A deep learning seismic phase identification method based on knowledge constraints

Abstract

The present invention discloses a deep learning seismic phase identification method based on knowledge constraints, which is applied to the intersection of artificial intelligence and earth science. It addresses the problem that existing technologies cannot achieve seismic phase segmentation under conditions of small samples and weak seismic reflection characteristics. The present invention learns the information of seismic phase entities represented in a seismic phase knowledge graph G through a seismic phase knowledge graph entity relationship learning model, and finally obtains a knowledge graph entity representation vector V of each seismic phase entity. Then, label interpolation is used to obtain the spatial constraints of the middle part of the seismic data, and the obtained seismic phase representation vector is filled to obtain the spatial knowledge constraints of the seismic data. Finally, the seismic data and its spatial knowledge constraints are jointly input into a deep learning network for training, and the converged model is used to perform seismic phase segmentation prediction.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary field of artificial intelligence and earth science, and particularly relates to a seismic phase identification technology. Background Art

[0002] Seismic facies are the reflection of sedimentary facies on seismic sections. Seismic facies identification aims to characterize the seismic behavior of reservoirs by classifying seismic attributes into categories that may be related to geological heterogeneity. Its accurate identification provides an effective basis for the exploration and development of underground resources, especially oil and gas resources, and is of great significance to sedimentary facies research and oil and gas exploration and development. Traditional seismic facies identification, commonly known as the facies face method, is performed manually by professionals based on sedimentary facies characteristics on the section and their own experience. However, manual seismic facies identification is labor-intensive, often time-consuming, and somewhat subjective. With the continuous development of exploration and development, the demand for seismic facies interpretation has increased significantly. Researchers have begun to use computers to identify seismic facies by combining seismic attributes. Subsequently, with the rapid growth of seismic data and the increasing demand for interpretation accuracy, traditional seismic facies identification has become inadequate, leading researchers to focus on artificial intelligence. In recent years, deep learning (DL) models have been widely used for seismic facies identification. Deep learning is a branch of machine learning, which falls within the scope of artificial intelligence and focuses on improving the performance of specific computer algorithms through empirical learning. Since its introduction in 2006, it has been applied to various fields, including computer vision and natural language processing, and has led to the development of many excellent deep learning models. With the continuous development of artificial intelligence, the domestic geological industry has gradually focused on the transformation towards digitalization, informatization, and intelligence. In recent years, a series of studies on seismic facies identification based on DL models have been carried out both domestically and internationally. Chevitarese et al. proposed a technique using convolutional neural networks (CNNs) to interpret seismic data. Results showed that this method can produce segmentation results close to those obtained by manual interpretation, but faster than manual interpretation. Zhang et al. established U-net and SegNet networks and used ensemble learning methods to combine the models and optimize the classification results. Results showed that deep learning methods can accurately classify seismic facies. Islam et al. proposed a new deep learning method based on U-net and Se-ResNet to identify salt bodies in seismic images. The results show that it can automatically extract the required useful features from seismic data.

[0003] DL models can achieve good results when the sample size is sufficient. However, due to factors such as exploration depth and cost, drilling data is often limited, especially in new areas, which can lead to insufficient samples. Therefore, seismic phase identification is often a small-sample problem. Furthermore, deep learning is data-driven, and its "black box" nature makes it easy to use but also leads to lower reliability. DL models are effective when seismic reflection features are clear in seismic images. However, in complex areas or where seismic reflection features are not clear, DL models can present a series of problems, including confusing seismic phase results, inconsistent geological patterns, confusing interpretations, and poor generalization.

[0004] Therefore, how to achieve seismic phase segmentation under the conditions of small samples and weak seismic reflection characteristics is a problem that needs to be solved at present. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a deep learning seismic phase identification method based on knowledge constraints.

[0006] The technical solution adopted by the present invention is: a deep learning seismic phase identification method based on knowledge constraints, comprising the following steps:

[0007] Step 1: Select m images from the original seismic images, have geological experts interpret them to obtain the labels corresponding to each original seismic image, and number them according to Inline;

[0008] Step 2: Construct a seismic phase knowledge graph. Specifically, the labeled seismic images from step 1 are represented using triples of the "entity-relationship-entity" structure. This is mainly divided into two parts: entity extraction from the seismic phase knowledge graph and relationship extraction from the seismic phase knowledge graph.

[0009] Step 3: Use a graph embedding model to learn the attributes and relationships of related seismic phases to obtain seismic phase representation vectors to reflect the similarity between seismic phases;

[0010] Step 4: interpolate the label data obtained in step 1 to obtain the seismic phase distribution of the seismic data in the middle part of every two label data, and then fill the seismic phase representation vectors of each type of seismic phase to form a spatial knowledge constraint feature map;

[0011] Step 5: The original seismic image and the spatial knowledge constraint feature map are combined as input data for the deep semantic segmentation network. The deep semantic segmentation network is trained using a backpropagation algorithm based on the loss function.

[0012] Step 6: Perform segmentation prediction. Input all original seismic images and the spatial knowledge constraint feature maps corresponding to each original seismic image into the trained deep semantic segmentation network to predict the segmentation result map of the original seismic image.

[0013] The beneficial effects of the present invention are as follows: the present invention learns the information of the seismic phase entities represented in the seismic phase knowledge graph G through the seismic phase knowledge graph entity relationship learning model, and finally obtains the knowledge graph entity representation vector V of each seismic phase entity, and then uses label interpolation to obtain the spatial constraints of the middle part of the seismic data, and fills the obtained knowledge graph entity representation vector V to obtain the spatial knowledge constraints of the seismic data. Finally, the original seismic data and its spatial knowledge constraints are input into the deep learning network for training, and the converged deep learning network is used to perform seismic phase segmentation prediction; compared with the existing technology, the present invention can more accurately realize the fine identification of seismic phases under the conditions of small samples and weak seismic reflection characteristics. After applying the present invention, the problems of inaccurate recognition results and chaotic stratigraphic sequences that occur when using the traditional pure data-driven deep learning neural network model are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the method of the present invention;

[0015] Figure 2 is the earthquake phase label;

[0016] Figure 3 It is the earthquake phase knowledge graph;

[0017] Figure 4 Schematic diagram of knowledge extraction-seismic phase representation vector learning;

[0018] Figure 5 Form a schematic diagram for spatial knowledge constraints;

[0019] Figure 6 This is the seismic phase identification result of Inline 332; DETAILED DESCRIPTION

[0020] To facilitate those skilled in the art to understand the technical content of the present invention, the present invention is further explained below with reference to the accompanying drawings.

[0021] like Figure 1 Shown is a flowchart of a deep learning seismic phase identification method based on knowledge constraints provided in an embodiment of the present application.

[0022] The specific steps include:

[0023] 1. This embodiment mainly uses the 3D seismic data F3 collected in the North Sea. This area has obvious large-scale S-shaped beds of large deltaic sediments, obvious folds and fault structures, and the seismic phase labels are as follows: Figure 2As shown in the figure, since the present invention studies small sample data, it extracts 7 images from the original seismic images (cross-sections), and interprets them by geological experts to obtain the corresponding stratigraphic line maps for each original seismic image. The stratigraphic line maps are then numbered according to the inline. Inline represents the registration line.

[0024] 2. Constructing a seismic phase knowledge graph. The knowledge graph constructed in this embodiment focuses on the relationships between seismic phases, representing them using triples of the "entity-relationship-entity" structure. The seismic phase knowledge graph construction module is divided into two main parts: entity extraction from the seismic phase knowledge graph and relationship extraction from the seismic phase knowledge graph.

[0025] (1) Entity extraction from earthquake phase knowledge graph

[0026] The dataset in this example refers to the seismic facies of the Netherlands North Sea data provided by Yazeed Alaudah et al. (Alaudah Y, Michalowicz P, Alfarraj M, et al. AMachine Learning Benchmark for Facies Classification: Interpretation, 10.1190 / int-2018-0249.1[P]. 2019.). It is divided into 6 categories according to stratigraphy and period lithology, namely: The Upper North Sea Group, The Middle North Sea Group, The Lower North Sea Group, The Chalk and Rijnland Group, The Scruff Group, and The Zechstein Group. At the same time, considering that tectonic levels and components will generate new entities, the following tectonic level entities are added: Cenozoic, Mesozoic, and Permian; and component nodes: Claystones, Sandstones, Clayformations, Evaporites, and Carbonates.

[0027] (2) Extraction of seismic phase knowledge graph relationships

[0028] Relationship extraction requires extracting the semantic relationship between two or more entities from the text. Semantic relationships are usually used to connect two entities and, together with the entities, express the main meaning of the text. The subject-predicate-object (SPO) ternary structure is used to represent the relationship extraction results. By collecting relevant data, the present invention organizes the attributes and relationships of six types of seismic phases and constructs the relationship between entities from multiple perspectives, such as geological objects, components, structural levels, sedimentary relationships, and adjacency relationships.

[0029] After completing the entity extraction and relationship extraction of the seismic phase knowledge graph, the triples of the "entity-relationship-entity" structure are used to represent them, thus completing the construction of the seismic phase knowledge graph and obtaining the seismic phase knowledge graph triple set (S).

[0030] The final seismic phase knowledge graph constructed is as follows Figure 3 shown.

[0031] 3. If Figure 4 As shown, a graph embedding model is used to obtain a seismic phase representation vector by learning the attributes and relationships of related seismic phases, thereby reflecting the similarity between seismic phases. This embodiment adopts the TranE (Translating Embeddings) graph embedding method.

[0032] First, the constructed seismic phase knowledge graph triple set S is defined as the positive sample set of the present invention, as shown in formula (1):

[0033] S={(h,r,t)}⑴

[0034] Among them, h, t∈E, E is the set of seismic phase knowledge graph entities, r∈R, R is the set of seismic phase knowledge graph relations, (h, r, t)∈S, S is the set of knowledge graph triples.

[0035] On this basis, the present invention selects to modify the entity h or entity t in the triple (h, r, t) in the set S to obtain the negative sample (h', r, t) or (h, r, t'), thereby forming the triple negative sample set S' (h',r,t') , as shown in formula (2):

[0036] S' (h',r,t') ={(h',r,t),h'∈E}∪{(h,r,t'),t'∈E}⑵

[0037] Then define the specific model loss function as shown in formula (3)

[0038]

[0039] Where d(h+r,t) represents the distance between h+r and t, [γ+d(h+r,t)-d(h'+r,t')] + It refers to the positive part of γ+d(h+r,t)-d(h'+r,t'), and γ>0 is a marginal hyperparameter.

[0040] Finally, the present invention continuously updates the seismic phase characterization vector through the random gradient descent of the loss function, and finally minimizes the model loss function to obtain the seismic phase characterization vector V of each type of seismic phase.

[0041] 4. If Figure 5 As shown in the figure, the seven labeled data are interpolated to obtain the seismic phase distribution of the middle part of the earthquake data, and then the seismic phase characterization vectors of each type of seismic phase are filled to form a spatial knowledge constraint feature map. It is mainly divided into the following two steps:

[0042] (1) Use the interpolation method to process these 7 labeled data and generate the seismic phase distribution of the middle part of the earthquake data to estimate the seismic phase distribution of the earthquake data missing between the labeled data. For example, if the label data inline numbers are 0, 20, and 40, the seismic phase distribution of the earthquake data with inline numbers 1-19 and 21-39 can be obtained through interpolation.

[0043] (2) The seismic phase characterization vectors of various seismic phases are filled into the middle part of the seismic phase distribution map generated by interpolation, forming a spatial knowledge constraint feature map.

[0044] 5. Train the semantic segmentation network.

[0045] (1) The basic segmentation networks selected in this embodiment are U-net and SegNet networks.

[0046] (2) The original seismic image and the spatial knowledge constrained feature map are combined as the input data of the deep semantic segmentation network. The spatial knowledge constrained feature map can provide some prior knowledge to help the model better understand the distribution of seismic phases in the seismic image.

[0047] (3) The loss function of this embodiment adopts the cross entropy loss function. Based on the loss function, the deep semantic segmentation network is trained through the back propagation algorithm. In addition, the optimizer adopts the Adam optimizer.

[0048] 6. Perform earthquake phase segmentation prediction

[0049] (1) All original seismic images and their corresponding spatial knowledge constraint feature maps are input into the converged semantic segmentation network model to predict the seismic phase segmentation result map of the original seismic images.

[0050] (2) The prediction results of Inline 331 profile are as follows Figure 6 As shown. Figure 6 As shown in the first layer of images, the recognition results using the DL model have rough edges, confusing interpretations, and missing seismic phase interpretations. Since the strata have a certain degree of continuity, the distribution of the strata in the current section can be roughly known through interpolation. Therefore, the addition of this prior knowledge through spatial constraints actually means that the training results are offset towards the spatial constraints during the training process. In addition, although there is a certain degree of continuity between strata, there are also many mutations, such as pinch-out. Therefore, the entity representation vector is filled in the spatial constraint to complete the knowledge graph constraint and finally form the spatial knowledge constraint. The entity representation vector is learned from the knowledge graph, which can reflect the relationship between seismic phases and seismic phases, thereby reflecting the similarities and differences between seismic phases. Therefore, by learning the spatial knowledge constraint and the original seismic image together, it can perform well, such as Figure 6 As shown in the second layer image, after adopting the present invention, the problems of rough edges, confusing interpretation, and missing seismic phase interpretation in the DL model have been improved.

[0051] Those skilled in the art will appreciate that the embodiments described herein are intended to aid the reader in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific descriptions and embodiments. Various modifications and variations are readily apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A deep learning seismic phase identification method based on knowledge constraints, characterized by: The following steps are involved: Step 1: Geological experts interpret each original seismic image to obtain a label and number it according to Inline. Step 2: Construct a seismic phase knowledge graph. Specifically, the labeled seismic images from step 1 are represented using triples of the "entity-relationship-entity" structure. This is mainly divided into two parts: entity extraction from the seismic phase knowledge graph and relationship extraction from the seismic phase knowledge graph. Step 3: Use a graph embedding model to learn the attributes and relationships of related seismic phases to obtain the seismic phase representation vector; Step 4: interpolate the label data obtained in step 1 to obtain the seismic phase distribution of the seismic data in the middle part of every two label data, and then fill the seismic phase representation vectors of each type of seismic phase to form a spatial knowledge constraint feature map; Step 5: The original seismic image and the spatial knowledge constraint feature map are combined as input data for the deep semantic segmentation network. The deep semantic segmentation network is trained using a backpropagation algorithm based on the loss function. Step 6: Perform segmentation prediction. Input all original seismic images and the spatial knowledge constraint feature maps corresponding to each original seismic image into the trained deep semantic segmentation network to predict the segmentation result map of the original seismic image.

2. A deep learning seismic phase identification method based on knowledge constraints according to claim 1, characterized in that: In step 2, the seismic phase knowledge graph entity extraction is to classify the seismic phases according to the stratum and period lithology to obtain seismic phase entities of different categories.

3. A deep learning seismic phase identification method based on knowledge constraints according to claim 2, characterized in that: Step three is as follows: A graph embedding model is used to obtain seismic phase representation vectors by learning the attributes and relationships of seismic phases; Let the set obtained from the triples in step 2 be S, and define S as the positive sample set, expressed as: S={(h,r,t)} Among them, h and t are entities in the triple (h, r, t), h, t∈E, E is the set of entities in the earthquake phase knowledge graph, r∈R, R is the set of relations in the earthquake phase knowledge graph, (h, r, t)∈S, S is the set of knowledge graph triples; Select entity h or entity t in the triple (h, r, t) in the modified set S to obtain negative samples (h', r, t) or (h, r, t'), thus forming a triplet negative sample set S' (h',r,t') , the expression is: S' (h',r,t') ={(h',r,t),h'∈E}∪{(h,r,t'),t'∈E} Then define the specific model loss function as: Where d(h+r,t) represents the distance between h+r and t, [γ+d(h+r,t)-d(h'+r,t')] + It refers to the positive part of γ+d(h+r,t)-d(h'+r,t'), where γ is a marginal hyperparameter; Finally, the seismic phase characterization vector is updated by stochastic gradient descent of the loss function, and the seismic phase characterization vector V of each type of seismic phase is obtained by minimizing the model loss function.

4. A deep learning seismic phase identification method based on knowledge constraints according to claim 3, characterized in that: The graph embedding model adopts the TranE graph embedding method.

5. A deep learning seismic phase identification method based on knowledge constraints according to claim 4, characterized in that: Step 4 is as follows: For the label data obtained in step 1, if the inline numbers of two adjacent seismic images are discontinuous, the seismic phase distribution of the seismic data between the two inline numbers is generated by interpolation; thereby obtaining the interpolated seismic phase distribution map; The seismic phase characterization vectors of various seismic phases are filled into the seismic phase distribution map generated by interpolation, forming a spatial knowledge constraint feature map.

6. A deep learning seismic phase identification method based on knowledge constraints according to claim 5, characterized in that: The deep semantic segmentation networks used in step 5 are U-net and SegNet networks.

7. The deep learning seismic phase identification method based on knowledge constraints according to claim 6 is characterized in that: The loss function used in step 5 is the cross entropy loss function.

Images