CN116755145A

CN116755145A - Three-dimensional seismic horizon intelligent tracking method based on multi-seismic attribute regression network

Info

Publication number: CN116755145A
Application number: CN202310883390.0A
Authority: CN
Inventors: 钱峰; 何宇; 华浩为; 温宇航; 胡光岷
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2023-07-18
Filing date: 2023-07-18
Publication date: 2023-09-15
Anticipated expiration: 2043-07-18
Also published as: CN116755145B

Abstract

The invention discloses a three-dimensional seismic layer intelligent tracking method based on a multi-attribute regression network, constructs a multi-attribute regression network, expresses the three-dimensional seismic layer tracking problem as a multi-attribute regression model, and then uses the DCAE encoder structure Extract depth features and use them as input sequences, perform sequence regression analysis through LSTM, then reconstruct the seismic horizon image through the predicted value output by LSTM and the DCAE decoder structure to complete the data reconstruction, and finally output the seismic horizon image to complete the three-dimensional seismic Layer tracking. The method of the present invention introduces system state equations to establish time relationships, captures time dynamics and improves the accuracy of regression analysis. It extracts depth features through DCAE and uses them as input sequences, so that LSTM can effectively simulate the relationship between multiple seismic attributes and target layers. Spatiotemporal relationship. Compared with directly using LSTM, the DCAE‑LSTM structure can fully leverage spatiotemporal correlation and utilize multiple seismic attributes to achieve accurate and robust tracking results.

Description

Three-dimensional seismic horizon intelligent tracking method based on multi-seismic attribute regression network

Technical Field

The invention belongs to the technical field of seismic image interpretation, and particularly relates to a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network.

Background

Horizon tracking plays a vital role in seismic interpretation because it is an important and time consuming challenge. Accurate and reliable seismic horizon tracking is critical to determining the structure and stratigraphic framework of subsurface geologic formations. Thus, many horizon tracking methods using seismic attributes have been developed with favorable results, such as local slope-based least squares and waveform similarity-based methods.

Summary of these approaches, however, shows that a fundamental challenge is how to effectively select and exploit seismic attributes to improve horizon tracking accuracy, such as across discontinuous formations (e.g., faults) and continuous reflective surfaces in the lateral direction (e.g., local slopes). The core challenge is to effectively utilize the appropriate seismic attributes to improve tracking accuracy, and many existing model-based or Deep Learning (DL) tracking methods rely on mathematical or data-driven mappings between seismic data and target layers, respectively, both of which utilize only seismic data or a single attribute, which can lead to discontinuous and inaccurate tracking.

Current horizon tracking methods based on seismic attributes can be roughly divided into three main categories: methods based on instantaneous phase, on waveform similarity, and on local slope. The instantaneous phase-based approach involves an initial unpacking of the instantaneous phases to derive a representative relative geological time volume (RGT), and then extracting the contour as a horizon. In these methods based on waveform similarity, the position where waveform similarity is maximum is recursively tracked starting from one seed point. Most local slope-based methods estimate horizons using local reflection slopes by employing least squares. However, these methods require the creation of a strict mathematical mapping between the seismic attributes and the target horizons, relying on certain model assumptions, such as least squares (linear assumptions).

Unlike the model-driven approach described above, the Deep Learning (DL) -based data-driven approach attempts to intuitively learn the direct mapping between seismic attributes and corresponding target horizons without requiring manual making assumptions. As a typical DL network, a deep Convolutional Neural Network (CNN) can map seismic data directly to target horizons through supervised training. In the indirect mapping aspect, the CNN may automatically acquire a two-dimensional RGT image from a two-dimensional seismic image by training, and then extract the seismic horizon from the RGT contour. Unlike using only a single seismic data, the dual-branch network simultaneously takes the original seismic data and the converted time-frequency domain data as input, and can map more comprehensively with the target horizon. However, these DL-based methods rely primarily on seismic data or a single correlation attribute, lacking guidance for other attributes, and thus may not correlate reflections on both sides of a fault in the absence of fault attributes.

Disclosure of Invention

In order to solve the technical problems, the invention provides a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network, provides a multi-attribute regression network (MARN) for three-dimensional seismic horizon tracking, utilizes a plurality of seismic attributes to realize accurate and steady tracking results, constructs a multi-attribute regression model, then introduces a system state equation to establish a time relationship, and designs a Deep Convolutional Automatic Encoder (DCAE) to automatically learn spatial correlation from the attributes so as to extract depth features.

The invention adopts the technical scheme that: a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network comprises the following specific steps:

s1, constructing a multi-attribute regression network, and expressing a three-dimensional seismic horizon tracking problem as a multi-attribute regression model;

s2, extracting depth features through a DCAE encoder structure;

s3, based on the depth features extracted in the step S2 as an input sequence, performing sequence regression through a long-short-term memory network LSTM, and analyzing time correlation among various attribute depth features;

s4, reconstructing a seismic horizon image through a DCAE decoder structure based on the predicted value output by the LSTM in the step S3, and completing data reconstruction;

s5, outputting the seismic horizon image to complete three-dimensional seismic horizon tracking.

Further, the step S1 specifically includes the following steps:

constructing a multi-attribute regression network by combining DCAE and LSTM, wherein the main objective of the three-dimensional seismic horizon tracking problem is to extract a target horizon from seismic data

wherein ,representing the target horizon->Representing real space, n ₁ 、n ₂ and n₃ The number of inline, crossline, and timeline seismic interpretation profiles are shown, respectively.

Converting the three-dimensional seismic horizon tracking problem into a multi-attribute regression model that is mathematically represented as:

where Θ represents the network weight,represents a fitted multiple regression function with Θ as a parameter,represents a plurality of seismic attributes, N represents the number of seismic attributes, [ N ]]The abbreviation for {1,2,...

Providing a state equation, analyzing state information related to each seismic attribute in each time step, and expressing a state prediction process and a measurement process as follows:

wherein ,representing the current state by adding various seismic attributes +.>With previous state variablesTaken together, the output of the system at the current time step t is derived> Representing the time-varying process noise,representing measurement noise +.> andRepresenting state transitions and observation functions, respectively.

Further, the step S2 specifically includes the following steps:

give a pair of training samples andRepresents-> andRespectively representing ground reality and input training data, and k represents the number of sections.

Complex depth features are extracted using the encoder structure of DCAE:

wherein ,representing the encoder module, extracting depth features +.>θ represents an implicit parameter of the DCAE network encoder.

Further, the step S3 specifically includes the following steps:

sequence regression by LSTM analyzes the temporal correlation between various attribute depth features, and equation (2) is restated as:

wherein ,respectively represent the nth seismic attributeState vector, observation vector, process noise vector and measurement noise vector at time step t, +.>Representing the depth profile of the nth seismic attribute at time step t. Then have the same parameter ψ _n Is +.>Obtained by:

wherein ,Ψ_n Parameters representing the nth seismic attribute LSTM branch.

Further, the step S4 specifically includes the following steps:

in deriving predicted response vectorsThese vectors are then concatenated by the decoder structure of the DCAE and converted into a visually interpretable image:

wherein ,representing a seismic horizon image, ++>Representing an abstract decoder function.

System modeling using DCAE-LSTM architecture, updating network weights Θ, including θ and ψ _n By minimizing the loss functionRealizing output convergence of the neural network:

wherein ,representing an abstract decoder function.

Further, the step S5 specifically includes the following steps:

the mathematical representation of the state transition for each time step is as follows:

wherein ,represents an internal state vector, delta (·) represents a sigmoid function, < >>Respectively, representing weights associated with the internal state vectors. Current state +.>The expression is as follows:

wherein the symbol +.is the Hadamard product between the two vectors. Then equation (4) may be rewritten as:

wherein, tan h represents the hyperbolic tangent function.

Finally, carrying out mathematical abstraction on the formula (4) through LSTM,obtained prediction resultsIs used as an input to a decoding moduleFinal output seismic horizon image +.>And (5) completing three-dimensional seismic horizon tracking.

The invention has the beneficial effects that: firstly, constructing a multi-attribute regression network, expressing a three-dimensional seismic horizon tracking problem as a multi-attribute regression model, extracting depth features through a DCAE encoder structure and taking the depth features as an input sequence, carrying out sequence regression analysis through LSTM, reconstructing a seismic horizon image through a predicted value output by LSTM and a DCAE decoder structure, completing data reconstruction, and finally outputting the seismic horizon image, thus completing three-dimensional seismic horizon tracking. According to the method, a system state equation is introduced to establish a time relation, time dynamics is captured, accuracy of regression analysis is improved, depth features are extracted through DCAE and are used as input sequences, so that LSTM can effectively simulate space-time relations between a plurality of seismic attributes and target horizons.

Drawings

FIG. 1 is a flow chart of a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network.

FIG. 2 is a diagram of an embodiment of a MARN according to the present invention.

FIG. 3 is a graph comparing original seismic data, corresponding local gradients, and fault attributes in an embodiment of the invention.

FIG. 4 is a graph comparing two-dimensional profile trace results of the method of the present invention with the single input DCAE-LSTM method in the Qian dataset of the present invention.

FIG. 5 is a graph of three-dimensional horizon surface comparisons extracted by the method of the present invention and a single input DCAE-LSTM method in an embodiment of the present invention.

FIG. 6 is a graph comparing the two-dimensional profile trace results of the method of the present invention and the single input DCAE-LSTM method in the F3 dataset of the present invention.

Fig. 7 is a graph of a three-dimensional horizon surface comparison of the method of the present invention with a single input DCAE-LSTM method.

FIG. 8 is a graph comparing Absolute Errors (AE) produced by the method of the present invention and the single input DCAE-LSTM method in the examples of the present invention.

Detailed Description

The invention is further described below with reference to the drawings and examples.

As shown in fig. 1, the method for intelligent tracking of three-dimensional seismic horizons based on multi-seismic attribute regression network comprises the following specific steps:

s2, extracting depth features through a DCAE encoder structure;

In this embodiment, the step S1 is specifically as follows:

Considering that higher accuracy of tracking results is obtained by using various seismic attributes, the problem is translated into a multi-attribute regression model expressed mathematically as:

In order to exploit the temporal correlation between target layer sites, state equations are proposed for analyzing the state information associated with each seismic attribute in each time step, which involve a state prediction process and a measurement process:

Considering that LSTM has the ability to build long-term dependencies and capture critical temporal dynamics, LSTM is utilized to identify systems.

It is not neglected that various seismic attributes are highly spatially correlated with target layer sites, such as discrete lateral variations and faults. In order to more comprehensively grasp the space-time mapping relation between different seismic attributes and target layer sites, the MARN provided by the method of the invention is established on the basis of DCAE and LSTM by using a DCAE-LSTM architecture.

In this embodiment, the step S2 is specifically as follows:

depth feature extraction:

Subsequently, complex depth features are extracted using the encoder structure of the DCAE:

In this embodiment, the step S3 is specifically as follows:

sequence regression analysis:

sequence regression by LSTM analyzes the temporal correlation between various attribute depth features, and thus equation (2) is restated as:

wherein ,state vector, observation vector, process noise vector and measurement noise vector, respectively representing nth seismic attribute at time step t,/->Representing the depth profile of the nth seismic attribute at time step t. Thus, having the same parameter ψ _n Is +.>Obtained by:

wherein ,Ψ_n Parameters representing the nth seismic attribute LSTM branch.

In this embodiment, the step S4 is specifically as follows:

and (3) data reconstruction:

The purpose of system modeling using the DCAE-LSTM architecture is to update network weights Θ, including θ and ψ _n Aimed at minimizing the loss functionRealizing output convergence of the neural network:

wherein ,representing an abstract decoder function.

By combining theta and psi _n Equation (7) involves depth feature extraction, sequential regression analysis, and data reconstruction, aiming at achieving a global minimum in theory. FIG. 2 is a diagram showing a MARN according to the method of the present invention.

In this embodiment, the step S5 is specifically as follows:

the DCAE is trained to find the best weights and other coefficients for the neural network. Therefore, emphasis is on the solution defined in the formula (4)An LSTM cell is detailed in fig. 2, and the mathematical representation of the state transitions for each time step is as follows:

wherein ,represents an internal state vector, delta (·) represents a sigmoid function, < >>Respectively, representing weights associated with the internal state vectors. Then, the current state +.>Is obtained by the following steps:

wherein the symbol +.is the Hadamard product between the two vectors. Therefore, the formula (4) can be rewritten as:

wherein, tan h represents the hyperbolic tangent function.

Finally, carrying out mathematical abstraction on the formula (4) through LSTM to obtain a prediction resultIs used as an input to a decoding moduleFinal output seismic horizon image +.>And (5) completing three-dimensional seismic horizon tracking.

The invention also provides a method for further verifying the invention in example 2, which is specifically as follows:

attribute selection is first performed in order to ensure that horizons across faults are correctly tracked and that temporal samples belonging to the same horizon exhibit similar local gradients while taking into account the original seismic data, the corresponding local gradients and fault attributes, as shown in fig. 3.

Fig. 3 (a) is an original seismic data map, fig. 3 (b) is a local slope attribute map, fig. 3 (c) is a fault attribute map, fig. 3 (d) is a result map using only the original seismic data, and fig. 3 (e) is a result map using three seismic attributes.

The cross-horizon phenomenon is evident in the rectangular area highlighted in fig. 3 (d), mainly in the area of lateral discontinuous reflection, concurrent with the local gradient and fault properties in fig. 3 (b) and 3 (c). Fig. 3 demonstrates the necessity to utilize these three properties.

The parameter N is set to 3 in view of the input using three different attributes. The encoder structure consists of 2 convolutional blocks (Conv 2D) and 1 fully-concatenated block (Dense). Each Conv2D includes a convolutional layer, a batch normalization layer, a leaky ReLU layer, a max-pooling layer, and a culling layer, facilitating depth feature extraction and preparing time series data for LSTM. The decoder receives and concatenates the output sequences, reconstructs them using Dense, convolution and upsampling (ConvTranspose 2D), and restores the image to the original input size. Finally, obtaining the probability of the target layer site through a sigmoid function.

The dataset is then trained, in this example, by performing an experiment on two real field three-dimensional seismic datasets: qiNan dataset and netherlands F3 dataset. And using the line profile and the gradient and fault attribute generated by the line profile as input training data.

The method of the invention trains 200 epochs in 4 batches of size and realizes the following in TensorFlow on the Google Colab platformhttps://colab.research.google.com). Ground truth was generated by commercial software and the results were compared to single input DCAE-LSTM.

1) QiNan dataset: it includes 401 lines, 441 Crossline lines, each with 651 samples sampled at 2 millisecond intervals, from which the necessary subset (200 [ line ] $/time $200[ Crossline ] $/time $300[ time ] samples) is extracted. Data enhancement operations such as rotation, flipping, and addition of gaussian noise are performed to mitigate overfitting. Specifically, 20 pairs were randomly selected from a total of 200 pairs of data and an additional 60 pairs were generated using the enhancement method. The enhanced data set is then divided into training and validation data sets using a 9:1 ratio.

2) F3 dataset: it is provided by SEG Wiki websitehttps://terranubis.com/ datainfo/F3-Demo-2020) Comprises 651 lines, 951 Crossline lines, time range 1,848ms, sampling interval 4ms, from which the necessary subset is extracted (300 [ lines]$400[Crossline]$136[Time]A sample). Likewise, the data enhancement operation is performed on 30 pairs of sample data sets, resulting in 120 pairs, of which 108 pairs are assigned to training and 12 pairs are used for verification.

For verification of the three-dimensional QiNan dataset, fig. 4 shows the tracking results of two-dimensional profiles of both methods.

Wherein, fig. 4 (a) is a 40 th line cross-section of single input DCAE-LSTM extraction, fig. 4 (b) is a 40 th line cross-section of single input DCAE-LSTM extraction, fig. 4 (c) is a 126 th cross-section of single input DCAE-LSTM extraction, and fig. 4 (d) is a 126 th cross-section of single input DCAE-LSTM extraction.

The rectangular areas in fig. 4 (a) and 4 (c) show some severe cross-layer phenomena, especially at faults. In contrast, in FIGS. 4 (b) and 4 (d), the MARN proposed by the method of the present invention improves the accuracy of the entire fault.

Fig. 5 is an amplitude-labeled horizon surface extracted by two methods, fig. 5 (a) is a three-dimensional horizon surface map extracted by single input DCAE-LSTM, and fig. 5 (b) is a three-dimensional horizon surface map extracted by the method of the present invention. Since the target tag is selected at the amplitude peak, the horizon surface is expected to cross the amplitude peak and remain uniformly varied throughout the process. There are some obvious discontinuities in fig. 5 (a), indicating that the extracted horizons do not follow the seismic amplitude correctly. The horizon extracted in fig. 5 (b) follows a consistent peak and is more accurate, especially at two faults, as is evident by the large number of white zero points that appear within the circle of fig. 5 (a).

Further validation of the three-dimensional F3 dataset, fig. 6 shows two complex two-dimensional seismic profile tracking results, where many local faults make the reflection highly discontinuous.

Wherein, FIG. 6 (a) is the 50 th Inline section extracted by the single input DCAE-LSTM, FIG. 6 (b) is the 50 th Inline section extracted by the method of the present invention, FIG. 6 (c) is the 150 th Crossline section extracted by the single input DCAE-LSTM, and FIG. 6 (d) is the 150 th Crossline section extracted by the method of the present invention.

In this case, the tracking result of the single input DCAE-LSTM cannot follow the laterally discontinuous reflection, and there is a severe cross-horizon phenomenon in FIGS. 6 (a) and 6 (c), especially in rectangular areas. In contrast, the MARN proposed by the method of the present invention can extract all target horizons more precisely, these horizons follow identical phases, such as peaks or valleys in FIGS. 6 (b) and 6 (d).

Fig. 7 shows one of the two extracted horizon surfaces, which is colored in an amplitude manner, fig. 7 (a) is a three-dimensional horizon surface map extracted by a single input DCAE-LSTM, and fig. 7 (b) is a three-dimensional horizon surface map extracted by the method of the present invention. In fig. 7 (a), many discontinuous structures are clearly visible. In contrast, the horizon surface shown in fig. 7 (b) reasonably follows the expected seismic amplitude peaks and is significantly more accurate.

To further verify the tracking performance of the MARN proposed by the method of the present invention, a map view is drawn in this example and the Absolute Error (AE) generated based on both methods is calculated as shown in fig. 8.

Wherein fig. 8 (a) is a map view of one horizon surface of commercial software, fig. 8 (b) is a map view of one horizon surface of the method of the present invention, and fig. 8 (c) is a map view of one horizon surface of a single input DCAE-LSTM; FIG. 8 (d) shows the AE between the commercial software and the MARN proposed by the method of the present invention, and FIG. 8 (e) shows the AE between the commercial software and the single input DCAE-LSTM.

The single input DCAE-LSTM trace results were significantly discontinuous or anomalous, especially in rectangular areas. It can be observed that the horizon map view extracted from the MARN proposed by the method of the present invention in fig. 8 (d) and 8 (e) is closer to ground truth than the result obtained with single input DCAE-LSTM.

In the embodiment, the MARN proposed by the method is fully compared with the DCAE-LSTM with single attribute input on two real three-dimensional data sets, and the MARN proposed by the method is verified through experimental results on two real field three-dimensional data sets, and compared with the DCAE-LSTM method with single input, the MARN proposed by the method has obvious improvement.

To sum up, in order to cope with serious cross-layer phenomena and realize high-precision cross-fault tracking, the method of the invention expresses the problem as a multi-attribute regression model through the proposed MARN. The time relationship between the point and the point on the target horizon is then modeled using the system state equation. Subsequently, by combining DCAE and LSTM, depth feature extraction, sequence regression analysis and data reconstruction are effectively realized. In this regard, the time-space correlation of various seismic attributes is fully utilized to achieve accurate tracking results across faults. Compared with a single-input deep learning method, the MARN provided by the method provided by the invention has strong fault crossing capability and tracking precision on two real field three-dimensional data sets.

Those of ordinary skill in the art will recognize that the embodiments described herein are for the purpose of aiding the reader in understanding the principles of the present invention and should be understood that the scope of the invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific modifications and combinations from the teachings of the present disclosure without departing from the spirit thereof, and such modifications and combinations remain within the scope of the present disclosure.

Claims

1. A three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network, the specific steps of which are as follows:

S1. Construct a multi-attribute regression network to express the three-dimensional seismic horizon tracking problem as a multi-attribute regression model;

S2. Extract depth features using the encoder structure of DCAE;

S3. Based on the deep features extracted in step S2, the sequence is used for sequence regression through a Long Short-Term Memory (LSTM) network to analyze the temporal correlation between various attribute deep features.

S4. Based on the predicted values output by LSTM in step S3, the seismic horizon image is reconstructed through the DCAE decoder structure to complete the data reconstruction.

S5. Output seismic horizon images to complete three-dimensional seismic horizon tracking.

2. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 1, wherein step S1 is specifically as follows:

The main objective of the 3D seismic horizon tracking problem is to extract target horizons from seismic data by constructing a multi-attribute regression network using a combination of DCAE and LSTM.

in, Indicates the target layer. Let _n1 , _n2 , and _n3 represent the number of seismic interpretation profiles along the main survey line, the seismic interpretation profiles along the connecting line, and the seismic interpretation profiles along the timeline, respectively.

The problem of 3D seismic horizon tracking is transformed into a multi-attribute regression model, which is mathematically represented as follows:

Where Θ represents the network weight, Let represent a well-fitted multiple regression function with Θ as the parameter. This represents multiple seismic attributes, where n represents the number of seismic attributes, and [N] is a shorthand for the set {1,2,...,N}.

A state equation is proposed, and the state information related to each seismic attribute at each time step is analyzed. The expressions for the state prediction process and the measurement process are as follows:

in, Indicates the current state by using various earthquake attributes. Compared with the previous state variables Combining these, we can obtain the system's output at the current time step t. Indicates noise in a time-varying process. Indicates measurement noise. and These represent the state transition and observation functions, respectively.

3. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 1, wherein step S2 is specifically as follows:

Given a pair of training samples and represent and represent the real ground data and the input training data, respectively, and k represents the number of profiles;

Extracting complex depth features using the encoder structure of DCAE:

in, This represents the encoder module, which extracts depth features. θ represents the implicit parameter of the DCAE network encoder.

4. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 1, wherein step S3 is specifically as follows:

By performing sequence regression using LSTM, the temporal correlation between various deep features of attributes is analyzed, and equation (2) is restated as follows:

in, Let these represent the state vector, observation vector, process noise vector, and measurement noise vector of the nth seismic attribute at time step t, respectively. Let Ψ<sub>n</sub> represent the depth characteristic of the nth seismic attribute at time step t; then the predicted response with the same parameter Ψ _<sub>n </sub> Obtained through the following methods:

Where _Ψn represents the parameters of the nth seismic attribute LSTM branch.

5. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 1, wherein step S4 is specifically as follows:

In deriving the predicted response vector Then, these vectors are concatenated and converted into a visually interpretable image using DCAE's decoder structure:

in, Represents seismic horizon image, Represents an abstract decoder function;

The system is modeled using a DCAE-LSTM architecture, and the network weights Θ, including θ and _Ψn , are updated by minimizing the loss function. Achieving output convergence in a neural network:

in, This represents an abstract decoder function.

6. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 1, wherein step S5 is specifically as follows:

The mathematical representation of the state transition at each time step is as follows:

in, Let δ(·) represent the internal state vector, and let δ(·) represent the sigmoid function. These represent the weights associated with the internal state vector; then the current state... The expression is as follows:

Where the symbol ⊙ represents the Hadman product between two vectors; then equation (4) can be rewritten as:

Where tanh represents the hyperbolic tangent function;

Finally, LSTM is used to perform mathematical abstraction on equation (4) to obtain the prediction results. Used as input to the decoding module Final output seismic horizon image Complete 3D seismic horizon tracing.