Formation-constrained well control GRAN seismic signal broadband reconstruction method

Abstract

The invention provides a formation-constrained well control GRAN seismic signal broadband reconstruction method, belongs to the technical field of geophysical exploration, and solves the problem of seismic signal resolution in an oil-gas exploration process. Comprising the following steps: S1, acquiring well location data, and generating and expanding a tag training data set by combining a sequential Gaussian simulation method according to target reservoir features; s2, constructing a GRAN model; training a data set by using the expanded label, training the nonlinear mapping capability of the model in the process of converting the seismic signal from a narrow band to a wide band, and performing model testing by using a distributed external sample after training is completed; s3, using the GRAN model passing the test to perform a conversion process from a narrowband to a broadband on an actual seismic signal to obtain a seismic signal with high resolution; according to the method, seismic narrow-band signals are converted into broadband signals, the signal-to-noise ratio of the seismic signals can be kept while the seismic signal resolution is improved, and meanwhile the model accuracy and generalization ability are guaranteed.

Description

Technical Field

[0001] The invention belongs to the technical field of geophysical exploration, is applied to the high-resolution processing of seismic signals, and is specifically a formation-constrained well-controlled GRAN seismic signal broadband reconstruction method. Background Art

[0002] The detection of high-resolution seismic signals can make methods such as well-seismic calibration, structural interpretation, attribute analysis, and lithology inversion more accurate in identifying favorable reservoir areas; thin-layer resolution is the key to the resolution problem, which is determined by the main frequency and bandwidth of the seismic record, and is usually quantitatively analyzed by wavelet energy distribution and vibration length. The amplitude and bandwidth of seismic signals change non-stationarily with time, and their non-stationarity is mainly caused by wavefront divergence, frequency-related attenuation and scattering; geometric expansion compensation and time-varying spectrum whitening algorithms are usually used for correction.

[0003] In the prior art, the proposed technical solutions for the related spectrum problem are: decomposing the seismic trace with overlapping windows and correcting the attenuation of each window to widen the spectrum; the adaptive spectrum widening method uses a variable window size to divide the seismic trace into different time periods, and assumes that the waveform is stable in each section, and then performs spectrum widening and amplitude compensation to create a seismic trace with a wider spectrum. At present, the methods for improving the resolution of seismic signals mainly include: zero-phase deconvolution, predictive deconvolution, inverse Q filtering, spectrum simulation deconvolution, well-constrained frequency enhancement, frequency division deconvolution and absorption attenuation compensation. These methods have played an important role in various links of actual seismic data processing and interpretation, and have achieved good application results. However, this type of traditional processing method for improving the resolution of seismic signals, when focusing on improving the bandwidth, faces the digital contradiction between the signal-to-noise ratio and the resolution, which brings a bottleneck problem that is difficult to break through.

[0004] Seismic signal resolution has always been a hot topic in seismic exploration research. Different criteria for distinguishing between resolvable and indistinguishable signals and definitions of vertical resolution limits can be given from different observation angles. For example, the Rayleigh criterion defines the vertical resolution limit as λ / 4 (λ is the seismic wavelength), and there are also cases where vertical resolution is defined as λ / 4.6, λ / 8, and λ / 12. As exploration objects become more and more complex, such as the exploration and development of lithologic oil and gas reservoirs and unconventional oil and gas, high-resolution algorithms are required to not only distinguish thin layers, but also provide richer waveform and phase information. Therefore, it is necessary to completely get rid of the constraints of traditional ideas in theoretical methods and realize broadband reconstruction of seismic signals.

[0005] In recent years, with the breakthroughs in computing power and algorithm theory of deep learning, a branch of machine learning, deep learning technology has developed rapidly. Its characteristics are complex neural network architectures with many "hidden" computing layers, which enable the network to learn and establish highly nonlinear relationships between input variables and output predicted variables. Deep learning technology has been proven to be good at image processing, especially after the introduction of generative adversarial networks (GANs), its application scope has become more extensive. The characteristic of GANs is that they can generate real high-quality images by inputting a noisy or low-quality image. The existing technology has gradually increased the application of GANs in the field of geophysics, such as: using GANs for seismic velocity inversion modeling; applying generator networks to seismic full waveform inversion to reduce the number of model parameters; using spatial generative adversarial networks to achieve geostatistical probability inversion; using generative adversarial networks to increase the sampling density of actual seismic data and improve the accuracy of fault detection results based on deep learning; using low-pass filters to filter broadband seismic signals to obtain high- and low-frequency seismic data for adversarial training. These applications fully verify the effectiveness of GANs in improving the resolution of seismic signals.

[0006] However, although GAN has the potential to greatly improve the resolution and interpretability of seismic signals, the results obtained are still subject to various questions; such as the physical limits of seismic data resolution, the computational overhead of obtaining high-resolution images, and the challenges of coherent and incoherent noise related to processing and imaging. Today, the GAN method in deep learning can provide a completely data-driven approach, and continued research is expected to overcome these challenges to a certain extent. Through a trained deep learning network, we can try to infer the relationship between seismic waveforms and small-scale anomalies to improve the quality of low-resolution or distorted seismic images. Therefore, the research of this technical solution has become the focus of technical personnel in this field; a trained deep learning network can be migrated to new similar work areas to help geophysical and geological experts obtain more accurate understanding from higher-quality seismic images. Summary of the invention

[0007] Under the background technology, the purpose of the present invention is to achieve the broadband reconstruction effect of seismic signals through the formation-constrained well-controlled GRAN; under the idea of ​​the present invention, training data of the same position of the seismic data body can be provided, and these data are composed of low-resolution and high-resolution seismic data pairs, so as to train the neural network model. Due to the limited number of wells drilled and the serious shortage of training samples, it is also necessary to combine the geological conditions of the actual work area, based on the fine well-seismic calibration, and use the logging data to obtain the correspondence between the seismic trace and the reflection coefficient near the well; then through sequential Gaussian random simulation, a large number of reflection coefficient sequences are simulated, and then a three-parameter broadband wavelet is used. Based on the convolution model, a large number of narrowband and broadband seismic signal pairs can be artificially constructed for the training of the cyclic generative adversarial network (GRAN) model; the trained model can be used for broadband reconstruction of narrowband seismic signals.

[0008] The present invention adopts the following technical solutions to achieve the purpose:

[0009] A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method, the method comprising the following steps:

[0010] S1. Obtain well location data, and generate and expand a label training data set with similar statistical significance based on the rock physics analysis method and the forward simulation label construction method, combined with the sequential Gaussian simulation method, for the target reservoir characteristics in the well location data;

[0011] S2. Construct a recurrent generative adversarial network model, namely the GRAN model; use the expanded labeled training data set to train the GRAN model's nonlinear mapping ability for the seismic signal conversion process from narrowband to broadband; after training, use samples outside the distribution of the labeled training data set to test the GRAN model;

[0012] S3. Using the trained and tested GRAN model, the actual seismic signal is converted from narrowband to broadband to obtain a seismic signal with high resolution.

[0013] Specifically, in step S1, when obtaining well location data based on well logging data, a well seismic calibration process with high accuracy is performed; in this process, a preliminary label training data set is generated, including narrow-band synthetic seismic records in the well logging data.

[0014] Specifically, during the well seismic calibration process, when extracting wellside tracks along the well trajectory, the weighted average of several wellside tracks is adopted; for the three-dimensional data volume, the wellside seismic records are extracted along different azimuths to extract labels.

[0015] Furthermore, after completing the well-seismic calibration process, since the training sample set is constructed through well logging data, the well logging data of the same work area are standardized using an environmental correction method, so that the reflection coefficient sequences constructed by all well logging data are in the same scale range.

[0016] Furthermore, the rationality of the environmental correction is judged by the synthetic seismic records near the well, and the convolution model is used to synthesize the seismic records. During the correction process, the matching degree between the mudstone top position and the seismic records is compared and analyzed through records synthesized at different frequencies to determine the frequency band range of the optimal broadband signal, that is, the broadband synthetic seismic records in the well logging data. The broadband wavelet of the broadband synthetic seismic record is used as the high-frequency component of the label constructed, corresponding to the low-frequency component of the label constructed in the narrow-band wavelet synthetic seismic record, as the preliminary label training data set.

[0017] Specifically, after the frequency band range of the broadband signal is determined, the reflection coefficient sequence of the known logging data is simulated by a sequential Gaussian simulation method to obtain a new reflection coefficient sequence.

[0018] Furthermore, in the new reflection coefficient sequence, the narrowband wavelet synthetic seismic record extracted near the well is used as the low-frequency component of the label, and the wavelet synthetic seismic record with the same frequency band as the determined broadband signal is used as the high-frequency component of the label, thereby expanding the label training data set.

[0019] Specifically, in step S2, the architecture of the GRAN model is improved based on the GAN model; the generator of the GRAN model is composed of a simplified U-net multi-stage series connection, all with the same loop structure; the discriminator of the GRAN model is implemented by introducing the Wasserstein distance to construct the corresponding loss function.

[0020] Specifically, the GRAN model uses a global variable time step T to control the size of the generator; in a single generator structure, it includes an encoder g(·) part and a decoder f(·) part, and the structures of the encoder and the decoder are symmetrical.

[0021] Furthermore, the narrowband seismic signal of the logging data in the labeled training data set is input into the generator of the GRAN model. The generator structure starts from the decoder. At each time t, the corresponding encoder generates a sample, and the outputs of all times are superimposed as the final output.

[0022] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows:

[0023] As a new method for realizing broadband reconstruction of seismic signals, the present invention can be used as a multivariate nonlinear mapping model for converting seismic signals from narrowband signals to broadband signals with the help of deep learning technology.

[0024] In the broadband reconstruction process of the seismic signal of the present invention, the signal-to-noise ratio of the seismic signal can be maintained while the resolution of the seismic signal is greatly improved, thus successfully solving the difficult problem in the prior art.

[0025] In the process of training the recurrent generative adversarial network model, the present invention systematically tests the robustness of the training network on samples outside the distribution of the training data set, which can ensure the accuracy and generalization ability of the reconstruction model. Through the trained GRAN model, broadband reconstruction of seismic signals is achieved, and finally the resolution of seismic signals is successfully improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;

[0027] Figure 2 It is a schematic diagram of the framework composition and process principle of the method of the present invention;

[0028] Figure 3 This is a schematic diagram of the structure of the traditional GAN ​​network;

[0029] Figure 4 It is a structural schematic diagram of the GRAN model generator in the present invention;

[0030] Figure 5 It is a schematic diagram of the overall structure of the GRAN model in the present invention;

[0031] Figure 6 Schematic diagram of extraction of two situations in the process of well seismic calibration of the present invention;

[0032] Figure 7 A schematic diagram showing the comparison of synthetic seismic records during the environmental correction process of the present invention;

[0033] Figure 8 is a schematic diagram of a frequency band range preferably used to construct a broadband sample pair in the present invention;

[0034] Fig. 9 This is a schematic diagram of the fine calibration of the reflection characteristics of the seismic profile before and after the broadband reconstruction of a certain well;

[0035] Fig.10 This is a schematic diagram showing the comparison of seismic sections before and after broadband reconstruction;

[0036] Fig.11 This is a schematic diagram of the comparison of seismic spectrum profiles before and after broadband reconstruction;

[0037] Fig.12 A schematic diagram showing the comparison of time slices before and after broadband reconstruction of the GRAN model;

[0038] Fig.13A schematic diagram showing the comparison of the ant body cross-section before and after broadband reconstruction;

[0039] Fig.14 Schematic diagram of the comparison of the Ant Body stratigraphic slices before and after broadband reconstruction. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0041] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] Example 1

[0043] A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method. The overall process of this method can be found in Figure 1 For example, Figure 2 At the same time, an overview of its framework composition and process principles is given. The steps of this method are summarized as follows:

[0044] S1. Obtain well location data, and generate and expand a label training data set with similar statistical significance based on the rock physics analysis method and the forward simulation label construction method, combined with the sequential Gaussian simulation method, for the target reservoir characteristics in the well location data;

[0045] S2. Construct a recurrent generative adversarial network model, namely the GRAN model; use the expanded labeled training data set to train the GRAN model's nonlinear mapping ability for the narrowband to broadband conversion process of seismic signals; after training, systematically test the robustness and generalization ability of the GRAN model on samples outside the distribution of the labeled training data set;

[0046] S3. Using the trained and tested GRAN model, the actual seismic signal is converted from narrowband to broadband to obtain a seismic signal with high resolution.

[0047] This embodiment will explain the details of the above method and its GRAN model in detail. First, the deep learning neural network model used in this embodiment, that is, the recurrent generative adversarial network model (GRAN model), is an improved network architecture based on the GAN model. Figure 3 A schematic diagram of the network structure of the traditional GAN ​​model is given.

[0048] In this embodiment, the generator of the GRAN model is composed of a simplified U-net multi-stage series connection, and the discriminator is implemented by introducing the Wasserstein distance to construct the corresponding loss function to ensure that there is no problem of gradient disappearance at any time. When this embodiment uses the convolution model of the prior art to synthesize seismic records, it is necessary to strictly ensure the environmental correction of the logging data and the accurate well-seismic calibration to obtain an accurate reflection coefficient sequence, and to preferably determine the frequency band range of the broadband signal through well-seismic comparative analysis. For the data set used for training, although it is synthesized and expanded manually, the effect in actual application is also satisfactory enough.

[0049] The following is a detailed introduction to the label training data set part of this embodiment.

[0050] The GRAN model used in this embodiment can be found in Figure 5 As a supervised artificial neural network, its prediction accuracy depends largely on the complexity of the structure and the amount of sample data available for training. Unlike other fields where artificial neural networks have achieved great success (such as image classification, etc.), where labeled data is relatively easy to obtain, the availability of actual labeled data in the field of exploration geophysics is often limited.

[0051] In the process of model training, training data is very important: first, there must be as many training samples as possible, and second, the distribution of samples must be uniform enough, that is, the number of samples in each category must be sufficient. Therefore, for the extraction of samples in this embodiment, relying on well location data, fine well seismic calibration is required as a basis. Figure 6 (a) is a schematic diagram of extracting wellside channels along the well trajectory. In order to improve the signal-to-noise ratio, this embodiment adopts a weighted average method for several wellside channels; Figure 6 (b) On the three-dimensional data volume, seismic records near the well are extracted along different azimuths. If broadband reconstruction is to be performed on the data volume of the stacked gathers at different azimuths, this method of extracting labels by azimuth can be considered.

[0052] In the actual data collection process, insufficient sample size often occurs, which will lead to overfitting of the trained model and insufficient generalization ability. Therefore, it is necessary to increase the number of samples through sample optimization and expansion based on relevant sample enhancement techniques.

[0053] In the seismic data of this embodiment, the number of wells drilled is limited, so the problem of big data and small samples is faced. How to use limited drilling and logging information to construct a label training data set with uniform distribution and diverse categories is the key to the training results of the GRAN model. However, the geological structure and sedimentary environment vary widely, and it is difficult to construct an all-encompassing label training data set under the realistic conditions faced by this embodiment. Therefore, when constructing the sample, it is necessary to understand the sedimentary environment and use the model of the target reservoir as a constraint to reduce the complexity of the sample set. In this embodiment, since the training sample set is constructed by logging data, the logging data in the same work area is standardized by environmental correction to ensure that the reflection coefficient sequence constructed by the logging data is within the same scale range. Of course, the specific environmental correction operation is well known to those skilled in the art as a method of standardization processing and will not be repeated; but the comparison judgment and frequency band selection in this process will be introduced in detail below.

[0054] In this embodiment, the rationality of the environmental correction is mainly determined by the effect of the synthetic seismic record near the well. Figure 7 (a) is the synthetic seismic record before environmental correction. The upper part is the seismic record synthesized by the main frequency 25Hz Ricker wavelet, and the lower part is the seismic record synthesized by the main frequency 5-60Hz broadband wavelet. Judging from the synthetic seismic record of 25Hz Ricker wavelet alone, it matches well with the well side channel. However, looking at the synthetic seismic record of 5-60Hz broadband wavelet, the position of the mudstone top does not match the seismic record, which is mainly caused by environmental factors, such as well wall collapse.

[0055] Figure 7 (b) is a synthetic seismic record after environmental correction, the upper part is a seismic record synthesized by a Ricker wavelet with a main frequency of 25 Hz, and the lower part is a seismic record synthesized by a broadband wavelet with a main frequency of 5-60 Hz; it can be seen that the synthetic record with a main frequency of 25 Hz is not much different from that before correction, but the synthetic seismic record with a broadband wavelet of 5-60 Hz has a higher match with the mudstone top than that before correction. Therefore, if you want to artificially construct reasonable high- and low-frequency sample pairs and expand the label training data set, then the preprocessing of logging data and fine well-seismic calibration are essential basic work. In addition, how to select a suitable frequency band and perform broadband reconstruction of seismic data to achieve a detailed description of geological structures and thin reservoirs is also a problem that will be considered in this embodiment. The solution to this problem is introduced below.

[0056] In this embodiment, by using synthetic seismic records with different main frequencies and combining well logging lithology interpretation results or logging information, the frequency band range that can most accurately describe the changes in thin reservoirs, that is, the frequency band range of broadband signals, can be selected. Figure 8 It is demonstrated how to optimize the frequency band range for constructing broadband sample pairs, which can be viewed simultaneously. Figure 8In the analysis, using 5-60 Hz broadband wavelet can well describe the reflection of the mudstone top, making the description of thin layers in the broadband reconstructed seismic records closer to the actual drilling results; increasing the bandwidth on this basis does not improve much.

[0057] Therefore, in order to ensure the fidelity of the reconstructed signal, this embodiment adopts the principle of choosing a lower value rather than a higher value, and preferably uses a 5-60 Hz broadband wavelet as the broadband wavelet for constructing the label.

[0058] In the actual process of label set construction, if only the information of the drilled wells is used, there will be too few known samples. Therefore, this embodiment combines the characteristics of the target reservoir, limits the maximum and minimum layer thicknesses, and uses the reflection coefficient sequence of the known wells to simulate a new reflection coefficient sequence based on sequential Gaussian simulation; the sequential Gaussian simulation technology is well known to those skilled in the art and will not be described in detail. After simulating the new reflection coefficient sequence, the wavelet synthesized seismic record extracted from the well is used as the low-frequency component of the label, and the 5-60Hz broadband wavelet synthesized seismic record selected in this embodiment is used as the high-frequency component. In this way, a large number of training labels can be artificially constructed to expand the label training data set, and used as input training data for the GRAN model.

[0059] The above content is a detailed introduction to the label training data set part of this embodiment; the following content is a detailed introduction to the GRAN part of the cyclic generative adversarial network model in this embodiment.

[0060] First, it is necessary to briefly describe the relevant content of the traditional GAN ​​model. The structure of the GAN model is inspired by game theory. Based on the principle that the sum of the two is zero and what one party gains is what the other party loses, it constructs two opposing parts, namely the generator and the discriminator. Figure 3 .

[0061] The generator is used to capture the potential distribution of real data and generate new samples; the discriminator is used to determine whether the input data comes from real data or generated samples. The minimax optimization principle is adopted to make both the generator and the discriminator reach Nash equilibrium. Figure 5 You can also refer to understand the GAN model, where G and D represent the generator and discriminator respectively.

[0062] The input of G is a random variable (or a low-quality signal, which can be a narrowband seismic signal in this embodiment) z; the input of D is mixed data, including real data marked as 1 and generated data marked as 0. Z(z) represents the sample generated by G. When the mixed data is input into D, D performs a binary classification on the data. If D determines that the current data is real data P data , the output is closer to 1; if it is judged to be generated data P g, the output is closer to 0. In order to make the data generated by G deceive D as much as possible, the generator and the discriminator need to be trained against each other so that their respective capabilities can be continuously improved. The GAN model is trained using a maximum and minimum principle, and its objective function is shown in the following formula:

[0063]

[0064] By training D to maximize the objective function, we can ensure that the data is judged as accurately as possible. At the same time, by training G to minimize log(1-D(G(z))), we can train the entire GAN model and improve application performance.

[0065] In this embodiment, the GRAN model used improves the generator of the GAN model and uses the same loop structure as the generator. Figure 4 In the GRAN model, the size of the generator is controlled by the global variable time step T. In a single generator structure, it contains two parts, g(·) and f(·). f(·) is equivalent to the decoder, and g(·) is equivalent to the encoder, and the two are symmetrical structures. The generator of the GRAN model is different from other loop structures. The model starts from the decoder f(·) part. At each time t, the corresponding encoder obtains a generated sample, and finally the outputs of all times are superimposed as the final output.

[0066] In this embodiment, z represents a low-quality signal (such as a narrowband seismic signal), t represents the current time step, and h represents the current time step. c,t Indicates encoder output, h z,t That is, the result of tanh calculation for low-quality signals is scaled to [1,1], ΔC (t) represents the output of the decoder, and Θ represents the final output of the generator. Therefore, the specific calculation process of the generator is as follows:

[0067] z t =P(z)

[0068] h c,t =g(ΔC t-1 )

[0069] h z,t =tanh(Wz t +b)

[0070] ΔC t =f([h c,t ,h z,t ])

[0071]

[0072] Among them, [h c,t,h z,t ] is h c,t and h z,t cascade, σ represents the activation function sigmoid, and compresses the final output value to [0,1].

[0073] In this embodiment, the objective function of the GRAN model is the same as the original GAN ​​model, which is used to guide the training of the entire model. When the discriminator is in the optimal state, the objective function of the generator in the GAN model and the GRAN model is equivalent to measuring the JS divergence between the generated distribution and the real data distribution. However, the JS divergence cannot help the generator generate high-quality samples, because in high-dimensional space, if two high-dimensional distributions do not intersect, or the intersecting parts can be ignored, then the distance measured by the JS divergence can neither reflect the distance between the two high-dimensional distributions nor provide a gradient, but the Wasserstein distance can provide a meaningful gradient.

[0074] In this embodiment, Wasserstein distance is used to guide the training of the entire GRAN model. When the loss function of the GRAN model is constructed using Wasserstein distance, it is to approximate the similarity between two distributions, which belongs to the regression task, so the sigmoid function is not needed. Therefore, the true distribution P data and generate distribution P g The Wasserstein distance between them is:

[0075]

[0076] In order to generate the distribution P g As close as possible to the true distribution P data , we need to make D in the formula w As small as possible. Since the Wasserstein distance is differentiable at any time, there is no problem of gradient disappearance, so the objective function of the discriminator can be obtained:

[0077]

[0078] In addition, the generator and discriminator have obj G =-obj D The relationship between the two is that the first term in the Wasserstein distance formula has nothing to do with the generator, so the objective function of the generator can be obtained:

[0079]

[0080] The GRAN model with the above structural features can complete the conversion process of actual seismic signals from narrowband to broadband after being trained with an expanded labeled training data set and further tested, and will be able to obtain seismic signals with high resolution in subsequent applications.

[0081] Example 2

[0082] This embodiment is based on the embodiment 1 and is combined with the display of the actual scene in the accompanying drawings to more clearly present the actual application and effect of the present invention.

[0083] Through the broadband reconstruction method of seismic signals in Example 1, the seismic signal frequency band is effectively broadened under the premise of preserving the fidelity of the original signal, so as to effectively improve the resolution of seismic data while ensuring that the signal-to-noise ratio of the seismic signal remains unchanged. This example is suitable for the identification of small interlayer faults of level 3 and 4 with a scale of 10-50m in the study area.

[0084] In this embodiment, in order to ensure the validity and fidelity of broadband reconstructed data, quality control is performed from several aspects such as points, lines, and surfaces. Fig. 9 It is a detailed calibration diagram and related data of the reflection characteristics of the seismic profile before and after the broadband reconstruction of a certain well; among them, the No. 4 layer is the top of the mudstone, which has the characteristics of low density and low velocity, and its top should correspond to the reflection coefficient and trough. In the original data, the top interface of the mudstone is a zero point and is difficult to track. Fig. 9 (b) The corresponding broadband data mudstone top interface is a stably distributed trough, which is consistent with the well geophysical response and can be stably tracked in the entire area. Therefore, from the well point of view, the broadband signal has a better response to thin layer reflection, which means that the 4 small layers of reflection appearing in the broadband signal have actual physical significance and are true and reliable.

[0085] The role of broadband reconstruction data in improving thin layer resolution and identifying small faults is further explained online. Fig.10 From the broadband seismic profile (b), the broadband seismic data has a higher resolution than the original seismic signal in the response of the No. 4 sublayer and the small fault. Fig.10 (a). Fig.11 is the comparison of seismic spectrum profiles before and after broadband reconstruction. Fig.11 From the original spectrum profile of (a), we can see that the spectrum characteristics vary greatly horizontally, and the bandwidth is between 9-50Hz; Fig.11 From the broadband spectrum profile of (b), the lateral changes of the spectrum characteristics are relatively consistent, and the frequency band is widened to 5-65Hz.

[0086] Fig.12 The time slices (T2100ms) before and after the wideband reconstruction of the GRAN model are shown. Fig.12 (a) is the isochronous slice of the original seismic data. Fig.12 (b) is the isochronous slice of broadband seismic data. From the comparison of the slices, we can see that Fig.12 The resolution of (b) is significantly improved, and the overall pattern of the slices is consistent.

[0087] Fig.13 The comparison of the ant body cross-section before and after broadband reconstruction is shown. Fig.13 (a) is the cross section of the original data ant body, and there is no fracture at the point where the arrow is; Fig.13 (b) is the profile of the ant body of broadband data, with a fracture line at the arrow point; compared with the original data, the ant body tracked by broadband data matches the actual drilling breakpoint better. Such small-scale fractures are the main cause of geological engineering events such as compression stringing. In this sense, broadband data can better find the design of well trajectories in shale oil and gas development and avoid the occurrence of geological engineering events such as compression stringing.

[0088] Fig.14 It shows the stratigraphic slices of the Ant Body before and after broadband reconstruction. Fig.14 (a) is the original data ant body. The circled position in the figure is the location where the fracturing event occurred during the actual drilling, but there is no crack display on the ant body slice of the original seismic data; Fig.14 (b) is the broadband data ant body. The circled position in the figure is the location where the fracturing event occurred during actual drilling. From the ant body slice of the broadband seismic data, obvious cracks can be seen. It can be determined that the effects of improving the resolution of seismic signals brought by broadband data are of great guiding significance for guiding the design of horizontal well trajectories and avoiding geological engineering events such as fracturing.

Claims

1. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method, characterized in that: The method comprises the following steps: S1. Obtain well location data, and generate and expand a label training data set with similar statistical significance based on the rock physics analysis method and the forward simulation label construction method, combined with the sequential Gaussian simulation method, for the target reservoir characteristics in the well location data; S2, build a cyclic generative adversarial network model, namely the GRAN model; Use the expanded labeled training data set to train the GRAN model's nonlinear mapping capability for the seismic signal conversion process from narrowband to broadband. After training, use samples outside the distribution of the labeled training data set to test the GRAN model. S3. Using the trained and tested GRAN model, the actual seismic signal is converted from narrowband to broadband to obtain a seismic signal with high resolution.

2. The formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 1, characterized in that: In step S1, when obtaining well location data based on well logging data, a well seismic calibration process with high accuracy is performed; in this process, a preliminary label training data set is generated, including narrow-band synthetic seismic records in the well logging data.

3. The formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 2, characterized in that: During the well seismic calibration process, when extracting wellside tracks along the well trajectory, the weighted average of several wellside tracks is used for processing; for three-dimensional data volumes, the wellside seismic records are extracted along different azimuths to extract labels.

4. The formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 2, characterized in that: After completing the well-seismic calibration process, the logging data in the same work area are standardized using the environmental correction method, so that the reflection coefficient sequences constructed by all logging data are in the same scale range.

5. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 4, characterized in that: The rationality of environmental correction is judged by the synthetic seismic records near the well, and the convolution model is used to synthesize the seismic records. During the correction process, the matching degree between the mudstone top position and the seismic records is compared and analyzed through records synthesized at different frequencies to determine the frequency band range of the optimal broadband signal, that is, the broadband synthetic seismic records in the well logging data. The broadband wavelet of the broadband synthetic seismic record is used as the high-frequency component of the constructed label, corresponding to the low-frequency component of the constructed label in the narrow-band wavelet synthetic seismic record, as the preliminary label training data set.

6. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 5, characterized in that: After the frequency band of the broadband signal is determined, the reflection coefficient sequence of the known logging data is simulated by the sequential Gaussian simulation method to simulate a new reflection coefficient sequence.

7. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 6, characterized in that: In the new reflection coefficient sequence, the narrowband wavelet synthetic seismic record extracted near the well is used as the low-frequency component of the label, and the wavelet synthetic seismic record with the same frequency band as the determined broadband signal is used as the high-frequency component of the label to expand the label training data set.

8. The formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 1, characterized in that: In step S2, the architecture of the GRAN model is improved based on the GAN model; the generator of the GRAN model is composed of a simplified U-net multi-stage series connection, all with the same loop structure; the discriminator of the GRAN model is implemented by introducing the Wasserstein distance to construct the corresponding loss function.

9. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 8, characterized in that: The GRAN model uses a global variable time step T to control the size of the generator; in a single generator structure, it includes an encoder and a decoder, and the structures of the encoder and the decoder are symmetrical.

10. A formation-constrained well-controlled GRAN seismic signal broadband reconstruction method according to claim 9, characterized in that: The narrowband seismic signal of the logging data in the label training data set is input into the generator of the GRAN model. The generator structure starts from the decoder. At each time t, the corresponding encoder generates a sample, and the outputs of all times are superimposed as the final output.

Images