A structurally constrained depth-domain multichannel collaborative seismic inversion method and system

Abstract

This invention discloses a tectonically constrained depth-domain multichannel collaborative seismic inversion method and system. First, a depth-domain wavelet matrix is ​​constructed and combined with an impedance model to generate synthetic records consistent with actual seismic characteristics, thus building a pre-training dataset. Gaussian noise is superimposed on the synthetic data to simulate stratigraphic displacement. Then, plane wave decomposition is used to extract tectonic dip information from the enhanced data, establishing lateral spatial constraints. Tectonic dip constraint terms are embedded in a spatiotemporal neural network, and a multichannel collaborative inversion model is trained using synthetic data. Dip data from real seismic records is extracted and fused with well logging and stratigraphic interpretation results to establish a depth-domain low-frequency initial impedance model. Finally, by combining measured seismic data, tectonic constraint information, the initial model, and well data, network parameters are adjusted through transfer learning, ultimately achieving high-precision and high-stability impedance inversion. This method solves the problems of insufficient result stability, low resolution, and error accumulation effects in existing seismic inversion methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic inversion, and particularly relates to a structure-constrained deep domain multi-channel cooperative seismic inversion method and system. BACKGROUND

[0002] Seismic exploration excites seismic waves by artificial sources, and records the propagation characteristics of wave field in the underground medium by using geophones, and then inverts the physical properties of the underground geological structure. As the core link connecting seismic data and geological interpretation, seismic impedance (defined as the product of rock density and P-wave velocity) can comprehensively reflect key parameters such as reservoir porosity, fluid saturation and lithology, and provide quantitative basis for oil and gas reservoir identification and mineral resource evaluation. With the continuous innovation of seismic exploration technology, the reservoir prediction method has developed from the early experience type analysis to a high-precision technical system of multi-disciplinary integration. The current exploration focus is on complex hidden targets, such as thin interbedded reservoirs with a thickness less than the seismic resolution limit and micro-structural traps blocked by faults. Such targets put forward higher requirements on the signal-to-noise ratio of seismic data and the resolution of inversion algorithm.

[0003] In the oil and gas exploration under complex geological conditions, the traditional seismic inversion method has long relied on single-channel independent processing mode, which has the advantages of high computational efficiency and simple algorithm implementation. This kind of method is usually based on the assumption of horizontal layered medium, and uses the convolution model of fixed wavelet form to solve the seismic trace. However, this simplified processing mode has significant limitations when facing complex structures: the lack of lateral resolution leads to the ambiguity of thin interbed boundary, and the single-channel independent solution is sensitive to noise, which further reduces the reliability of reservoir parameter prediction. Especially when there is a lateral velocity jump or anisotropy in the subsurface medium, the traditional method cannot meet the current demand for fine characterization of thin reservoirs and complex structure imaging due to the neglect of the spatial correlation of seismic data and the geometric constraints of geological structure. It is urgent to break through the existing technical bottleneck by using more advanced inversion strategies. In addition, the traditional seismic inversion method is mainly carried out in the time domain, and its core technical bottleneck is the lack of characterization ability for thin reservoirs. This limitation is due to the irreversible loss of high-frequency components in the process of converting logging data from depth to time, which leads to a significant decrease in the vertical identification accuracy of thin layers (such as layers with a thickness less than the conventional seismic resolution). Based on the above limitations, the seismic inversion technology has experienced an important technical iteration from the time domain to the depth domain in the field of oil and gas reservoir characterization. In recent years, with the breakthroughs in high-performance computing and algorithm theory, prestack depth migration technology has been realized and has become a standard step in seismic data processing. However, seismic attribute inversion and reservoir parameter prediction based on depth migration results still face many challenges in methodology. In essence, this difficulty is due to the fundamental difference in the depth domain seismic response mechanism: when the velocity of the subsurface medium changes continuously with depth, the shape of the seismic wavelet will show spatial non-stationary characteristics. The forward modeling framework based on fixed wavelet and reflection coefficient convolution in traditional time-domain inversion is no longer applicable in the depth domain, which makes it difficult to directly migrate mature time-domain inversion algorithms to the depth domain. The current processing of depth domain seismic data in the field of oil and gas exploration generally adopts a makeshift strategy, which is essentially achieved by "time-depth conversion" to realize the technical transition: first, convert the depth domain migration imaging data to the time domain, use mature time-domain inversion technology to complete reservoir parameter prediction, and finally convert the results back to the depth domain for geological interpretation. This roundabout path fails to touch the core mechanism of depth domain inversion, and the error is amplified in the conversion process due to multiple domain conversion, which leads to information distortion of geological details such as thin interbed in the repeated resampling process, significantly reducing the characterization accuracy of complex reservoirs.

[0004] In summary, the current seismic impedance inversion method mainly has the following problems: 1. It relies on single-channel independent processing mode, ignores the spatial correlation of seismic data and the geological structure constraint, leads to ambiguous thin interbed boundary, distorted fault zone structure, and is sensitive to noise, and is difficult to depict the spatial distribution of complex reservoirs; 2. In the complex domain conversion process of "depth-time-depth", the high-frequency components are irreversibly lost, which significantly reduces the vertical identification accuracy of thin layers and restricts the fine characterization of hidden reservoirs; 3. The depth domain seismic wavelet presents spatial non-stationarity with the change of velocity, the fixed wavelet assumption in the time domain does not match the actual dynamic change of the wavelet characteristics in the depth domain, leading to amplitude response error and reservoir parameter prediction deviation; 4. The high-resolution algorithm such as multi-channel cooperation and anisotropy correction has high computational complexity, and the multi-scale data (seismic, logging, structure information) fusion capability is weak, making it difficult to predict complex reservoirs with high precision. SUMMARY

[0005] The purpose of the present application is to provide a structure-constrained depth-domain multi-channel cooperative seismic inversion method and system, which adopts a depth learning-driven multi-channel joint inversion method, establishes the spatial correlation between channels by introducing a structure dip angle constraint term, and realizes the inversion modeling of stratigraphic continuity by using a dynamic window alignment mechanism. A global mean square error constraint term based on dip angle guidance is designed, the spatial alignment relationship of adjacent channel window data is dynamically adjusted by calculating the stratigraphic offset based on the structure dip angle data, the longitudinal data matching degree and the lateral geological structure continuity are simultaneously optimized in the loss function, and the fusion of the dip angle constraint and the inversion process is realized based on the space-time neural network (STNN) framework. The network parameters and geological structure characteristics are simultaneously updated through back propagation, which maintains high resolution while enhancing the reservoir characterization capability in complex structural background, and provides a high-reliability geophysical solution for fine description of oil and gas reservoirs in thin interbed and complex structural background.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] On the one hand, the present application provides a structure-constrained depth-domain multi-channel cooperative seismic inversion method, comprising the following steps:

[0008] Comprising the following steps:

[0009] S1, constructing a depth domain wavelet matrix by using an interpolation method, synthesizing depth domain seismic records based on non-stationary convolution model using impedance model data, providing a benchmark data set for subsequent network training, and ensuring the consistency of the synthesized data and the real depth domain seismic data;

[0010] S2, adding Gaussian noise to the seismic records to simulate random noise and acquisition interference in actual seismic data, and performing up and down layer dislocation on the noisy seismic records to enhance the robustness of the model to complex structures;

[0011] S3, structural dip data of noisy seismic records are extracted by using plane wave decomposition method (PWD), to provide high reliability lateral constraint information for subsequent seismic inversion;

[0012] S4, spatial correlation constraints of inversion results are established by using space-time neural network (STNN) and loss function, and the network is pre-trained by using noisy seismic records and extracted structural dip data;

[0013] S5, structural dip data of real seismic records are extracted by using the method of S3, and low-frequency initial impedance parameter model in depth domain is established by using seismic data horizon interpretation data and well logging data in the working area;

[0014] S6, the final impedance inversion result is obtained by adjusting the network parameters of real seismic records and corresponding structural dip data and initial impedance data.

[0015] In some embodiments, in S1, the non-stationary convolution model is:

[0016] ;

[0017] The depth domain seismic record is:

[0018] ;

[0019] Wherein,

[0020] , , ;

[0021] In the formula, is a seismic wavelet varying with depth; is the displacement of the seismic wavelet in the depth direction; is the depth; is the reflection coefficient corresponding to the seismic wavelet in the depth direction; is the depth domain seismic record: is a non-stationary depth variable wavelet matrix; is a depth domain reflection coefficient; is; is the seismic wavelet at depth ; is the corresponding depth domain reflection coefficient at depth z.

[0022] In some embodiments, S2 includes the following steps:

[0023] S21, Gaussian noise intensity is randomly generated according to uniform distribution;

[0024] S22, the noisy seismic data is divided into several non-overlapping blocks, each 5 channels as a group.

[0025] S23. Set random offsets for non-overlapping blocks in areas of severe tectonic activity (where the absolute value of the tectonic dip angle in seismic data is greater than 30°), and fill in the missing values ​​after the offset using cubic spline interpolation.

[0026] In some embodiments, the calculation expression for superimposed Gaussian noise on seismic records is as follows:

[0027] ;

[0028] In the formula, This is noisy seismic data; This represents the number of depth sampling points; This refers to the number of earthquake traces; This is the normalization result for the synthetic depth-domain seismic record data; For noise intensity, ; It follows a standard Gaussian distribution.

[0029] In some embodiments, S3 includes the following steps:

[0030] The local slope is obtained by using the local plane wave differential equation. When the local slope is constant, the plane wave field is obtained.

[0031] The tilt angle solution equation is obtained by using the filtering operator;

[0032] In each iteration of the equation for constructing the tilt angle, the slope increment is obtained to update the initial slope. The solution is then solved cyclically until the solution converges, and finally the tilt angle field is obtained.

[0033] In some embodiments, the equation for constructing the tilt angle is:

[0034] ;

[0035] In the formula, For the first The increment of the slope in each iteration; Earthquake record data; For the first The filtering operator in the next iteration; for The first-order partial derivative with respect to, where This is the initial slope value. This represents the number of iterations.

[0036] In some embodiments, S4 includes the following steps:

[0037] The Huber loss constraint is used to predict the matching degree between impedance and label data. The auxiliary task dynamically calculates the waveform similarity of adjacent seismic traces under tectonic guidance by constructing dip constraint terms.

[0038] The structural dip angle constraint term is constructed and the structural dip angle data is used to calculate the structural offset of strata based on the sliding window mechanism, so as to dynamically align the adjacent seismic data and quantify the spatial continuity error of the structural direction;

[0039] The prediction result of the model impedance data is obtained by using the synthetic deep domain seismic record data, filtering the initial impedance model obtained from the impedance model data (the classic Overthrust nappe impedance model of the network open source) and the extracted structural dip angle data.

[0040] In some embodiments, the loss function is:

[0041] ;

[0042] wherein,

[0043] ;

[0044] ;

[0045] wherein, is a loss function in neural network training; is a label constraint term in the loss function; is a structural dip angle constraint term in the loss function; is a set of network parameters to be trained; , is a weight value of different losses in the training process; is label data corresponding to the seismic data; is a network to be trained; is seismic record data; is an initial impedance parameter model input into the network; is a robustness threshold value for controlling the Huber loss; is a sample point number; is a depth sampling point number; is a seismic trace number; is a relative offset index in the sliding window; is a local offset correlation range; is the first trace impedance value obtained in the network training process; is an offset of each sample point;

[0046] ;

[0047] is a structural dip angle value of each sampling point extracted according to the seismic data; is a rounding function.

[0048] In another aspect, the application is a deep domain multi-channel collaborative seismic inversion system for structural constraints, which uses the above method and comprises the following modules:

[0049] A synthesis module: a deep domain wavelet matrix is constructed, and a deep domain seismic record is synthesized based on a non-stationary convolution model;

[0050] A noise adding module: used for adding Gaussian noise to the seismic record and performing up and down layer dislocation on the noisy seismic record;

[0051] A structural dip angle module: structural dip angle data of the noisy seismic record are extracted by using a plane wave decomposition method;

[0052] A pre-training module: a spatial correlation constraint of an inversion result is established by using a loss function through a space-time neural network, and the network is pre-trained by using the noisy seismic record and the extracted structural dip angle data;

[0053] An impedance parameter module: structural dip angle data of a real seismic record are extracted, and a low-frequency initial impedance parameter model in the deep domain is established by using seismic data layer interpretation data and logging data in the work area;

[0054] An acquisition module: the final impedance inversion result is obtained by adjusting the pre-training network parameters by using the real seismic record, the corresponding structural dip angle data and the initial impedance data.

[0055] Compared with the prior art, the application has the following beneficial effects:

[0056] 1. The application synchronously optimizes longitudinal data matching and lateral geological structure continuity through the joint inversion of the structural dip angle constraint term and the dynamic window alignment mechanism, effectively suppresses the solution space divergence problem of the traditional inversion method, and significantly reduces the multi-solution property of the inversion result.

[0057] 2. The application improves the lateral continuity of the inversion result in complex structural areas such as faults and pinchouts by combining the fault processing strategy of the non-stationary wavelet matrix synthesis seismic record and the PWD dip angle extraction, and enhances the adaptability of the model to complex geological conditions.

[0058] 3. The application reduces the dependence on large-scale labeled data based on the two-stage strategy of pre-training based on synthetic data and fine-tuning based on real data, accelerates network convergence, and at the same time guarantees the high consistency of the inversion result and the well data.

[0059] 4. The application is completed in the deep domain from wavelet construction, seismic record synthesis to inversion modeling, avoids the accumulation of time-depth conversion errors, and improves the accuracy of the deep domain inversion result. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 It is a schematic diagram of the overall process of the application.

[0061] Figure 2 Fig. 2 is a schematic diagram of a depth domain seismic record synthesized based on a non-stationary convolution model according to single-channel model data of embodiment 1 of the present application;

[0062] Figure 3 Fig. 3 is a schematic diagram of synthetic post-stack depth domain seismic data with noise (a), an initial impedance parameter model (b), extracted structural dip data (c) and predicted results of the model impedance data (d) of pre-training input of embodiment 1 of the present application;

[0063] Figure 4 Fig. 4 is a schematic diagram of real post-stack depth domain seismic data (a), an initial impedance parameter model (b) and extracted structural dip data (c) of input of embodiment 1 of the present application;

[0064] Figure 5 Fig. 5 is a schematic diagram of impedance prediction results obtained by the structural constraint depth domain multi-channel collaborative intelligent seismic inversion method of embodiment 1 of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0066] Embodiment 1:

[0067] Please refer to Figures 1-5 A structural constraint depth domain multi-channel collaborative seismic inversion method, first constructs a depth domain wavelet matrix based on non-stationary convolution theory, generates synthetic records consistent with actual seismic characteristics in combination with an impedance model, and constructs a high-precision pre-training data set. Secondly, Gaussian noise is added to the synthetic data and layer position is simulated to strengthen the robustness of the model to actual noise interference and structural distortion; structural dip information is extracted from the enhanced data using a plane wave decomposition method (PWD) to establish a horizontal spatial constraint relationship; then, a structural dip constraint term is embedded in a space-time neural network (STNN), a multi-channel collaborative inversion model is trained through synthetic data to strengthen the geological correlation between channels; PWD dip angle data based on real seismic records are extracted, logging and layer interpretation results are fused to establish a depth domain low-frequency initial impedance model; finally, measured seismic data, structural constraint information, an initial model and well data are combined to fine-tune network parameters through transfer learning, and finally high-precision and high-stability impedance inversion is realized.

[0068] Specifically includes the following steps:

[0069] S1, the original time domain seismic wavelet is interpolated to obtain a depth domain wavelet matrix of uniform depth sampling points, and a depth domain seismic record convolution model is obtained through a non-stationary convolution model: the depth domain wavelet matrix is convolved with a depth domain reflectivity sequence obtained from the model impedance parameter data to synthesize a depth domain seismic record.

[0070] In the depth domain, due to the influence of the change of the formation velocity on the seismic wavelet, the depth domain seismic data can be regarded as a seismic profile of the change of the seismic wavelet with depth, and the convolution model of the depth domain seismic record can be expressed in the form of non-stationary convolution:

[0071] ;

[0072] In order to facilitate the construction of the inversion objective function, it can be written in the form of a matrix:

[0073] ;

[0074] Wherein,

[0075] , , ;

[0076] Wherein, the depth domain wavelet matrix is obtained by interpolating the original time domain wavelet, and each seismic wavelet in the depth domain wavelet matrix is different, which reflects the characteristics of the change of the depth domain seismic wavelet with depth. As shown in Figure 2 (a) is a one-dimensional impedance parameter model constructed by experience, (b) is the reflectivity corresponding to the model shown in (a), and (c) is a one-dimensional depth domain seismic record synthesized by using a non-stationary convolution model, and it can be seen that the waveform is not symmetrical at the impedance interface and presents non-stationarity.

[0077] S2, the seismic data synthesized in S1 is superimposed with Gaussian noise (SNR=1-5dB) to simulate random noise and system interference in field acquisition. Secondly, the layer position error processing is carried out on the noisy record to simulate the stratum dislocation caused by the fault (the dislocation distance is 1-7 sampling points).

[0078] The Gaussian noise of each data of the synthesized depth domain seismic record is superimposed through the following formula:

[0079] ;

[0080] where the noise intensity is randomly generated according to uniform distribution, and the signal-to-noise ratio of the noisy seismic data is 1-5 dB; the noisy seismic data is divided into several non-overlapping blocks, each block containing 5 traces; a random offset is set for the blocks in the tectonic activity area (the absolute value of the structural dip of the seismic data is greater than 30°), and the missing values after the offset are filled by the cubic spline interpolation method. The two-dimensional simulated seismic record obtained by applying artificial interference above is shown in Fig. Figure 3 (a).

[0081] S3, extracting the synthetic seismic dip angle data by using the plane wave decomposition method (PWD): based on the local plane wave assumption, the dip angle estimation is optimized by multi-scale window scanning, the horizontal smoothing constraint is performed on the combined multi-channel data, the local slope is solved by nonlinear inversion, and finally the high-precision dip angle data is converted.

[0082] The local slope is obtained by the local plane wave differential equation:

[0083] ;

[0084] The local plane wave differential equation is:

[0085] ;

[0086] In the formula, is the plane wave field; is the local slope.

[0087] When the local slope is constant, the plane wave field expression can be obtained:

[0088] ;

[0089] In the formula, is an arbitrary waveform; is the slope value; is the distance; is the time.

[0090] In order to estimate the local slope by the PWD filter, a filter operator (indicating the convolution of data and two-dimensional filter) is introduced, and finally the structural dip angle solution equation can be obtained:

[0091]

[0092] The slope increment is obtained in each iteration of the structural dip angle solution equation, which is used to update the initial slope, and then the solution is iteratively solved until the solution converges. The structural dip angle field corresponding to the two-dimensional simulated seismic record extracted according to the above method is shown in Fig. Figure 3 (c).

[0093] S4, a space-time neural network (STNN) is used and a structural dip angle constraint mechanism is introduced to establish a spatial correlation constraint for the inversion result through a loss function. The loss function is designed in a multi-task fusion mode: a main task adopts a Huber loss to constrain the matching degree of the predicted impedance and the label data, and an auxiliary task dynamically calculates the similarity of the inversion impedance data of adjacent seismic traces in the structural direction through a structural dip angle constraint term. The structural constraint module is based on a sliding window mechanism, calculates the stratigraphic offset through dip angle data, dynamically aligns the seismic data of adjacent traces, and quantifies the spatial continuity error of the structural direction. Then, the data obtained in S2 and S3 are used as input data for network pre-training.

[0094] The network training adopts an Adam optimizer (initial learning rate 0.01, momentum parameters 0.9 / 0.99) and the total training period is 1500 rounds.

[0095] The loss function is in a multi-task fusion mode: a task adopts a Huber loss (weight 0.975) to constrain the matching degree of the predicted impedance and the label data, and an auxiliary task dynamically calculates the waveform similarity of adjacent seismic traces in the structural direction through a structural dip angle constraint term (weight 0.025), and the calculation expression is:

[0096] ;

[0097] Among them,

[0098] ;

[0099] ;

[0100] The structural dip angle constraint term is based on a sliding window mechanism (window size 9) and calculates the stratigraphic offset through dip angle data:

[0101] ;

[0102] Dynamically aligning the seismic data of adjacent traces quantifies the spatial continuity error of the structural direction.

[0103] Then, the above synthetic seismic record data, the initial impedance model obtained by filtering the impedance model data, and the extracted structural dip angle data are used for network pre-training to obtain the prediction result of the model impedance data, as shown in Figure 3 (a), (b), (c), as shown in Figure 3 (d), and the network parameters are saved. In particular, the seismic records participating in the label loss calculation in the pre-training stage are 3 traces (100th, 600th, and 2500th traces).

[0104] S4, constructing the dip angle data from the measured seismic data according to the method in S3, and fusing the seismic horizon interpretation results and the well logging data in the work area to construct a low-frequency initial impedance model in the depth domain.

[0105] The structural dip angle data (as shown in Figure 4 (a)) are extracted from the real seismic records (as shown in Figure 4 (c)), and a geological structure framework model is constructed by using the seismic structure interpretation data (such as the spatial structural features of the fault system and the stratum contact relationship) and combining the regional sedimentary mode analysis, including the sedimentary facies belt distribution law, the stratum superimposed style, and the lithofacies lateral variation trend. The framework model represents the elements such as the geometric shape (such as the stratum thickness and occurrence), the structural unit division, and the spatial contact relationship of the stratum. Under the constraint of the structural framework, the wave impedance parameters obtained from the well logging data are extrapolated by the spatial interpolation algorithm controlled by the structural trend, the discrete well point data are expanded to the positions of each survey line in the seismic interpretation, and finally the wave impedance initial model (as shown in Figure 4 (b)) that matches the geological structure and the sedimentary features is established. The establishment of the impedance parameter model is based on the spatial interpolation method: the two-dimensional plane modeling is performed on the seismic interpretation horizon by discrete data interpolation, the geological horizon profile is formed to depict the spatial shape of the sedimentary body, then the point-shaped well logging information is expanded to the continuous distribution along the two-dimensional plane and extrapolated to the well-uncontrolled area by implementing the lateral interpolation of the well logging impedance data under the constraint of the horizon profile, and finally the two-dimensional initial impedance parameter model covering the whole range of the work area is generated.

[0106] S6, realizing the collaborative inversion of the multi-channel seismic data by using the transfer learning strategy: loading the pre-trained network weight, freezing the parameters of the feature extraction layer to retain the general features of the seismic data; based on the real seismic records, the structural dip angle data, and the initial impedance model in the depth domain, the network parameters are jointly optimized by the dynamic weighting mechanism, wherein the structural dip angle data constrain the spatial prediction direction and the well data correct the low-frequency trend error; finally, the inversion results meet the seismic waveform matching, the structural guidance continuity, and the well logging data hard constraint by iterative training, and the high-precision impedance inversion result is obtained.

[0107] Specifically, the parameters of the feature extraction layer (including two convolution modules and one GRU time series layer) are fixed under the transfer learning framework according to the measured seismic records, the structural dip angle data obtained by the above method, the initial impedance parameter model, and the measured well impedance data (the 80th, 127th, and 233rd channels), the weight matrix parameters of the linear layer of the neural network are updated by using the real seismic data, the training is terminated when the validation set loss does not decrease for 3 consecutive training rounds, and the output of the network is the depth domain impedance parameter model (as shown in Figure 5 ).

[0108] Example 2

[0109] A kind of structural constraint deep domain multi-channel collaborative seismic inversion system, using above-mentioned method, include following modules:

[0110] Synthesis module: by constructing depth domain wavelet matrix, and based on non-steady state convolution model synthesis depth domain seismic record;

[0111] Noise adding module: for seismic record superposition Gaussian noise, and to the noisy seismic record is up and down layer dislocation;

[0112] Structural dip module: by using plane wave decomposition method extraction noisy seismic record structural dip data;

[0113] Pre-training module: by using space-time neural network, by loss function the spatial correlation constraint of inversion result is established, and with noisy seismic record and the extracted structural dip data network pre-training is carried out;

[0114] Impedance parameter module: by extracting real seismic record structural dip data, and using the seismic data layer interpretation data and well logging data in work area, establish the low-frequency initial impedance parameter model in depth domain;

[0115] Acquisition module: using real seismic record and its corresponding structural dip data, initial impedance data are used for pre-training network parameter adjustment, and final impedance inversion result is obtained.

[0116] A kind of structural constraint deep domain multi-channel collaborative seismic inversion system of the application can be installed in computer equipment.The computer equipment includes processor, memory and computer program stored in the memory and can be run on the processor, for example, a kind of structural constraint deep domain multi-channel collaborative seismic inversion program.Various types of readable storage medium are included in the memory at least, the readable storage medium includes flash memory, mobile hard disk, multimedia card, card type memory (for example: SD or DX memory etc.), magnetic memory, disk, optical disc etc.The processor is the control core of the electronic equipment, utilizes various interfaces and line connections the components of entire computer equipment, by running or executing the program or module stored in the memory, and call the data stored in the memory, to execute the various functions of computer equipment and process data.

[0117] The module of the application refers to a series of computer program segments that can be executed by the processor of computer equipment and can complete fixed function, which is stored in the memory of computer equipment.

[0118] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the application embrace any and all variations of the present application that fall within the scope of the general inventive concept as defined in the appended claims and that the application may be practiced otherwise than is specifically recited in the detailed description.

Claims

1. A method of constructing a constrained deep domain multi-channel collaborative seismic inversion, characterized in that, The method comprises the following steps: S1, constructing a deep domain wavelet matrix, and synthesizing a deep domain seismic record based on a non-stationary convolution model; S2, adding Gaussian noise to the seismic record, and performing up and down layer dislocation on the noisy seismic record; comprising the following steps: S21, generating Gaussian noise intensity according to uniform distribution; S22, dividing the noisy seismic data into several non-overlapping blocks in groups of 5 channels; S23, setting a random offset for the non-overlapping blocks with an absolute value of dip angle greater than 30°, and filling the missing values after offsetting by a cubic spline interpolation method; S3, extracting structural dip angle data of the noisy seismic record by using a plane wave decomposition method; S4, in the space-time neural network, the spatial correlation constraint of the inversion result is established by a loss function, and the network is pre-trained by using the noisy seismic record and the extracted structural dip angle data; comprising the following steps: The matching degree of the predicted impedance and the label data is constrained by using a Huber loss, and the auxiliary task dynamically calculates the waveform similarity of adjacent seismic traces in the structural direction by constructing a dip angle constraint term; The dip angle constraint term is constructed and based on a sliding window mechanism, the stratigraphic offset is calculated by using the dip angle data, the seismic data of adjacent channels are dynamically aligned, and the spatial continuity error of the structural direction is quantified; The prediction result of the model impedance data is obtained by pre-training the network by using the synthesized deep domain seismic record data, the initial impedance model obtained by filtering the impedance model data, and the extracted structural dip angle data; S5, extracting the structural dip angle data of the real seismic record, and establishing a low-frequency initial impedance parameter model in the depth domain by using the seismic data layer interpretation data and the logging data in the work area; S6, adjusting the network parameters by using the real seismic record and the corresponding structural dip angle data and the initial impedance data, and obtaining the final impedance inversion result.

2. The method of claim 1, wherein, In S1, the non-stationary convolution model is: ； The deep domain seismic record is: ； wherein, ， ， ； wherein, is a seismic wavelet varying with depth; is the displacement of the seismic wavelet in the depth direction; is the depth; is the reflection coefficient corresponding to the seismic wavelet in the depth direction; is a depth domain seismic record: is a non-stationary depth varying wavelet matrix; is a depth domain reflection coefficient; is; is the seismic wavelet at depth is the seismic wavelet at depth is the depth domain reflection coefficient corresponding to depth z.

3. The method of claim 1, wherein, The calculation expression for adding Gaussian noise to the seismic record is: ； wherein is noisy seismic data; is the number of depth samples; is the number of seismic traces; is the normalized result of the synthetic depth domain seismic record data; is the noise intensity, ; is a standard Gaussian distribution.

4. The method of claim 1, wherein, S3 comprises the following steps: The local slope is obtained by a local plane wave differential equation, and the plane wave field is obtained when the local slope is constant; The structural dip angle solution equation is obtained by a filtering operator; The slope increment is obtained in each iteration of the structural dip angle solution equation, which is used to update the initial slope, and then the solution is repeatedly solved until the solution converges, and finally the structural dip angle field is obtained.

5. The method of claim 4, wherein, The structural dip angle solution equation is: ； wherein is the slope of the filter at the iteration number n; is the slope of the filter at the iteration number n; is the seismic record data; is the filter operator at the iteration number n; is the filter operator at the iteration number n; is the first order partial derivative of f with respect to x, where is the first order partial derivative of f with respect to x, where is the initial slope value, is the number of iterations.

6. The method of claim 1, wherein, The loss function is: ； wherein, ； ； wherein, is a loss function in neural network training; is a label constraint term in the loss function; is a structural dip constraint term in the loss function; is a set of network parameters to be trained; , is a weight for the contribution of different losses to the training process; is label data corresponding to seismic data; is a network to be trained; is seismic record data; is an initial impedance parameter model input to the network; is a robustness threshold for controlling Huber loss; is a number of sample points; is a number of depth sampling points; is a number of seismic traces; is a relative offset index within a sliding window; is a local offset correlation range; is a first trace impedance value obtained in the network training process; is an offset of each sample point; ； constructing dip values for each sample point extracted from the seismic data; is a rounding function.

7. A system for constructing a structure-constrained depth domain multi-channel simultaneous seismic inversion, using the structure-constrained depth domain multi-channel simultaneous seismic inversion method of any one of claims 1-6. It comprises: The synthesis module: synthesizing a deep domain seismic record by constructing a deep domain wavelet matrix and based on a non-stationary convolution model; The noise adding module: for adding Gaussian noise to the seismic record, and performing up and down layer dislocation on the noisy seismic record; The structural dip angle module: extracting the structural dip angle data of the noisy seismic record by using a plane wave decomposition method; The pre-training module: using a space-time neural network, establishing the spatial correlation constraint of the inversion result by a loss function, and pre-training the network by using the noisy seismic record and the extracted structural dip angle data; The impedance parameter module: extracting the structural dip angle data of the real seismic record, and establishing a low-frequency initial impedance parameter model in the depth domain by using the seismic data layer interpretation data and the logging data in the work area; The acquisition module adjusts the pre-training network parameters by using the real seismic record and the corresponding structural dip angle data and initial impedance data, and obtains the final impedance inversion result.

Images