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

Abstract

The application discloses a kind of three-dimensional seismic horizon intelligent tracking methods based on multiple seismic attribute regression network, constructs a multiple attribute regression network, three-dimensional seismic horizon tracking problem is described as a multiple attribute regression model, then the encoder structure of DCAE is extracted as input sequence with depth feature, sequence regression analysis is carried out by LSTM, the predicted value output by LSTM is reconstructed into seismic horizon image by the decoder structure of DCAE, data reconstruction is completed, finally seismic horizon image is output, and three-dimensional seismic horizon tracking is completed.The method of the application introduces system state equation to establish time relationship, captures time dynamics and improves the accuracy of regression analysis, extracts depth features by DCAE and as input sequence, so that LSTM can effectively simulate the space-time relationship between multiple seismic attributes and target horizon, compared with directly using LSTM, DCAE-LSTM structure can fully exert space-time correlation, and multiple seismic attributes are used to achieve accurate and robust tracking results.

Description

TECHNICAL FIELD

[0001] The present application 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

[0002] Horizon tracking plays a crucial role in seismic interpretation, as it is an important and time-consuming challenging task. Accurate and reliable seismic horizon tracking is essential for determining the structure of subsurface geological structures and stratigraphic framework. Therefore, many horizon tracking methods using seismic attributes have been developed, and have achieved gratifying results, such as the least squares method based on local slope and the method based on waveform similarity.

[0003] However, a summary of these methods shows that the basic challenge is how to effectively select and use seismic attributes to improve the accuracy of horizon tracking, such as across discontinuous strata (such as faults) and along laterally continuous reflectors (such as local slope). The core challenge is to effectively use appropriate seismic attributes to improve tracking accuracy, and many existing model-based or deep learning (DL) based tracking methods rely on mathematical or data-driven mapping relationships between seismic data and target layers, both of which only use seismic data or a single attribute, which can lead to discontinuous and inaccurate tracking.

[0004] Current seismic attribute-based horizon tracking methods can be roughly divided into three categories: methods based on instantaneous phase, waveform similarity and local slope. The method based on instantaneous phase involves initial unwrapping of instantaneous phase to obtain a representative relative geologic time (RGT), and then extracting contour outlines as horizons. In these waveform similarity-based methods, starting from a seed point, recursively track the position with the greatest waveform similarity. Most of the local slope-based methods use least squares to estimate the horizon using local reflection slope. However, these methods require a strict mathematical mapping relationship between seismic attributes and target horizons, and rely on certain model assumptions, such as least squares (linear assumption).

[0005] Unlike the model-driven methods described above, data-driven methods based on deep learning (DL) try to learn the direct mapping between seismic attributes and the corresponding target horizon intuitively without the need for hand-crafted assumptions. As a typical DL network, deep convolutional neural network (CNN) can directly map seismic data to the target horizon through supervised training. In terms of indirect mapping, CNN can automatically obtain 2D RGT images from 2D seismic images through training, and then extract seismic horizons from RGT contours. Unlike using only single seismic data, the dual-branch network simultaneously takes the original seismic data and its converted time-frequency domain data as input, which can more comprehensively map the target horizon. However, these DL-based methods mainly rely on seismic data or a single related attribute, lack the guidance of other attributes, and therefore may not be able to correlate the reflections on both sides of the fault in the absence of fault attributes. SUMMARY

[0006] To solve the above technical problems, the present application proposes a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network, proposes a multi-attribute regression network (MARN) for three-dimensional seismic horizon tracking, uses multiple seismic attributes to achieve accurate and robust tracking results, constructs a multi-attribute regression model, then introduces a system state equation to establish a time relationship, and designs a deep convolutional autoencoder (DCAE) to automatically learn spatial correlation from attributes to extract deep features.

[0007] The technical scheme adopted by the present application is as follows: a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network, the specific steps are as follows:

[0008] S1, constructing a multi-attribute regression network, expressing the three-dimensional seismic horizon tracking problem as a multi-attribute regression model;

[0009] S2, extracting deep features through the encoder structure of the DCAE;

[0010] S3, taking the deep features extracted in step S2 as input sequences, performing sequence regression through a long short-term memory network (LSTM), and analyzing the time correlation between various attribute deep features;

[0011] S4, based on the predicted value output by the LSTM in step S3, reconstructing the seismic horizon image through the decoder structure of the DCAE to complete data reconstruction;

[0012] S5, outputting the seismic horizon image to complete three-dimensional seismic horizon tracking.

[0013] Further, the step S1 is specifically as follows:

[0014] The multi-attribute regression network is constructed by combining DCAE and LSTM, and the main target of the three-dimensional seismic horizon tracking problem is to extract the target horizon in seismic data

[0015] wherein, represents the target horizon, represents a real space, and n1, n2 and n3 represent the number of seismic interpretation profiles along the main line, the contact line and the time line respectively.

[0016] The three-dimensional seismic horizon tracking problem is converted into a multi-attribute regression model, which is mathematically expressed as:

[0017]

[0018] wherein, Θ represents the network weight, represents a fitted multi-regression function with Θ as a parameter, represents a plurality of seismic attributes, n represents the number of seismic attributes, and [N] represents the abbreviation of the set {1, 2,..., N}.

[0019] The state equation is proposed, and the state information related to each seismic attribute in each time step is analyzed, and the state prediction process and the measurement process are expressed as follows:

[0020]

[0021] wherein, represents the current state, and various seismic attributes are combined with the previous state variable to obtain the output of the system at the current time step t represents a time-varying process noise, represents a measurement noise, and represent state transition and observation functions respectively.

[0022] Further, the step S2 is specifically as follows:

[0023] A pair of training samples and represent and respectively represent the ground truth and the input training data, and k represents the number of profiles.

[0024] The encoder structure of DCAE is used to extract complex deep features:

[0025]

[0026] wherein, denotes an encoder module that extracts deep features denotes the hidden parameters of the DCAE network encoder.

[0027] Further, the step S3 is specified as follows:

[0028] The sequence regression is performed by LSTM to analyze the time correlation between various attribute deep features, and formula (2) is re-expressed as:

[0029]

[0030] wherein, denote the state vector, observation vector, process noise vector and measurement noise vector of the nth seismic attribute at time step t, respectively, denotes the deep feature of the nth seismic attribute at time step t. The predicted response n with the same parameters Ψ is obtained by:

[0031]

[0032] wherein, Ψ n denotes the parameters of the LSTM branch of the nth seismic attribute.

[0033] Further, the step S4 is specified as follows:

[0034] After deriving the predicted response vector , these vectors are concatenated and converted into visually interpretable images by the decoder structure of the DCAE:

[0035]

[0036] wherein, denotes the seismic horizon image, denotes an abstract decoder function.

[0037] The system is modeled using the DCAE-LSTM architecture, and the network weights Θ, including θ and Ψ n , are updated by minimizing the loss function to achieve the convergence of the output of the neural network:

[0038]

[0039] wherein, denotes an abstract decoder function.

[0040] Further, the step S5 is specified as follows:

[0041] The mathematical representation of the state transition of each time step is as follows:

[0042]

[0043] wherein, represents the internal state vector, and δ(·) represents a sigmoid function, respectively represent weights related to the internal state vector. Then the current state is expressed as follows:

[0044]

[0045] wherein, the symbol ⊙ represents the Hadamard product between two vectors. Then formula (4) can be rewritten as:

[0046]

[0047] wherein, tanh represents a hyperbolic tangent function.

[0048] Finally, the prediction result obtained by mathematically abstracting formula (4) by the LSTM is used as the input of the decoding module to finally output the seismic horizon image to complete the three-dimensional seismic horizon tracking.

[0049] The method of the present application firstly constructs a multi-attribute regression network, expresses the three-dimensional seismic horizon tracking problem as a multi-attribute regression model, then extracts deep features through the encoder structure of the DCAE and takes them as input sequences, performs sequence regression analysis through the LSTM, reconstructs the seismic horizon image through the prediction value output by the LSTM and the decoder structure of the DCAE, completes data reconstruction, finally outputs the seismic horizon image, and completes the three-dimensional seismic horizon tracking. The method of the present application introduces a system state equation to establish the time relationship, capture the time dynamics and improve the accuracy of the regression analysis, extracts deep features through the DCAE and takes them as input sequences, so that the LSTM can effectively simulate the space-time relationship between multiple seismic attributes and the target horizon, compared with directly using the LSTM, the DCAE-LSTM structure can fully play the space-time correlation, and utilize multiple seismic attributes to realize accurate and robust tracking results. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network of the present application.

[0051] Figure 2 is a specific implementation diagram of the MARN in the embodiment of the present application.

[0052] Figure 3This is a comparison diagram of the original seismic data, corresponding local slope, and fault attributes in an embodiment of the present invention.

[0053] Figure 4 This is a comparison chart of the two-dimensional profile tracking results of the method of the present invention and the single-input DCAE-LSTM method in the QiNan dataset in the embodiments of the present invention.

[0054] Figure 5 This is a comparison image of the three-dimensional layer surface extracted by the method of the present invention and the single-input DCAE-LSTM method in an embodiment of the present invention.

[0055] Figure 6 This is a comparison chart of the two-dimensional profile tracking results of the method of the present invention and the single-input DCAE-LSTM method in the F3 dataset of the present invention embodiments.

[0056] Figure 7 This is a comparison image of the three-dimensional horizon surface extracted by the method of the present invention and the single-input DCAE-LSTM method in an embodiment of the present invention.

[0057] Figure 8 This is a comparison chart of the absolute error (AE) generated by the method of the present invention and the single-input DCAE-LSTM method in the embodiments of the present invention. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 The flowchart of a three-dimensional seismic horizon intelligent tracking method based on a multi-seismic attribute regression network of the present invention is shown below. The specific steps are as follows:

[0060] S1. Construct a multi-attribute regression network to express the three-dimensional seismic horizon tracking problem as a multi-attribute regression model;

[0061] S2. Extract depth features using the encoder structure of DCAE;

[0062] 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.

[0063] 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.

[0064] S5. Output seismic horizon images to complete three-dimensional seismic horizon tracking.

[0065] In this embodiment, step S1 is specifically as follows:

[0066] A multi-attribute regression network is constructed by combining DCAE and LSTM, and the main goal of the three-dimensional seismic horizon tracking problem is to extract the target horizon in seismic data

[0067] wherein, represents the target horizon, represents a real space, and n1, n2 and n3 represent the number of seismic interpretation profiles along the main survey line, the contact line seismic interpretation profile and the time line seismic interpretation profile respectively.

[0068] In view of the higher tracking result accuracy obtained by using various seismic attributes, the problem is converted into a multi-attribute regression model, which is mathematically expressed as:

[0069]

[0070] wherein, Θ represents a network weight, represents a fitted multi-regression function with Θ as a parameter, represents a plurality of seismic attributes, n represents the number of seismic attributes, and [N] represents a shorthand for the set {1, 2,..., N}.

[0071] In order to utilize the time correlation between the target horizon points, a state equation is proposed for analyzing the state information related to each seismic attribute in each time step, which involves a state prediction process and a measurement process:

[0072]

[0073] wherein, represents a current state, and various seismic attributes are combined with previous state variables to obtain the output of the system at the current time step t represents a time-varying process noise, represents a measurement noise, and represent state transition and observation functions respectively.

[0074] In view of the ability of LSTM to establish long-term dependencies and capture key time dynamics, LSTM is used to identify the system.

[0075] It should not be ignored that a plurality of seismic attributes are highly related to the target horizon point in space, such as discontinuous lateral variation and faults. In order to more comprehensively grasp the spatio-temporal mapping relationship between different seismic attributes and the target horizon point, the MARN proposed by the method of the present application utilizes the DCAE-LSTM architecture, which is based on DCAE and LSTM.

[0076] In the present embodiment, the step S2 is as follows:

[0077] Deep feature extraction:

[0078] A pair of training samples is given and represent and respectively, and k represents the number of profiles.

[0079] Subsequently, complex deep features are extracted by using the encoder structure of the DCAE:

[0080]

[0081] wherein, represents an encoder module, and deep features are extracted θ represents the hidden parameters of the DCAE network encoder.

[0082] In the present embodiment, the step S3 is as follows:

[0083] Sequence regression analysis:

[0084] The sequence regression is performed by using the LSTM to analyze the time correlation between various attribute deep features, and thus formula (2) is re-expressed as:

[0085]

[0086] wherein, respectively represent the state vector, the observation vector, the process noise vector and the measurement noise vector of the n-th seismic attribute at the time step t, represents the deep feature of the n-th seismic attribute at the time step t. Thus, the predicted response n with the same parameters Ψ is obtained by:

[0087]

[0088] wherein, Ψ n represents the parameters of the n-th seismic attribute LSTM branch.

[0089] In the present embodiment, the step S4 is as follows:

[0090] Data reconstruction:

[0091] After the predicted response vector is derived, these vectors are concatenated and converted into visually interpretable images by using the decoder structure of the DCAE:

[0092]

[0093] wherein, denotes the seismic horizon image, denotes an abstract decoder function.

[0094] The purpose of system modeling using the DCAE-LSTM architecture is to update the network weights Θ, which includes θ and Ψ n , aiming to achieve the convergence of the output of the neural network by minimizing the loss function

[0095]

[0096] wherein, denotes an abstract decoder function.

[0097] By simultaneous optimization of θ and Ψ n , formula (7) involves deep feature extraction, sequence regression analysis and data reconstruction, aiming to theoretically achieve a global minimum value. Figure 2 The specific implementation diagram of the MARN proposed for the method of the present application.

[0098] In the present embodiment, the step S5 is specifically as follows:

[0099] The DCAE is trained to find the optimal weights and other coefficients of the neural network. Therefore, the focus is on solving Figure 2 An LSTM unit is described in detail in the above-mentioned paper, and the mathematical representation of the state transition of each time step is as follows:

[0100]

[0101] wherein, denotes the internal state vector, and δ(·) denotes the sigmoid function, denotes the weight related to the internal state vector, respectively. Then, the current state is obtained by the following way:

[0102]

[0103] wherein, the symbol ⊙ denotes the Hadamard product between two vectors. Therefore, formula (4) can be rewritten as:

[0104]

[0105] wherein, tanh denotes the hyperbolic tangent function.

[0106] Finally, the prediction result obtained by mathematically abstracting formula (4) by the LSTM is used as the input of the decoding module​ Final output seismic horizon image The three-dimensional seismic horizon tracking is completed.

[0107] The present application also provides a method for further verifying the present application of Example 2, as follows:

[0108] First, attribute selection is performed, in order to ensure correct tracking of horizons across faults, and time samples belonging to the same horizon exhibit similar local slopes, while considering original seismic data, corresponding local slopes and fault attributes, such as Figure 3 as shown.

[0109] wherein, Figure 3 (a) is an original seismic data map, Figure 3 (b) is a local slope attribute map, Figure 3 (c) is a fault attribute map, Figure 3 (d) is a result map using only original seismic data, Figure 3 (e) is a result map using three seismic attributes.

[0110] The cross-horizon phenomenon is very obvious in the rectangular area highlighted in Figure 3 (d), mainly in the lateral discontinuous reflection domain, which coexists with local slope and fault attributes in Figure 3 (b) and 3(c). Figure 3 This indicates the necessity of using these three attributes.

[0111] Given the input using three different attributes, the parameter N is set to 3. The encoder structure consists of 2 convolutional blocks (Conv2D) and 1 fully connected block (Dense). Each Conv2D includes a convolutional layer, a batch normalization layer, a leaky ReLU layer, a max pooling layer and a dropout layer, facilitating deep feature extraction and preparing time series data for LSTM. The decoder receives the output sequence and concatenates it, applies Dense, convolution and upsampling (ConvTranspose2D) for reconstruction, restoring the image to the original input size. Finally, the probability of the target horizon point is obtained by the sigmoid function.

[0112] Then the training dataset, in this embodiment, two real field three-dimensional seismic datasets are experimented: QiNan dataset and Dutch F3 dataset. The Inline profile and the generated slope and fault attributes thereof are used as input training data.

[0113] The method of the present application is trained in 4 batches of size 200 epochs, and implemented in TensorFlow on the Google Colab platform https: / / colab.research.google.com). The ground truth is generated by commercial software and the results are compared with the single-input DCAE-LSTM.

[0114] 1) QiNan dataset: It contains 401 Inline lines, 441 Crossline lines, and 651 samples per trace with 2ms interval, from which the necessary subset (200[Inline] $ 200[Crossline] $ 300[Time] samples) is extracted. Data augmentation operations such as rotation, flipping, and adding Gaussian noise are performed to mitigate overfitting. Specifically, 20 pairs of data are randomly selected from the total of 200 pairs, and 60 additional pairs are generated using the augmentation method. Then the augmented dataset is divided into training and validation datasets with a ratio of 9:1.

[0115] 2) F3 dataset: It is publicly available from the SEG Wiki website ( https: / / terranubis.com / datainfo / F3-Demo-2020 ) and contains 651 Inline lines, 951 Crossline lines, and a time range of 1,848ms with a sampling interval of 4ms, from which the necessary subset (300[Inline] $ 400[Crossline] $ 136[Time] samples) is extracted. Similarly, data augmentation operations are performed on a dataset of 30 pairs of samples, resulting in 120 pairs, of which 108 are assigned to training and 12 to validation.

[0116] The three-dimensional QiNan dataset is validated, Figure 4 The tracking results of two two-dimensional profiles of the two methods are shown.

[0117] Among them, Figure 4 (a) is the 40th Inline profile extracted by the single-input DCAE-LSTM, Figure 4 (b) is the 40th Inline profile extracted by the method of the present application, Figure 4 (c) is the 126th Crossline profile extracted by the single-input DCAE-LSTM, Figure 4 (d) is the 126th Crossline profile extracted by the method of the present application.

[0118] Figure 4 The rectangular areas in (a) and 4(c) exhibit some serious cross-layer phenomena, especially at the fault. In contrast, Figure 4 (b) and 4(d), the MARN proposed by the method of the present application improves the accuracy of the entire fault.

[0119] Figure 5 are the horizon surfaces marked with amplitude extracted by the two methods, Figure 5(a) is a 3D horizon surface map extracted by the single-input DCAE-LSTM, Figure 5 (b) is a 3D horizon surface map extracted by the method of the present invention. Since the target label is selected at the amplitude peak, the horizon surface is expected to pass through the amplitude peak and keep consistent variation throughout the process. Figure 5 There are some obvious discontinuous structures in (a), indicating that the extracted horizon does not correctly follow the seismic amplitude. Figure 5 The horizon extracted in (b) follows consistent peaks and is more accurate, especially at the two faults, Figure 6 This is more obvious in (a) where a large number of white zero points appear in the circle.

[0120] The 3D F3 dataset is also validated, Figure 6 Two complex 2D seismic profile tracking results are shown, where many local faults make the reflection highly discontinuous.

[0121] where, Figure 6 (a) is the 50th Inline profile map extracted by the single-input DCAE-LSTM, Figure 6 (b) is the 50th Inline profile map extracted by the method of the present invention, Figure 6 (c) is the 150th Crossline profile map extracted by the single-input DCAE-LSTM, Figure 6 (d) is the 150th Crossline profile map extracted by the method of the present invention.

[0122] In this case, the tracking result of the single-input DCAE-LSTM cannot follow the laterally discontinuous reflections, in Figure 6 There are serious cross-horizon phenomena in (a) and 6(c), especially in the rectangular area. In contrast, the MARN proposed by the method of the present invention can more accurately extract all the target horizons, which follow consistent stages, such as Figure 7 the wave peaks or troughs in (b) and 6(d).

[0123] Figure 7 One of the horizon surfaces extracted by the two methods is shown, which is colored in the amplitude way, Figure 7 (a) is a 3D horizon surface map extracted by the single-input DCAE-LSTM, Figure 7 (b) is a 3D horizon surface map extracted by the method of the present invention. In Figure 7 In (a), many discontinuous structures are clearly visible. In contrast, Figure 8 The horizon surface shown in (b) reasonably follows the expected seismic amplitude peak and is obviously more accurate.

[0124] To further verify the tracking performance of the MARN proposed by the method of the present application, the map view is drawn in this embodiment, and the absolute error (AE) based on the two methods is calculated, as shown in Figure 8

[0125] wherein, Figure 8 (a) is a map view of a horizon surface of a commercial software, Figure 8 (b) is a map view of a horizon surface of the method of the present application, Figure 8 (c) is a map view of a horizon surface of a single-input DCAE-LSTM; Figure 8 (d) is the AE between the commercial software and the MARN proposed by the method of the present application, Figure 8 (e) is the AE between the commercial software and the single-input DCAE-LSTM.

[0126] The tracking results of the single-input DCAE-LSTM have obvious discontinuity or abnormality, especially in the rectangular area. It can be observed that, ​ (d) and 8(e) the horizon map view extracted from the MARN proposed by the method of the present application is closer to the ground truth than the results obtained by the single-input DCAE-LSTM.

[0127] This embodiment fully compares the MARN proposed by the method of the present application with the single-attribute input DCAE-LSTM on two real three-dimensional data sets, and verifies that the MARN proposed by the method of the present application has a significant improvement compared with the single-input DCAE-LSTM method through the experimental results on two real field three-dimensional data sets.

[0128] In summary, in order to deal with the serious cross-layer phenomenon and realize high-precision cross-fault tracking, the method of the present application formulates the problem as a multi-attribute regression model through the proposed MARN. Then, the system state equation is used to simulate the time relationship between the point and the point on the target horizon. Subsequently, the DCAE and LSTM are combined to effectively realize deep feature extraction, sequence regression analysis and data reconstruction. In this regard, the time-space correlation of multiple seismic attributes is fully utilized to realize accurate cross-fault tracking results. Compared with the single-input deep learning method, the MARN proposed by the method of the present application shows strong cross-fault ability and tracking accuracy on two real field three-dimensional data sets.

[0129] ​Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and that the inventive principles are not limited to these particular embodiments. Other variations and modifications can be made to the embodiments without departing from the spirit and scope of the inventive principles.

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, characterized in that, The specific steps of S1 are 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. Represents the space of real numbers. , and These represent the number of seismic interpretation profiles along the main survey line, seismic interpretation profiles along the connecting line, and 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: （1）； in, Indicates network weights, Indicates a For the parameters, a well-fitted multiple regression function. Indicates multiple earthquake attributes, Indicates the number of earthquake attributes. express Short for set; 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: （2）； in, Indicates the current state by using various earthquake attributes. Compared with the previous state variables Combining these, we can derive the system's steps at the current time. Output ; 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 2, characterized in that, Step S2 is as follows: Given a pair of training samples and ,represent and , representing real ground data and input training data, respectively. Indicates the number of cross sections; Extracting complex depth features using the encoder structure of DCAE: （3）； in, This represents the encoder module, which extracts depth features. , This represents the implicit parameters 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 3, characterized in that, Step S3 is 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: （4）； in, They represent the first Seismic attributes in time steps The state vector, observation vector, process noise vector, and measurement noise vector. Indicates the first Seismic attributes in time steps The depth features; then they have the same parameters. Predicted response Obtained through the following methods: （5）； in, Indicates the first Parameters of each seismic attribute LSTM branch.

5. The intelligent three-dimensional seismic horizon tracking method based on a multi-seismic attribute regression network according to claim 4, characterized in that, Step S4 is 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: （6）； in, Represents seismic horizon image, Represents an abstract decoder function; System modeling and network weight updates are performed using the DCAE-LSTM architecture. including and By minimizing the loss function Achieving output convergence in a neural network: （7）； 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 5, characterized in that, Step S5 is as follows: The mathematical representation of the state transition at each time step is as follows: （8）； in, Represents the internal state vector. This represents the sigmoid function. These represent the weights associated with the internal state vector; then the current state... The expression is as follows: (9)； Among them, symbols Let the Hadman product between two vectors be represented; then equation (4) can be rewritten as: (10)； in, 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 The final output is a seismic horizon image. Complete three-dimensional seismic horizon tracking.

Images