High resolution automatic recovery method based on new and old seismic data

Abstract

The application provides a high-resolution automatic recovery method based on new and old seismic data, which comprises the following steps: step 1, fine calibration of new and old seismic data; step 2, construction of a high-resolution automatic recovery three-dimensional sample database of seismic data; step 3, establishment of a high-resolution automatic recovery model based on a deep convolutional neural network; step 4, high-resolution seismic recovery model training and output; and step 5, high-resolution automatic recovery of old seismic data in the whole work area. The high-resolution automatic recovery method based on new and old seismic data directly establishes the mapping relationship between the low-resolution data body and the high-resolution data body according to the actual data based on the characteristics of high seismic resolution of new seismic data, and uses the trained high-resolution seismic automatic recovery model to perform high-resolution automatic recovery of old seismic data in the whole work area, so that the method is a new method for effectively improving seismic resolution.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a high-resolution automatic recovery method based on both new and old seismic data. Background Technology

[0002] Older seismic data covered a large area, but due to limitations in acquisition capabilities at the time, the resolution was low, and it did not clearly reflect the characteristics of complex geological bodies, failing to meet the interpretation needs of micro-structures and small fault systems. Newly acquired seismic data has smaller cell sizes and higher density, effectively improving resolution and better reflecting true geological information, but the cost is relatively high, and it has not yet achieved full coverage of the entire work area, making it difficult to meet the needs of exploration, development, and production.

[0003] For older data, traditional techniques such as deconvolution, compressed sensing, and curvelet transform are commonly used to improve resolution. These techniques are mostly based on single-channel resolution enhancement, and are affected by adjacent seismic traces, resulting in poor lateral continuity of seismic bodies after frequency upsampling. Furthermore, they involve numerous parameters requiring manual adjustment, leading to long processing times, low efficiency, and low reliability. New and old data often overlap, and there is a one-to-one mapping relationship between the two types of data. Finding effective correlation analysis methods to upgrade the resolution of old data to the level of new data will facilitate further mining and utilization of information from old data.

[0004] Chinese patent application CN201510364114.9 discloses a method for improving the resolution of seismic data, which involves pre-stack processing of seismic data with multiple development. The method first derives a multidimensional wavelet deconvolution model using multiples based on SRME and focusing transformation theory. Then, it uses an unsteady regression adaptive matched filtering method with shaping regularization to separate the high-resolution data recovered by multiples in the focusing domain, thereby achieving high-resolution conversion of the original data.

[0005] Chinese patent application CN201611259346.9 discloses a method for improving the resolution of seismic data. The specific steps are as follows: applying continuous wavelet transform (CWT) time series analysis to the original seismic record; utilizing the multi-resolution characteristics of the continuous wavelet domain, calculating the harmonic information of wavelets at each scale through harmonic analysis and inverse transforming the harmonic information back to the time domain; adding the obtained time threshold information to the original seismic record, effectively expanding the bandwidth of the seismic data.

[0006] Chinese patent application CN200710017029.0 discloses a method for improving seismic data resolution based on system identification. The method mainly includes the following steps: (i) high-resolution profile preprocessing; (ii) construction of a stratigraphic absorption system model; (iii) implementation of system identification of the stratigraphic absorption model, which includes ① the basic structure of system identification, ② the implementation of the non-parametric system model of system identification, and ③ the implementation of the parametric system model of system identification; (iv) verification of the system structure of the stratigraphic absorption model; (v) calculation of the system response of the stratigraphic absorption model; (vi) three-dimensional spatial extrapolation of the response of the stratigraphic absorption model; and (vii) implementation of high-frequency extension of ground seismic data.

[0007] The existing technologies described above are significantly different from the present invention and have failed to solve the technical problem we want to address. Therefore, we have invented a new high-resolution automatic recovery method based on both new and old seismic data. Summary of the Invention

[0008] The purpose of this invention is to provide a new method for high-resolution automatic seismic restoration based on both new and old seismic data. This method involves training and learning from both new and old seismic data to establish a three-dimensional high-resolution seismic restoration model, thereby improving the resolution of old seismic data.

[0009] The objective of this invention can be achieved through the following technical measures: a high-resolution automatic recovery method based on both new and old seismic data, comprising:

[0010] Step 1: Perform fine calibration of old and new seismic data;

[0011] Step 2: Construct a high-resolution, automatically recoverable 3D earthquake sample database;

[0012] Step 3: Establish a high-resolution autorecovery model based on a deep convolutional neural network;

[0013] Step 4: Train and output a high-resolution earthquake recovery model;

[0014] Step 5: Perform high-resolution automatic restoration of the entire seismic data area.

[0015] The objective of this invention can also be achieved through the following technical measures:

[0016] In step 1, the post-stack seismic data volumes with different accuracies processed in the two steps within the work area are compared, and the dominant seismic frequency and bandwidth information are analyzed. The old and new seismic data analyzed in the comparison are uniformly and finely labeled with the stratigraphic level. Multiple stratigraphic levels are interpreted based on the consistency of the dominant frequency and the reflection interface characteristics of the geological marker layer.

[0017] In step 2, based on the fine calibration of old and new seismic data, within the new data work area, the old and new data volumes are segmented along multiple interpreted stratigraphic levels to ensure that the segmented seismic data volumes have a certain spectral similarity within the stratigraphic segments. To meet training requirements, the old and new data within each stratigraphic segment are slidably segmented using a fixed-size three-dimensional time window. The size of the segmented three-dimensional seismic volume data is N1×N2×N3, where N1 represents the line number, N2 represents the trace number, and N3 represents the time window size. To ensure the integrity of the analysis data, the line number and trace number should ideally be kept at a fixed value, and the size of N3 should be at least no less than the length of a seismic wavelet.

[0018] In step 2, the old data segmentation data volume is used as the sample input data X, and the new data segmentation data volume at the corresponding position is used as the sample output data Y. In order to reduce the difficulty of training and the applicability of the model, the input data and output data are standardized, and the amplitude value is standardized to between -1 and 1.

[0019] In step 3, based on the high-resolution restoration network structure of the 3D deep convolutional autoencoder, the key parameters of the network are optimized, and the model is trained on the 3D seismic sample database; the deep convolutional autoencoder network is sampled to perform end-to-end high-resolution restoration.

[0020] In step 3, chunks of old data are used as input, and chunks of new data are used as output. A three-layer network is used to construct the encoding and decoding processes respectively. In the encoding process, each layer includes two 3×3×3 convolutional kernels, a Leaky ReLU activation function (with leakage correction linear unit), and a 2×2×2 max pooling layer to extract data features from the old data. In the decoding process, each layer includes a 2×2×2 upsampling layer, two 3×3×3 convolutional kernels, and a Leaky ReLU activation function (with leakage correction linear unit) to restore the feature extraction results. The layers corresponding to encoding and decoding are concatenated along the channel direction to retain richer data information.

[0021] In step 3, the network's loss function is defined using the L1 norm, ensuring that the data recovered from the old data by the network model is closest to the new data. Its specific form is as follows:

[0022]

[0023] To cut the old data into blocks of 3D volume data after standardization, The new data is a 3D volumetric data segmented after standardization. F(·) represents the forward propagation process of the deep convolutional autoencoder network model, and ||·||1 represents the 1-norm fitting.

[0024] In step 4, when the loss function value reaches its optimum, the earthquake dataset to be tested is input into the high-resolution earthquake restoration model and cross-validated with the actual new earthquake data. If the accuracy requirements are met, the high-resolution earthquake restoration model is output.

[0025] In step 4, the convolutional kernel weight parameters are randomly initialized, the Adam method is used to optimize the network parameters, and the learning rate and number of iterations are set. The old data chunks are used as input for forward propagation to calculate the high-resolution restoration result and calculate the loss function value with the corresponding new data. The error is then mapped to the modification process of the convolutional kernel weight parameters through backpropagation. After multiple iterations, the loss function value of the entire model no longer decreases, the model training is completed, the optimal weight parameters corresponding to the sample dataset are obtained, and the high-resolution earthquake restoration model is output.

[0026] In step 5, the trained automatic seismic recovery model is used to perform high-resolution automatic recovery of the old seismic data across the entire work area.

[0027] The high-resolution automatic recovery method based on both old and new seismic data in this invention utilizes deep neural networks, which possess powerful nonlinear representation capabilities. By using old data as input and new data as the overlay target, this technique establishes the correlation between old and new data, effectively overcoming human interference and avoiding the influence of manual rules, thus achieving more reliable high-resolution recovery of 3D seismic data volumes. The beneficial effects of this invention are as follows:

[0028] This invention proposes a high-resolution automatic recovery method based on both new and old seismic data. Taking advantage of the high seismic resolution of new seismic data, a mapping relationship between low-resolution and high-resolution data volumes is directly established based on actual data. A trained high-resolution automatic seismic recovery model is then used to perform high-resolution automatic recovery of old seismic data across the entire seismic area. This is a novel method that effectively improves seismic resolution. Attached Figure Description

[0029] Figure 1 This is a flowchart of a specific embodiment of the high-resolution automatic recovery method based on new and old seismic data of the present invention;

[0030] Figure 2 This is a schematic diagram of the seismic data work area in a specific embodiment 1 of the present invention;

[0031] Figure 3 This is a structural diagram of a three-dimensional deep convolutional neural network model in a specific embodiment 1 of the present invention;

[0032] Figure 4 This is a cross-sectional view of old earthquake data in a specific embodiment 2 of the present invention;

[0033] Figure 5This is a high-resolution restored profile of old earthquake data in a specific embodiment 2 of the present invention;

[0034] Figure 6 This is a seismic profile of old earthquake data in a specific embodiment 3 of the present invention;

[0035] Figure 7 This is a high-resolution restored profile of old earthquake data in a specific embodiment 3 of the present invention. Detailed Implementation

[0036] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0037] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0038] The high-resolution automatic recovery method based on both old and new seismic data of the present invention includes the following steps:

[0039] Step 1: Perform fine calibration of old and new seismic data. Compare post-stack seismic data volumes with different accuracies processed in the two phases within the work area, and analyze the dominant frequency and bandwidth information of the earthquakes;

[0040] The old and new seismic data were compared and analyzed to perform unified and detailed stratigraphic calibration, and multiple stratigraphic sets were interpreted based on the consistency of dominant frequency and the reflection interface characteristics of geological marker layers.

[0041] Step 2, Construction of a high-resolution automatic seismic recovery 3D sample database, including:

[0042] Based on the fine calibration of old and new earthquakes, within the new data work area, the old and new data volumes are segmented along multiple interpreted stratigraphic levels to ensure that the segmented earthquake data volumes have a certain spectral similarity within the stratigraphic segment.

[0043] To meet training requirements, the old and new data within each segment are divided into sliding segments using a fixed-size three-dimensional time window;

[0044] The size of the 3D seismic volume data after being segmented is N1×N2×N3, where N1 represents the line number, N2 represents the trace number, and N3 represents the time window size.

[0045] To ensure the integrity of the analysis data, the line number and track number should ideally be kept at a fixed value, and the value of N3 should be at least no less than the length of a seismic wavelet.

[0046] Use the old data segmentation data volume as the sample input data X, and use the new data segmentation data volume at the corresponding position as the sample output data Y;

[0047] To reduce the difficulty of training and improve the applicability of the model, the input and output data are standardized, and the amplitude values ​​are standardized to between -1 and 1.

[0048] Step 3: Establish a high-resolution autorecovery model based on a deep convolutional neural network, including:

[0049] Based on a high-resolution recovery network structure of 3D deep convolutional autoencoder, key network parameters are optimized, and the model is trained on a 3D seismic sample database.

[0050] A depthwise convolutional autoencoder network is used for end-to-end high-resolution restoration;

[0051] The old data chunks are used as input, and the new data chunks are used as output. A three-layer network is used to construct the encoding and decoding processes respectively.

[0052] In the encoding process, each layer unit includes two 3×3×3 convolutional kernels, a Leaky ReLU activation function (with leakage correction linear unit), and a 2×2×2 max pooling layer to complete the extraction of data features from old data;

[0053] During the decoding process, each layer unit includes 2×2×2 upsampling, two 3×3×3 convolutional kernels, and a Leaky ReLU activation function (with leakage correction linear unit) to restore the feature extraction results;

[0054] The layers corresponding to encoding and decoding are spliced ​​along the channel direction to retain richer data information;

[0055] The network's loss function is defined using the L1 norm, ensuring that the data recovered from the old data through the network model is closest to the new data. Its specific form is as follows:

[0056]

[0057] To cut the old data into blocks of 3D volume data after standardization, The new data is a 3D volumetric data segmented after standardization. F(·) represents the forward propagation process of the deep convolutional autoencoder network model, and ||·||1 represents the 1-norm fitting.

[0058] Step 4, training and output of the high-resolution earthquake recovery model, including:

[0059] When the loss function value reaches its optimum, the earthquake dataset to be tested is input into the high-resolution earthquake recovery model and cross-validated with the actual new earthquake data. If the accuracy requirements are met, the high-resolution earthquake recovery model is output.

[0060] Randomly initialize the convolutional kernel weights, use the Adam method to optimize the network parameters, and set the learning rate and number of iterations;

[0061] The old data chunks are used as input for forward propagation, the high-resolution restoration result is calculated, and the loss function value is calculated with the corresponding new data. The error is then mapped to the modification process of the convolution kernel weight parameters through backpropagation.

[0062] Through multiple iterations, the loss function value of the entire model no longer decreases, the model training is completed, the optimal weight parameters corresponding to the sample dataset are obtained, and a high-resolution earthquake recovery model is output.

[0063] Step 5: High-resolution automatic restoration of the entire seismic data area, including:

[0064] A trained automatic seismic recovery model was used to perform high-resolution automatic recovery of old seismic data across the entire seismic field.

[0065] The following are several specific embodiments of the application of the present invention.

[0066] Example 1:

[0067] In a specific embodiment 1 of the present invention, Figure 1 This is a flowchart illustrating a high-resolution automatic recovery method based on both new and old seismic data according to the present invention. Figure 1 As shown, the method includes the following steps:

[0068] The first step is to finely calibrate both old and new seismic data.

[0069] This case study first compares the old and new seismic data volumes, analyzing the dominant frequency and bandwidth information. Then, it performs fine calibration on the old and new seismic data, determining the corresponding seismic profiles based on the consistency of dominant frequencies and the reflection interface characteristics of geological marker layers. For example... Figure 2 As shown, the new seismic data volume area is part of the old seismic data volume area.

[0070] The second step is to construct a high-resolution, automatically recoverable 3D sample database of earthquake samples.

[0071] Based on precise calibration, within the new seismic data work area, the old and new data volumes are segmented along the layer segments to ensure that the segmented seismic data volumes have a certain degree of spectral similarity within the layer segments.

[0072] To meet training requirements, the old and new data within each segment are slidably divided using a fixed-size three-dimensional time window.

[0073] Assume the size of the 3D seismic volume data after being segmented is: N1×N2×N3, where N1 represents the line number, N2 represents the trace number, and N3 represents the time window size.

[0074] To ensure the integrity of the analysis data, the line number and track number should ideally be kept at a fixed value, and the value of N3 should be at least no less than the length of a seismic wavelet.

[0075] The old data segments are used as the input data X, and the corresponding segments of the new data are used as the output data Y. To reduce the difficulty of training and improve the applicability of the model, the input and output data are standardized, and the amplitude values ​​are standardized to between -1 and 1.

[0076] The third step is to establish a high-resolution autorecovery model based on a deep convolutional neural network.

[0077] Sampling depthwise convolutional autoencoders are used for end-to-end high-resolution restoration, such as... Figure 3 ;

[0078] The old data chunks are used as input, and the new data chunks are used as output. A three-layer network is used to construct the encoding and decoding processes respectively.

[0079] In the encoding process, each layer unit includes two 3×3×3 convolution kernels and a Leaky ReLU activation function (with leakage correction linear unit) to extract data features from the old data;

[0080] During the decoding process, each layer unit includes two 3×3×3 convolution kernels for deconvolution and a Leaky ReLU activation function (with leakage correction linear unit) to restore the feature extraction results;

[0081] The layers corresponding to encoding and decoding are spliced ​​along the channel direction to retain richer data information;

[0082] The network's loss function is defined using the L1 norm, ensuring that the data recovered from the old data through the network model is closest to the new data. Its specific form is as follows:

[0083]

[0084] To cut the old data into blocks of 3D volume data after standardization, The new data is a 3D volumetric data segmented after standardization. F(·) represents the forward propagation process of the deep convolutional autoencoder network model, and ||·||1 represents the 1-norm fitting.

[0085] The fourth step is to train and output a high-resolution earthquake recovery model.

[0086] When the loss function value reaches its optimum in the third step, the earthquake dataset to be tested is input into the high-resolution earthquake recovery model and cross-validated with the actual new earthquake data. If the accuracy requirements are met, the high-resolution earthquake recovery model is output.

[0087] Randomly initialize the convolution kernel weights, use the Adam method to optimize the neural network parameters, and set the learning rate and number of iterations.

[0088] The old data chunks are used as input for forward propagation to calculate the high-resolution restoration result. The loss function value is then calculated by comparing the result with the corresponding new data. The error is then mapped to the modification process of the convolution kernel weight parameters through backpropagation.

[0089] Through multiple iterations, the loss function value of the entire model no longer decreases, the model training is completed, the optimal weight parameters corresponding to the sample dataset are obtained, and a high-resolution earthquake recovery model is output.

[0090] Step 5: High-resolution automatic recovery of the entire seismic data area.

[0091] The trained automatic seismic recovery model was used to perform high-resolution automatic recovery of old seismic data across the entire seismic area.

[0092] Example 2:

[0093] In a specific embodiment 2 of the present invention, the following steps are included:

[0094] The first step is to finely calibrate both old and new seismic data.

[0095] This case study first compares the old and new earthquake data volumes to analyze the dominant frequency and bandwidth information of the earthquakes. Then, it performs fine calibration on the old and new earthquake data and determines the corresponding profiles of the old and new earthquakes based on the consistency of the dominant frequency and the reflection interface characteristics of the geological marker layer.

[0096] The second step is to construct a high-resolution, automatically recoverable 3D sample database of earthquake samples.

[0097] Based on precise calibration, within the new seismic data work area, the old and new data volumes are segmented along the layer segments to ensure that the segmented seismic data volumes have a certain degree of spectral similarity within the layer segments.

[0098] To meet training requirements, the old and new data within each segment are slidably divided using a fixed-size three-dimensional time window.

[0099] Assume the size of the 3D seismic volume data after being segmented is: N1×N2×N3, where N1 represents the line number, N2 represents the trace number, and N3 represents the time window size.

[0100] To ensure the integrity of the analysis data, the line number and track number should ideally be kept at a fixed value, and the value of N3 should be at least no less than the length of a seismic wavelet.

[0101] The old data segments are used as the input data X, and the corresponding segments of the new data are used as the output data Y. To reduce the difficulty of training and improve the applicability of the model, the input and output data are standardized, and the amplitude values ​​are standardized to between -1 and 1.

[0102] The third step is to establish a high-resolution autorecovery model based on a deep convolutional neural network.

[0103] A depthwise convolutional autoencoder network is used for end-to-end high-resolution restoration;

[0104] The old data chunks are used as input, and the new data chunks are used as output. A three-layer network is used to construct the encoding and decoding processes respectively.

[0105] In the encoding process, each layer unit includes two 3×3×3 convolution kernels and a Leaky ReLU activation function (with leakage correction linear unit) to extract data features from the old data;

[0106] During the decoding process, each layer unit includes two 3×3×3 convolution kernels for deconvolution and a Leaky ReLU activation function (with leakage correction linear unit) to restore the feature extraction results;

[0107] The layers corresponding to encoding and decoding are spliced ​​along the channel direction to retain richer data information;

[0108] The network's loss function is defined using the L1 norm, ensuring that the data recovered from the old data through the network model is closest to the new data. Its specific form is as follows:

[0109]

[0110] To cut the old data into blocks of 3D volume data after standardization, The new data is a 3D volumetric data segmented after standardization. F(·) represents the forward propagation process of the deep convolutional autoencoder network model, and ||·||1 represents the 1-norm fitting.

[0111] The fourth step is to train and output a high-resolution earthquake recovery model.

[0112] When the loss function value reaches its optimum in the third step, the earthquake dataset to be tested is input into the high-resolution earthquake recovery model and cross-validated with the actual new earthquake data. If the accuracy requirements are met, the high-resolution earthquake recovery model is output.

[0113] Randomly initialize the convolution kernel weights, use the Adam method to optimize the neural network parameters, and set the learning rate and number of iterations.

[0114] The old data chunks are used as input for forward propagation to calculate the high-resolution restoration result. The loss function value is then calculated by comparing the result with the corresponding new data. The error is then mapped to the modification process of the convolution kernel weight parameters through backpropagation.

[0115] Through multiple iterations, the loss function value of the entire model no longer decreases, the model training is completed, the optimal weight parameters corresponding to the sample dataset are obtained, and a high-resolution earthquake recovery model is output.

[0116] Step 5: High-resolution automatic recovery of the entire seismic data area.

[0117] Using a trained automatic earthquake recovery model to analyze old earthquake data ( Figure 4 The high-resolution automatic recovery results for the entire work area are used to obtain corresponding high-resolution seismic profiles (such as...). Figure 5 ).

[0118] Example 3:

[0119] In a specific embodiment 3 of the present invention, the following steps are included:

[0120] The first step is to finely calibrate both old and new seismic data.

[0121] This case study first compares the old and new earthquake data volumes to analyze the dominant frequency and bandwidth information of the earthquakes. Then, it performs fine calibration on the old and new earthquake data and determines the corresponding profiles of the old and new earthquakes based on the consistency of the dominant frequency and the reflection interface characteristics of the geological marker layer.

[0122] The second step is to construct a high-resolution, automatically recoverable 3D sample database of earthquake samples.

[0123] Based on precise calibration, within the new seismic data work area, the old and new data volumes are segmented along the layer segments to ensure that the segmented seismic data volumes have a certain degree of spectral similarity within the layer segments.

[0124] The third step is to establish a high-resolution autorecovery model based on a deep convolutional neural network.

[0125] A depthwise convolutional autoencoder network is used for end-to-end high-resolution restoration.

[0126] The fourth step is to train and output a high-resolution earthquake recovery model.

[0127] When the loss function value reaches its optimum in the third step, the earthquake dataset to be tested is input into the high-resolution earthquake recovery model and cross-validated with the actual new earthquake data. If the accuracy requirements are met, the high-resolution earthquake recovery model is output.

[0128] Step 5: High-resolution automatic recovery of the entire seismic data area.

[0129] Using a trained automatic earthquake recovery model to analyze old earthquake data ( Figure 6 The high-resolution automatic recovery results for the entire work area are used to obtain corresponding high-resolution seismic profiles (such as...). Figure 7 ).

[0130] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0131] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.

Claims

1. A high-resolution automatic recovery method based on old and new seismic data, characterized in that, Comprise: Step 1, carry out new and old seismic data fine calibration; Step 2, construct seismic high-resolution automatic recovery three-dimensional sample database, including: Based on new and old seismic fine calibration, in the scope of new data work area, along the interpreted multiple sets of horizon, the old data and new data volume are divided, and the divided seismic data volume has a certain frequency spectrum similarity in the layer section; in order to meet the training requirements, the old data and new data in each layer section are divided according to the sliding division of fixed size three-dimensional time window; the size of the three-dimensional seismic data after cutting is N1xN2xN3, N1 represents the number of lines, N2 represents the number of channels, and N3 represents the size of time window; in order to ensure the integrity of the analysis data, the line number and channel number are preferably kept as a fixed value, and the size of N3 is not less than the length of a seismic wavelet; The old data division data volume is used as sample input data X, and the new data division data volume at the corresponding position is used as sample output data Y; in order to reduce the difficulty of training and the applicability of the model, the input data and output data are standardized to-1~1; Step 3, establish high-resolution automatic recovery model based on depth convolution neural network, including: Based on three-dimensional depth convolution auto-encoding high-resolution recovery network structure, the key parameters of the network are optimized, and the model training is carried out on the three-dimensional seismic sample database; the sampling depth convolution auto-encoding network is used for end-to-end high-resolution recovery; The cutting data of old data is used as input, and the cutting data of new data is used as output, and 3 layers of network are used to construct coding and decoding process respectively; in the coding process, each layer unit includes 2 3x3x3 convolution kernels, leaky ReLU activation function with leakage correction and 2x2x2 maximum pooling layer, which completes the extraction of data characteristics in old data; in the decoding process, each layer unit includes 2x2x2 up sampling, 2 3x3x3 convolution kernels and leaky ReLU activation function with leakage correction, which completes the recovery of feature extraction results; the layers corresponding to coding and decoding are spliced along the channel direction to retain more rich data information; Step 4, high-resolution seismic recovery model training and output; Step 5, carry out high-resolution automatic recovery of old seismic data in the whole work area.

2. The method of claim 1, wherein, In step 1, the different precision poststack seismic data volume of the two times in the comparison work area is compared, the seismic main frequency and bandwidth information is analyzed; the new and old seismic data after comparison and analysis are fine calibrated, and multiple horizons are interpreted according to the consistency of main frequency and the reflection interface characteristics of geological marker layer.

3. The method of claim 1, wherein the method is characterized by, In step 3, the loss function of the network is defined as L 1 The norm is defined to satisfy that the data recovered by the network model from the old data is closest to the new data, and the specific form is: ； standardized old data slice three-dimensional volume data, standardized new data slice three-dimensional volume data, representing a forward propagation process of a deep convolutional auto-encoder network model, representing a 1-norm fit.

4. The method of claim 1, wherein, In step 4, when the loss function value reaches the optimum, the test seismic data set is input into the high-resolution seismic recovery model to cross validate with the actual new seismic data, and if the accuracy requirement is met, the high-resolution seismic recovery model is output.

5. The method of claim 4, wherein, In step 4, the convolution kernel weight parameters are randomly initialized, the Adam method is selected to optimize the network parameters, the learning rate and the iteration number are set, the old data cut data is taken as the input for forward propagation, the high-resolution recovery result is calculated, and the loss function value is calculated with the corresponding new data; through back propagation, the error is mapped to the modification process of the convolution kernel weight parameters; Through multiple iterations, the loss function value of the entire model no longer decreases, the model training is completed, the optimal weight parameters corresponding to the sample data set are obtained, and the high-resolution seismic recovery model is output.

6. The method of claim 1, wherein, In step 5, the trained seismic automatic recovery model is used for full-area high-resolution automatic recovery of old seismic data.

Images