them, μ x and μ y are the means of the data x and y respectively.

[0068] The trace correlation loss function can enhance the correspondence between the traces of the output data of the bidirectional recurrent multi-scale neural network. The core of the trace correlation loss function is to ensure that each trace in the prediction result is aligned with the data of the corresponding trace, thereby enhancing the correlation between single-trace data, so that the low-frequency features generated by the bidirectional recurrent multi-scale neural network are not only consistent with the low-frequency data in the sample as a whole, but also highly matched in the local details of each trace. The trace correlation loss function not only ensures the morphological similarity between the prediction result and the sample data, but also further improves the accuracy of low-frequency compensation, generating more realistic seismic data on the basis of trace-by-trace alignment.

[0069] The frequency domain loss function calculates the frequency curve of the predicted data and the sample data, calculates the error between the average frequencies of the predicted data and the sample data, and constrains the predicted data to align with the sample data in the frequency domain. Amplitude weighting adds higher weights in weak amplitude regions, so that the bidirectional recurrent multi-scale neural network pays more attention to the accuracy of prediction when processing weak amplitude regions; amplitude weighting applies lower weights in strong amplitude regions to reduce overfitting of strong amplitudes, that is, amplitude weighting is inversely proportional to the strength of the amplitude region, so that the bidirectional recurrent multi-scale neural network can more accurately predict subtle features in low amplitude regions, making the results closer to the sample data. The introduction of amplitude weighting plays a key role, especially in the reconstruction of weak signals, and effectively improves the prediction performance of the network in weak amplitude regions.

[0070] When training based on the bidirectional recurrent multi-scale neural network and the multi-angle mixed loss function, for example, 16 data are put into each batch and trained for 100 rounds until the curve obtained by the training result converges, and the trained bidirectional recurrent multi-scale neural network is obtained. Figure 4As shown in the figure, the upper part uses synthetic data (5-14Hz) and predicted data (5-14Hz), that is, the high-frequency data in the synthetic data and the predicted high-frequency data are input into CNN1 (a network in which high-frequency data is predicted by a bidirectional cyclic multi-scale neural network) to obtain predicted data (0-6Hz), that is, the predicted low-frequency data, and the label data (0-6Hz) is the low-frequency data in the synthetic data, and the predicted data (0-6Hz) is input into CNN2 (a network in which low-frequency data is predicted by a bidirectional cyclic multi-scale neural network) to obtain predicted data (5-14Hz). Among them, the synthetic data (5-14Hz) and the predicted data (5-14Hz) are calculated using a multi-angle mixed loss function, and CNN1 and CNN2 are optimized based on the loss calculation result model, and the predicted data (0-6Hz) and the label data (0-6Hz) are calculated using a multi-angle mixed loss function, and CNN1 and CNN2 are optimized based on the loss calculation result model. The lower part of the figure shows the use of actual data (5-14Hz) and predicted data (5-14Hz), that is, high-frequency operation data and predicted high-frequency data are input into CNN1 (a network in which high-frequency data is predicted by a bidirectional cyclic multi-scale neural network) to obtain predicted data (0-6Hz), that is, predicted low-frequency data. The predicted data (0-6Hz) is input into CNN2 (a network in which low-frequency data is predicted by a bidirectional cyclic multi-scale neural network) to obtain predicted data (5-14Hz). Among them, the actual data (5-14Hz) and the predicted data (5-14Hz) are calculated using a multi-angle mixed loss function, and CNN1 is optimized based on the loss calculation result model. The prediction results obtained based on synthetic data are compared. Figure 5 As shown, the input (5-14Hz) is the high-frequency data in the synthetic data, the predicted value (0-6Hz) is the low-frequency data, and the label (0-6Hz) is the low-frequency data in the synthetic data. The bidirectional recurrent multi-scale neural network can be tested, and the predicted output result is consistent with the low-frequency data of the synthetic data. Among them, the horizontal axis is the seismic trace and the vertical axis is time. The above is an example, which is set according to the implementation situation and is not limited here. The curve obtained by the training result converges as shown Figure 6 As shown, the horizontal axis is frequency, the vertical axis is energy, the blue is the label data, that is, the low-frequency data in the synthetic data, and the yellow is the prediction result, that is, the output result of the bidirectional recurrent multi-scale neural network. Based on the convergence and overlap of the two curves, it can be determined whether the bidirectional recurrent multi-scale neural network is trained.

[0071] Furthermore, the trained bidirectional recurrent multi-scale neural network can be used to perform inference and full waveform inversion verification based on synthetic data in order to verify the bidirectional recurrent multi-scale neural network. Based on the verification results, in the time domain, the predicted low-frequency waveform is consistent with the main amplitude and phase characteristics of the real data; in the frequency domain, spectral analysis shows that the predicted results are completely consistent with the frequency band of the sample data.

[0072] Obtaining low-frequency data through prediction for inversion can improve the resolution of the velocity model in deep strata and complex geological structure areas, and more accurately restore the global characteristics and local details of the velocity model, thereby significantly compensating for the information loss problem when only high-frequency data is used for inversion, and providing a reliable foundation for processing such as full waveform inversion of seismic data.

[0073] According to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method provided by the embodiment of the present invention, based on the implicit correlation between high-frequency data and low-frequency data, a bidirectional cyclic multi-scale neural network is used to learn the implicit correlation to ensure that the reconstruction of low-frequency data is more accurate and to avoid the weakening or loss of low-frequency features during the learning process. The adaptability and stability of the bidirectional cyclic multi-scale neural network are enhanced by pre-training of the complete spectral characteristics of synthetic data and supplementary training of actual operating data, which makes up for the lack of low-frequency data in actual operating data. The loss function adopts multi-angle joint constraints, which not only ensures the morphological similarity between the prediction and the operating data, but also further improves the accuracy of low-frequency compensation, so that the network can generate more realistic seismic data on the basis of alignment, solving the problem of insufficient low-frequency data.

[0074] Figure 7 The schematic diagram of the structure of the multi-scale neural network multi-angle constrained seismic data low-frequency extension device provided by the embodiment of the present invention is shown. Figure 7 As shown, the device comprises:

[0075] A velocity model module 710, adapted to construct a plurality of velocity models;

[0076] The synthetic data module 720 is suitable for constructing a broadband wavelet to drive multiple velocity models for forward modeling to obtain synthetic data; the synthetic data includes high-frequency data and low-frequency data, and there is an implicit correlation between the high-frequency data and the low-frequency data;

[0077] A pre-training module 730 is adapted to construct a bidirectional recurrent multi-scale neural network and pre-train the bidirectional recurrent multi-scale neural network based on synthetic data;

[0078] The model optimization module 740 is suitable for inputting the collected high-frequency operation data into a pre-trained bidirectional recurrent multi-scale neural network to predict low-frequency prediction data, and inputting the low-frequency prediction data into a pre-trained bidirectional recurrent multi-scale neural network to predict high-frequency prediction data; performing multi-angle loss calculation on the high-frequency prediction data and the high-frequency operation data, constraining the bidirectional recurrent multi-scale neural network to perform model optimization processing, and utilizing the bidirectional recurrent multi-scale neural network after model optimization processing to perform low-frequency extension on the seismic data; wherein, the loss calculation is performed using a multi-angle mixed loss function consisting of a time domain loss function, a trace correlation loss function, and a frequency domain loss function.

[0079] Optionally, the bidirectional recurrent multi-scale neural network includes multi-size convolution kernels, dilated convolutions and multi-column convolutions, and a bandpass filter at the output end.

[0080] Optionally, the pre-training module 730 is further adapted to:

[0081] The synthetic data is filtered in different frequency bands to obtain high-frequency data and low-frequency data;

[0082] The bidirectional recurrent multi-scale neural network is pre-trained using high-frequency data and low-frequency data so that the bidirectional recurrent multi-scale neural network can learn the implicit correlation between the high-frequency data and the low-frequency data.

[0083] Optionally, the device also includes: a loss function module 750, suitable for constructing a channel correlation loss function; processing the time domain loss function, the channel correlation loss function and the frequency domain loss function based on a preset balance coefficient, and obtaining a multi-angle mixed loss function using amplitude weighted calculation.

[0084] Optionally, the loss function module 750 is further adapted to:

[0085] Determine the similarity between the target data and the predicted data according to the variance between the target data and the predicted data of each channel of seismic data;

[0086] The correlation loss function is determined according to the similarity between the target data and the predicted data.

[0087] Optionally, the amplitude weighting is inversely proportional to the strength of the amplitude region.

[0088] The description of each module above refers to the corresponding description in the method embodiment and will not be repeated here.

[0089] An embodiment of the present invention also provides a non-volatile computer storage medium, which stores at least one executable instruction, and the executable instruction can execute operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method in any of the above method embodiments.

[0090] An embodiment of the present application provides a computer program product, which includes at least one executable instruction or computer program, which enables a processor to perform operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method in any of the above-mentioned method embodiments.

[0091] Figure 8 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown, and the specific implementation of the embodiment of the present invention does not limit the specific implementation of the computing device.

[0092] like Figure 8 As shown, the computing device may include: a processor 802 , a communication interface 804 , a memory 806 , and a communication bus 808 .

[0093] in:

[0094] The processor 802 , the communication interface 804 , and the memory 806 communicate with each other via a communication bus 808 .

[0095] The communication interface 804 is used to communicate with other devices such as clients or other servers.

[0096] The processor 802 is used to execute the program 810, and specifically can execute the relevant steps in the above-mentioned multi-scale neural network multi-angle constrained seismic data low-frequency extension method embodiment.

[0097] Specifically, the program 810 may include program codes, which include computer operation instructions.

[0098] The processor 802 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. The one or more processors included in the computing device may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0099] The memory 806 is used to store the program 810. The memory 806 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0100] Program 810 can be specifically used to enable processor 802 to execute the multi-scale neural network multi-angle constrained seismic data low-frequency extension method in any of the above-mentioned method embodiments. The specific implementation of each step in program 810 can refer to the corresponding descriptions in the corresponding steps and units in the above-mentioned multi-scale neural network multi-angle constrained seismic data low-frequency extension embodiment, which will not be repeated here. Technical personnel in the relevant field can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described equipment and modules can refer to the corresponding process description in the aforementioned method embodiment, which will not be repeated here.

[0101] The algorithm or display provided herein is not inherently related to any particular computer, virtual system or other device. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing such systems. In addition, the embodiment of the present invention is not directed to any specific programming language either. It should be understood that various programming languages ​​can be utilized to realize the content of the embodiment of the present invention described herein, and the description of the above specific language is to disclose the preferred implementation of the embodiment of the present invention.

[0102] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0103] Similarly, it should be understood that in order to streamline the embodiments of the present invention and to aid in understanding one or more of the various inventive aspects, in the above description of the exemplary embodiments of the present invention, the various features of the embodiments of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, the disclosed method should not be interpreted as reflecting the following intention: that the claimed embodiments of the present invention require more features than the features explicitly recited in each claim. More specifically, as reflected in the claims below, the inventive aspects lie in less than all the features of the single embodiment disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, wherein each claim itself serves as a separate embodiment of the present invention.

[0104] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0105] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0106] The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The embodiments of the present invention may also be implemented as a device or apparatus program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing an embodiment of the present invention may be stored on a computer-readable medium, or may have the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0107] It should be noted that the above embodiments illustrate rather than limit the present invention, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim listing a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be understood as limitations on the order of execution.

Claims

1. A multi-scale neural network multi-angle constrained seismic data low-frequency extension method, characterized in that: Methods include: Construct multiple velocity models; Constructing a broadband wavelet to drive the plurality of velocity models for forward modeling to obtain synthetic data; The synthetic data includes high-frequency data and low-frequency data, and there is an implicit association relationship between the high-frequency data and the low-frequency data; Constructing a bidirectional recurrent multi-scale neural network, and pre-training the bidirectional recurrent multi-scale neural network based on the synthetic data; The collected high-frequency operation data is input into a pre-trained bidirectional recurrent multi-scale neural network to predict low-frequency prediction data, and the low-frequency prediction data is input into a pre-trained bidirectional recurrent multi-scale neural network to predict high-frequency prediction data; multi-angle loss calculation is performed on the high-frequency prediction data and the high-frequency operation data, and the bidirectional recurrent multi-scale neural network is constrained to perform model optimization processing, so as to use the bidirectional recurrent multi-scale neural network after the model optimization processing to perform low-frequency extension on the seismic data; wherein, the loss calculation is performed using a multi-angle mixed loss function composed of a time domain loss function, a track correlation loss function and a frequency domain loss function.

2. The method according to claim 1, characterized in that The bidirectional recurrent multi-scale neural network includes multi-size convolution kernels, dilated convolutions and multi-column convolutions, as well as a bandpass filter at the output end.

3. The method according to claim 1, characterized in that The pre-training of the bidirectional recurrent multi-scale neural network based on the synthetic data further comprises: The synthesized data is subjected to filtering processing in different frequency bands to obtain high-frequency data and low-frequency data; The bidirectional recurrent multi-scale neural network is pre-trained using the high-frequency data and the low-frequency data, so that the bidirectional recurrent multi-scale neural network learns an implicit association relationship between the high-frequency data and the low-frequency data.

4. The method according to claim 1, characterized in that: The method further comprises: Construct the channel correlation loss function; The time domain loss function, channel correlation loss function and frequency domain loss function are processed based on the preset balance coefficient, and the multi-angle hybrid loss function is obtained by amplitude weighted calculation.

5. The method according to claim 4, characterized in that The constructing of the channel correlation loss function further comprises: Determine the similarity between the target data and the predicted data according to the variance between the target data and the predicted data of each channel of seismic data; The correlation loss function is determined according to the similarity between the target data and the predicted data.

6. The method according to claim 4, characterized in that The amplitude weighting is inversely proportional to the strength of the amplitude region.

7. A multi-scale neural network multi-angle constrained seismic data low-frequency extension device, characterized in that: The device includes: Velocity model module, suitable for building multiple velocity models; A synthetic data module is suitable for constructing a broadband wavelet to drive the plurality of velocity models for forward modeling to obtain synthetic data; the synthetic data includes high-frequency data and low-frequency data, and there is an implicit correlation between the high-frequency data and the low-frequency data; A pre-training module, adapted to construct a bidirectional recurrent multi-scale neural network, and pre-train the bidirectional recurrent multi-scale neural network based on the synthetic data; The model optimization module is suitable for inputting the collected high-frequency operation data into a pre-trained bidirectional recurrent multi-scale neural network to predict low-frequency prediction data, and inputting the low-frequency prediction data into a pre-trained bidirectional recurrent multi-scale neural network to predict high-frequency prediction data; performing multi-angle loss calculation on the high-frequency prediction data and the high-frequency operation data, constraining the bidirectional recurrent multi-scale neural network to perform model optimization processing, so as to use the bidirectional recurrent multi-scale neural network after the model optimization processing to perform low-frequency extension on the seismic data; wherein, the loss calculation is performed using a multi-angle mixed loss function composed of a time domain loss function, a track correlation loss function and a frequency domain loss function.

8. A computing device, characterized in that include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method as described in any one of claims 1-6.

9. A computer storage medium, characterized in that The storage medium stores at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method as described in any one of claims 1-6.

10. A computer program product, characterized in that It includes at least one executable instruction, which enables the processor to execute operations corresponding to the multi-scale neural network multi-angle constrained seismic data low-frequency extension method as described in any one of claims 1-6.

Images