A method for making earthquake data labels for deep learning

Abstract

The present invention belongs to the field of exploration geophysics technology and relates to a method for producing seismic data labels for deep learning. The method comprises the following steps: Step 1. Loading seismic data and its existing interpretation results, and cutting and processing the seismic data; Step 2. Determining the position in the seismic data according to the interpretation results and assigning and labeling; Step 3. Slicing the seismic data. Slicing the seismic data, performing local transformation on the sliced label data body, and realizing the expansion of the training samples. The production method of the present invention can effectively improve the efficiency of label production with the help of existing manual interpretation results. The slicing operation can reduce redundant background information, and the local transformation can realize sample expansion. The method of the present invention can realize efficient labeling of seismic labels.

Description

Technical Field

[0001] The present invention belongs to the technical field of exploration geophysics and relates to a method for producing seismic data labels for deep learning. Background Art

[0002] Deep learning is a new technology that has developed rapidly in recent years. By constructing complex neural networks, it can effectively build a nonlinear mapping relationship from input to output. Therefore, it is widely used in medical imaging, unmanned driving, natural language processing and other fields.

[0003] Deep learning can be broadly categorized into supervised and unsupervised approaches, depending on whether or not expert labels are available. Unsupervised approaches lack the prior constraints of expert labels, while supervised approaches use expert labels to guide model performance. While theoretically, unsupervised approaches have a higher upper limit, supervised approaches currently offer more promising results.

[0004] In the field of exploration geophysics, supervised deep learning can be used for fault identification, first arrival picking, and seismic facies delineation. Appropriate model training can significantly save manpower and resources, reducing the workload for processing and interpretation personnel. Furthermore, well-trained deep learning networks can effectively avoid subjective errors compared to human interpreters, improving efficiency while also achieving considerable accuracy.

[0005] The effectiveness of supervised deep learning network models depends on the training process. An ideal training process requires good training samples and labels, as well as appropriate loss functions, optimization techniques, and hyperparameters. The quality of the training samples and labels is paramount. Training samples must be sufficient in number and contain as much potential feature information as possible. The corresponding sample labels must be accurate and have clear geological or geophysical meaning.

[0006] However, the particularity of geophysical data makes it difficult to produce labels for supervised deep learning, especially for supervised deep learning labels of 3D seismic data. It is unrealistic to manually label large amounts of data pixel by pixel.

[0007] Chinese patent application CN112034512A discloses a method for detecting discontinuities in seismic data based on a deep learning model. The method comprises dividing seismic data into three categories: simple, medium, and difficult, and processing the seismic data differently according to different categories. An initial deep learning model is established by generating initial data labels for simple seismic data, and the initial deep learning model is trained using the initial data labels for simple seismic data to obtain a final deep learning model. The seismic wave signal to be detected is substituted into the final deep learning model, and the discontinuity of the seismic wave signal to be detected is detected.

[0008] Chinese invention patent CN111562611B discloses a semi-supervised deep learning seismic data inversion method driven by wave equations, which can realize a deep learning inversion network for some seismic data that lacks a corresponding geological model. First, a convolution-fully connected network is used to enhance the seismic data based on the characteristics of pre-stack seismic data, and the mapping relationship between the seismic data and the underground multi-layer medium model is completed by extracting the feature map and finally obtaining the geological velocity model; at the same time, the wave equation is added to the network structure, and for seismic data without a corresponding geological model, the velocity loss function is replaced by the data loss function, physical laws are introduced, and a semi-supervised learning strategy is implemented. Through the semi-supervised deep learning seismic data inversion network, the inversion effect of the deep learning network is improved when there is less labeled data.

[0009] However, there is currently no method for efficiently labeling seismic data for deep learning. Summary of the Invention

[0010] The main purpose of the present invention is to provide a method for producing seismic data labels for deep learning. With the help of existing manual interpretation results, the efficiency of label production can be effectively improved. The cutting process can reduce redundant background information, and the local transformation can achieve sample expansion, thereby realizing efficient annotation of seismic labels.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] The present invention provides a method for preparing earthquake data labels for deep learning, comprising the following steps:

[0013] Step 1. Load the seismic data and its existing interpretation results, and perform segmentation processing on the seismic data;

[0014] Step 2. Determine the location in the seismic data based on the interpretation results and assign values and mark them;

[0015] Step 3: Slice the seismic data and perform local transformation on the sliced labeled data volume to expand the training samples.

[0016] Furthermore, in step 1, the loaded seismic data includes post-stack amplitude data and various seismic attributes derived therefrom, and can be either single-channel data or multi-channel composite data.

[0017] Furthermore, in step 1, the existing interpretation results include horizon interpretation information, fault interpretation information, and salt dome interpretation information.

[0018] Furthermore, in step 1, the minimum data volume that envelops all interpretation results is obtained by traversing and cutting it from the original data volume, thereby removing as much redundant background information as possible and reducing the imbalance between the target class and the background class.

[0019] Furthermore, in step 2, the seismic data is labeled at the pixel level, and different values are assigned to the seismic data to represent different background classes. Multiple background classes need to be distinguished from each other and strictly distinguished from the target class.

[0020] Furthermore, in step 2, the unexplained areas are masked and given outlier values, and are ignored in subsequent training.

[0021] Furthermore, in step 3, the label data volume is sliced according to the input requirements of the model, wherein the three-dimensional model corresponds to rectangular parallelepiped slices of the same size, and the two-dimensional model corresponds to rectangular slices of the same size.

[0022] Furthermore, the local transformation in step 3 includes rotation, inversion, replication and local deformation, thereby achieving the effect of data enhancement and realizing the expansion of training samples.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] The present invention provides a method for producing seismic data labels for deep learning, filling a gap in the prior art. The production method of the present invention can achieve efficient labeling of seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A schematic flow chart of a method for producing seismic data labels for deep learning according to an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of the original three-dimensional seismic data volume provided in Example 1 of the present invention;

[0027] Figure 3 A schematic diagram of the original three-dimensional seismic data volume provided in Example 1 of the present invention;

[0028] Figure 4 This is a schematic diagram of the cut 3D seismic data volume according to Example 1 of the present invention;

[0029] Figure 5 A schematic diagram of a three-dimensional label data volume corresponding to a three-dimensional data volume after replacement is completed in an embodiment of the present invention;

[0030] Figure 6 Comparison diagram of the deformed 3D data volume and the undeformed 3D data volume provided in Example 1 of the present invention: (a) 3D seismic data volume before deformation, (b) 3D labeled data volume before deformation, (c) 3D seismic data volume after deformation, and (d) 3D labeled data volume after deformation;

[0031] Figure 7The two-dimensional seismic label slice provided in Example 1 of the present invention;

[0032] Figure 8 This is a schematic diagram of the original 3D seismic data volume provided in Example 2 of this specification;

[0033] Figure 9 A schematic diagram of a three-dimensional label data volume corresponding to a three-dimensional data volume after replacement provided in Example 2 of this specification;

[0034] Figure 10 This is the two-dimensional seismic label slice provided in Example 2 of this specification. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations and / or combinations thereof.

[0037] Figure 1 This is a flow chart of a method for making earthquake labels for deep learning of the present invention. The specific steps are as follows:

[0038] Step 101: Load the training sample seismic data and interpretation results. The input seismic data includes post-stack amplitude data and various seismic attributes derived therefrom. It can be either single-channel data or multi-channel composite data. The input previous interpretation results include target interpretation results such as horizon interpretation information, fault interpretation information, and salt dome interpretation information. In the following description, the input seismic data is single-channel post-stack amplitude data, and the input previous interpretation results are interpretation information for multiple horizons.

[0039] Step 102: After the seismic data and interpretation results are loaded, all input horizons are traversed to obtain the minimum data volume containing all horizons. The traversal method is:

[0040] X∈[min(x i ),max(x i )]

[0041] Y∈[min(y i ),max(y i )]

[0042] Z∈[min(z i ),max(z i )]

[0043] Where X, Y, Z are the ranges of the smallest data volume along the x, y, and z axes, and xi, yi, and zi are the x, y, and z values of the interpretation points in the interpretation results. These are extracted from the larger work area data volume as a new 3D data volume.

[0044] Step 2: Determine the location in the seismic data and perform value assignment based on the interpretation results. The specific method is as follows:

[0045]

[0046] When there are some seismic traces in the data volume without manual interpretation results, the unexplained areas can be masked and given outliers, and ignored in subsequent training. The specific method is as follows:

[0047]

[0048] Step 3: Slice the seismic data. The slicing process is performed according to the input requirements of the model. The three-dimensional model corresponds to rectangular slices of the same size, and the two-dimensional model corresponds to rectangular slices of the same size.

[0049] Slicing the obtained three-dimensional data volume can obtain corresponding two-dimensional labels, which is particularly applicable when the required labels are all two-dimensional and concentrated in one area.

[0050] Perform local transformations on the sliced seismic data to augment the training sample. There are no specific requirements for the local transformations applied to the labeled information; rotation, inversion, replication, and local deformation can all achieve data augmentation. There are many methods for deformation, the most commonly used being Gaussian function deformation, as follows:

[0051]

[0052] where Z shift is the displacement of the point in the data volume in the Z direction, Z is the Z-axis coordinate value of the point, that is, when traveling in both directions, Z max is its maximum value; X is the X-axis coordinate value of the point, i.e., the inline number; Y is the Y-axis coordinate value of the point, i.e., the crossline number; N is the number of Gaussian functions involved in the deformation, and a, bk, ck, dk, and σk are the corresponding Gaussian function parameters. After calculating the Z-direction displacement and adding it to the original Z coordinate, we obtain the 3D data volume deformed by the Gaussian function.

[0053] In order to enable those skilled in the art to more clearly understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0054] Example 1

[0055] In this embodiment, there are no seismic traces without manual interpretation, so mask processing is not required. Figure 2 This is a schematic diagram of the seismic data volume used in this embodiment. Figure 3 This is a schematic diagram of the manual interpretation results used in this embodiment. First, the seismic data volume is cut to fit the manual interpretation results. Figure 4 This is a schematic diagram of the seismic data volume cut in this embodiment. Then, the seismic data volume is assigned a value. When assigning values, each layer is traversed and the amplitude value of the seismic data volume is replaced with a number to distinguish the layer to which the position belongs, and the label data volume corresponding to the seismic data volume is obtained, such as Figure 5 As shown. The method for label replacement is:

[0056]

[0057] The labels thus produced have a total of i+1 categories, namely 1 majority class and i minority classes. Then the same local transformation is added to the earthquake data volume and the corresponding label data volume to achieve data enhancement. Figure 6 The comparison of the three-dimensional data volume after deformation and the three-dimensional data volume before deformation in this embodiment. The corresponding two-dimensional labels can be obtained by slicing the obtained three-dimensional data volume, such as Figure 7 shown.

[0058] Example 2

[0059] The manual interpretation results used in this example are very sparse, and there are a large number of seismic traces with no interpretation results, so a masking operation is required. When assigning values, each layer is traversed and the amplitude value of the 3D data volume is replaced with a number to distinguish the layer to which the position belongs, thereby obtaining a label data volume corresponding to the 3D data volume. The label replacement method is as follows:

[0060]

[0061] The seismic data used in this embodiment is as follows Figure 8 As shown, the label data body obtained in this embodiment is as follows Figure 9 As shown, the two-dimensional labels extracted from the label data body of this embodiment are as follows Figure 10 shown.

[0062] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for producing seismic data labels for deep learning, characterized in that: The following steps are involved: Step 1. Load the seismic data and its existing interpretation results, and perform segmentation processing on the seismic data; Step 2. Determine the location in the seismic data based on the interpretation results and assign values and mark them; Step 3. Slice the seismic data and perform local transformation on the sliced labeled data volume to expand the training samples; In step 1, the loaded seismic data includes post-stack amplitude data and various seismic attributes derived therefrom, which can be either single-channel data or multi-channel composite data; In step 1, the existing interpretation results include horizon interpretation information, fault interpretation information, and salt dome interpretation information; In step 1, the smallest data volume that encloses all interpretation results is obtained by traversing and cutting it from the original data volume, thereby removing as much redundant background information as possible and reducing the imbalance between the target class and the background class; In step 2, pixel-level labeling is performed on the seismic data, and different values are assigned to the seismic data to represent different background classes. Multiple background classes must be distinguished from each other and strictly distinguished from the target class. In step 2, the unexplained areas are masked and assigned outliers, which are then ignored in subsequent training. In step 3, the label data volume is sliced according to the input requirements of the model, where the 3D model corresponds to cuboid slices of the same size, and the 2D model corresponds to rectangular slices of the same size; The local transformation includes rotation, inversion, replication and local deformation.

Images