A method and system for interpreting a data body

Abstract

The present invention relates to the technical field of data volume interpretation, and discloses a data volume interpretation method and system. A semi-supervised self-training method is introduced in the geological anomaly recognition task, and training is carried out by using a small amount of labeled data and a large amount of unlabeled data. During the self-training process of the model, the pseudo-labels are continuously corrected through three-dimensional smoothing operations, and supervised learning is carried out using the pseudo-labels. The experimental data uses real three-dimensional seismic data, and experts annotate the geological anomalies. Different numbers of labeled samples are divided in the experiment, and the average F1 and average IoU values on the test data exceed 96% and 93% respectively, proving the effectiveness of the present invention.

Description

Technical Field

[0001] The present invention relates to the technical field of data volume interpretation, and particularly relates to a data volume interpretation method and system. Background Art

[0002] Seismic interpretation is a key link in geological exploration and development. Its function is to translate seismic information into geological language or symbols, and its interpretation quality determines the exploration and development direction and process of a block. Seismic interpretation obtains underground information from seismic data and reveals geological significance. Seismic data contains various information in a multi-dimensional environment, including geometric, kinematic, dynamic, and statistical characteristics. Traditional seismic interpretation tasks, such as seismic facies interpretation, detection of abnormal geological bodies, and reservoir prediction, are quite dependent on manual work and time-consuming processes.

[0003] Geological bodies that cause geophysical anomalies and have a certain spatial form are usually called geological anomaly bodies, including different density bodies that cause gravity anomalies, magnetic bodies that cause magnetic anomalies, etc. Their formation and distribution are often related to geological conditions. Faults, river channels, and salt domes are common geological anomaly bodies. The interpretation of geological anomaly bodies is beneficial to analyzing the underground geological environment and further predicting oil and gas reservoirs. However, traditional seismic data interpretation depends to a large extent on the domain knowledge and experience of interpreters and visual cues of geological structures, such as textures and geometries. With the sharp increase in the scale of seismic data for geological exploration, the interpretation of geological structures has become more time-consuming and laborious. In recent years, researchers have used advanced image processing and machine learning technologies to assist interpretation to reduce manual and time consumption and improve interpretation accuracy.

[0004] The continuous development of digital technology has led to an explosive growth of seismic data. Seismic data has realized the conversion from two-dimensional structure to three-dimensional structure. The acquisition of modern wide-azimuth and high-density seismic data provides higher-quality and higher-resolution seismic data for seismic data interpreters, but the change in seismic data has in turn further increased the time and effort of manual interpretation. In addition, the dependence of traditional seismic interpretation methods on manual participation results in the inevitable influence of their interpretation results by the experience and professional knowledge of interpreters, and the application to the refined description of geological anomaly bodies lacks objectivity and reliability. Although traditional machine learning methods reduce the partial dependence on manual work, the degree of automation for the identification of geological anomaly bodies is not high, and the reliability of the identification results is poor.

[0005] In recent years, in order to further reduce manual intervention and improve the accuracy of seismic interpretation, many researchers have tried to use deep learning methods to achieve automated seismic interpretation. Deep learning methods have shown good performance in fields such as Computer Vision (CV), natural language processing, and recommendation systems, and have achieved certain results in the field of geophysical science, providing new ideas for the identification of geological anomalies.

[0006] Given the similarity between the two-dimensional data profile interpretation of three-dimensional seismic data and image interpretation in the CV field, many researchers have tried to use CV algorithms to analyze geological structures. For example, unsupervised clustering and deep convolutional autoencoders in deep learning are used to achieve seismic facies analysis. Edge detection techniques are also often used for the identification of geological anomalies. For example, Carter et al. used edge detection and consistency measurement to detect geological anomalies of seismic faults; Sun Xiping et al. used edge detection methods to image and identify channels and faults. Currently, in the field of seismic interpretation, the application of deep learning has initially achieved results.

[0007] In seismic interpretation tasks, seismic data and their geological information labels are usually used as inputs to establish the relationship between seismic data and geological information. Among them, convolutional neural networks are widely used in classification tasks in seismic interpretation, such as fault identification, salt dome prediction, seismic facies identification, and geological anomaly identification. However, in seismic data, the data point set of geological anomalies accounts for a low proportion of the total data volume, that is, the data category ratio is seriously imbalanced, which greatly limits the performance of the classification network. Using a semantic segmentation network can achieve pixel-level classification of geological anomalies. Classic semantic segmentation networks include Fully Convolutional Networks (FCN) and U-Net networks. Zhang Yilun et al. proposed a multi-channel joint seismic facies identification and first arrival picking method based on U-Net to process seismic facies. Peng Jianwei et al. transformed the problem of underground salt dome identification into an image semantic segmentation problem and used a U-Net network with a residual structure to achieve end-to-end pixel-level salt dome segmentation. However, these methods do not solve the problem that the model performance is too dependent on the number of labels and their accuracy. In addition, in the task of geological anomaly detection, there is also a lot of research on artificially synthesized sample label data. For example, sample enhancement is performed on seismic images through brightness change, rotation, cropping, etc.; deep network models such as generative adversarial networks, variational autoencoders, and transfer learning networks are used for sample augmentation. Wu et al. trained a CNN model using three-dimensional seismic data artificially synthesized according to geometry and used the trained model to predict faults in real seismic data. Although the method of artificially synthesized sample label data can achieve sample data augmentation to a certain extent, there are obvious disadvantages. For example, conventional image sample enhancement is prone to overfitting; the samples generated by deep learning networks have the same feature distribution as the original data and cannot pay attention to more rich feature information. The cost of manual annotation of seismic data is high and the annotations of similar profiles are very close, which means that there is a lot of redundancy in manual annotation. Therefore, semi-supervised learning methods can be used to reduce manual annotation and train with a very small number of labeled data and a large number of unlabeled data in order to achieve higher accuracy.

[0008] The detection of geological anomalies belongs to the category of seismic interpretation. It describes the interfaces and positions of seismic facies. Given the wide application of semantic segmentation methods in seismic interpretation tasks, the present invention adopts a semantic segmentation method to predict geological anomalies in two-dimensional data profiles and obtains relatively accurate results. In practical applications, if the new three-dimensional seismic data (test set) and the labeled and unlabeled three-dimensional seismic data (training set) incorporated into the training set have the same geological conditions, that is, mathematically having the property of "independent and identically distributed", then the model can achieve a prediction effect approximating the training accuracy on the new three-dimensional seismic data. However, this condition is difficult to meet in reality, and it is also difficult to predict whether this condition is met in the case where the test set lacks annotations. Therefore, there are unpredictable risks in using neural networks for "extrapolation". Even using unsupervised domain adaptation learning to adjust the model on new three-dimensional seismic data also has the problem of negative transfer. Therefore, based on the characteristics of scarce geological anomaly label data and similar structures in the same area, the present invention proposes a method for identifying geological anomalies based on deep semi-supervised self-training. It uses a U-Net semantic segmentation network as the network model, combines the characteristics of geological body structure continuity and image processing knowledge, and adds a three-dimensional smoothing operation. The three-dimensional smoothing operation can continuously correct the pseudo-labels during the self-training process to make them more conform to the true shape of geological anomalies, thereby making the final network prediction result more accurate. Summary of the Invention

[0009] To solve the above technical problems, the present invention provides a method and system for data volume interpretation, which greatly reduces manual intervention and can achieve high accuracy.

[0010] To solve the above technical problems, the present invention adopts the following technical solutions:

[0011] A method for data volume interpretation predicts a specific structure in a three-dimensional data volume through a trained network model to obtain a prediction result; the three-dimensional data volume is composed of multiple two-dimensional data, and the training method of the network model is a semi-supervised self-training method, which specifically includes the following steps:

[0012] Step 1: Label part of the two-dimensional data in the three-dimensional data volume to form a labeled data set, and the remaining two-dimensional data forms an unlabeled data set;

[0013] Step 2: Initially train the network model through the labeled data set, and optimize the network model by backpropagation through calculating the loss function until the loss function converges;

[0014] Step 3: Input the unlabeled dataset into the network model, perform 3D smoothing operation on the output prediction results to correct the prediction results of the network model, and calculate the pseudo-labels of the unlabeled data; the 3D smoothing operation includes: performing a morphology-based smoothing operation within the prediction results of the unlabeled data, and performing a mean filter-based smoothing operation between the prediction results of adjacent unlabeled data;

[0015] Step 4: Assign the corresponding pseudo-labels to the unlabeled data in the unlabeled dataset to form a pseudo-labeled dataset, and use the pseudo-labeled dataset to perform supervised learning on the network model;

[0016] Step 5: Repeat Step 2 to Step 4 until the network model converges.

[0017] Furthermore, the 3D data volume is 3D seismic data; the specific structure is a geological anomaly in the 3D seismic data; the 2D data is a 2D data profile in the 3D seismic data.

[0018] Specifically, in Step 3, when performing the morphology-based smoothing operation within the prediction results of the unlabeled data, perform closing operation and opening operation on the prediction results of the unlabeled data in sequence.

[0019] Specifically, in Step 3, when performing the mean filter-based smoothing operation between the prediction results of adjacent unlabeled data, smooth the prediction result a of the current unlabeled data A to: the prediction result a and the mean of the prediction results of the two unlabeled data adjacent to the unlabeled data A:

[0020]

[0021] where f k represents the prediction result of the k-th unlabeled data after the morphology-based smoothing operation, represents f k after the mean filter-based smoothing operation, and the range of k is from 1 to n, where n represents the number of unlabeled data.

[0022] Specifically, the network model is a U-Net network.

[0023] A data volume interpretation system predicts a specific structure in a 3D data volume through a trained network model to obtain prediction results; trains the network model through a semi-supervised self-training system until the network model converges, and the semi-supervised self-training system includes:

[0024] A labeling module labels part of the 2D data in the 3D data volume to form a labeled dataset, and the remaining 2D data forms an unlabeled dataset;

[0025] The preliminary training module initially trains a network model using a labeled dataset and optimizes the network model through backpropagation by calculating a loss function until the loss function converges.

[0026] The 3D smoothing operation module inputs an unlabeled dataset into the network model, performs a 3D smoothing operation on the output prediction results to correct the prediction results of the network model, and calculates the pseudo-labels of the unlabeled data. The 3D smoothing operation includes: adopting a morphology-based smoothing operation within the prediction results of the unlabeled data, and adopting a mean filter-based smoothing operation between the prediction results of adjacent unlabeled data.

[0027] The supervised learning module assigns corresponding pseudo-labels to the unlabeled data in the unlabeled dataset to form a pseudo-labeled dataset, and uses the pseudo-labeled dataset to perform supervised learning on the network model.

[0028] The data volume interpretation system of the present invention corresponds to the data volume interpretation method, and the preferred technical solutions of the data volume interpretation method are equally applicable to the data volume interpretation system.

[0029] Compared with the prior art, the beneficial technical effects of the present invention are:

[0030] The present invention adopts a typical semantic segmentation network, designs a 3D smoothing operation for physical constraints such as the closure, continuity, and interpretability of geological anomalies. A semi-supervised self-training method is introduced in the geological anomaly recognition task. By using a small amount of labeled data and a large amount of unlabeled data for training, the pseudo-labels are continuously corrected through the 3D smoothing operation during the self-training process of the network model, and supervised learning is performed using the pseudo-labels. The experimental data uses real 3D seismic data, and experts annotate the geological anomalies. Different numbers of annotated samples are divided in the experiment, and the average F1 and average IoU values on the test data exceed 96% and 93% respectively, proving the effectiveness of the 3D geological anomaly recognition method in the present invention. Description of the Drawings

[0031] Figure 1 It is the structural diagram of the U-Net network adopted by the present invention. Detailed Embodiment

[0032] The following describes in detail a preferred embodiment of the present invention with reference to the drawings.

[0033] This embodiment provides a 3D geological anomaly recognition method based on a small amount of profile annotations. The network model is trained through a semi-supervised self-training method, and the geological anomalies in the 3D seismic data are recognized through the trained network model.

[0034] 1. Semi-supervised self-training method combined with 3D smoothing operation

[0035] The semi-supervised self-training method proposed by the present invention mainly relates to semi-supervised learning, the U-Net network, and three-dimensional smoothing operations.

[0036] 1.1 Semi-supervised learning

[0037] In the fields of seismic interpretation, lithofacies identification, well logging interpretation, etc., due to the high cost of data annotation, the labeled data is scarce. However, the performance of supervised learning depends on a large amount of labeled data. Therefore, the feasibility of supervised learning in practical tasks is relatively low. The core idea of semi-supervised learning (SSL) is to use a small number of labeled samples and a large amount of unlabeled data for model training, which can effectively overcome the overfitting problem caused by scarce labels and thus achieve higher accuracy. Using semi-supervised learning can reduce the dependence on a large amount of manual annotation, so it is attracting more and more attention from researchers in label-scarce tasks such as geological anomaly identification.

[0038] Self-Training (ST) is one of the commonly used semi-supervised learning methods. Its main idea is to use the unlabeled dataset to expand the labeled dataset. The unlabeled dataset is called the unlabeled dataset, and the labeled dataset is called the labeled dataset. The process is as follows: First, use the labeled data to train a preliminary model, then use this preliminary model to predict the unlabeled data. After obtaining the pseudo-labels, expand the training dataset and retrain. Iterate this process until a certain preset condition is reached. Since the prediction results of the model obtained by the preliminary training for the unlabeled data are not accurate, for classical self-training, a confidence score threshold is generally set to filter the unlabeled data with relatively low prediction results, so as to label pseudo-labels for the unlabeled data with higher confidence.

[0039] 1.2 U-Net network

[0040] Traditional CNN networks have achieved remarkable results in image classification tasks, but there are certain defects in the dense prediction semantic segmentation tasks. The proposed encoder-decoder architecture effectively solves the pixel-level prediction problem in semantic segmentation. Among them, the encoder backbone network is usually composed of a CNN network. The convolutional layer in the CNN network is used to extract features from the input data, and at the same time, the pooling layer is used to downsample the features. The decoder uses upsampling operations such as bilinear interpolation and transposed convolution to reconstruct the feature information and predict the category corresponding to each pixel point of the input data. The U-Net network modifies the encoding and decoding structure based on the fully convolutional neural network (FCN), and realizes dense prediction by connecting the feature maps in the encoder and the upsampling layer in the decoder, so that it can achieve more accurate prediction with very little training data. Therefore, the network model of the present invention selects the U-Net network.

[0041] As Figure 1 shown, the U-Net network in the present invention consists of an encoding path (left side) and a decoding path (right side), with a total of five layers of structure, including 23 convolutions. The encoding path follows the typical architecture of a convolutional network. Each layer of the encoding path includes two 3×3 convolutions (Conv), one 2×2 max pooling operation (Max pooling), and each convolution is followed by a rectified linear unit (ReLU). Each layer of the encoding path performs a downsampling operation. In each downsampling operation, the number of feature channels is doubled, while the size of the feature map is reduced to half of the original. Each layer of the decoding path includes a 2x2 transposed convolution (Up-conv) and two 3×3 convolutions (Conv), and each convolution is followed by a rectified linear unit (ReLU). Each layer of the decoding path performs an upsampling operation. In each upsampling operation, the number of feature channels is halved, and the result after upsampling is concatenated with the feature map of the corresponding level in the encoding path. It should be noted that the sizes of the corresponding feature maps in the encoding path may not be the same, and corresponding padding or cropping is required before concatenation. The concatenated features are subjected to two 3×3 convolution operations, and each convolution is followed by an activation unit (ReLU). In the last layer of structure, 1×1 convolution is used to fuse the channels to obtain the category corresponding to each pixel point. The encoding process goes from shallow to deep, and the semantic information contained in the feature map is richer, but the detailed information is less. To utilize more information, in the decoding process from deep to shallow, the detailed information is retained by concatenating the feature maps of the corresponding encoding layers. In the task of geological anomaly recognition, the high-level semantic information obtained by the deep network contains the overall shape information of the geological anomaly, and the low-level semantic information obtained by the shallow network contains the detailed information of the local area of the geological anomaly. The U-Net network can well fuse the overall shape information and local detailed information, realize the retrieval of feature edges, and thus more accurately predict the boundary of the geological anomaly.

[0042] 1.3 Three-dimensional smoothing operation

[0043] Geological anomaly bodies are usually special geological structures such as protrusions, depressions, and fractures. These structures have certain similar tomographic structures, and the geological layer structures generally have continuity. According to the morphological information of the geological anomaly bodies and the similarity of adjacent profiles within the geological anomaly bodies, the present invention designs a three-dimensional smoothing operation to utilize these characteristics to constrain the recognition results of the geological anomaly bodies, so that the recognition results reach a higher accuracy and are also more in line with the true shape of the geological anomaly bodies. Specifically, it includes two parts: First, introducing the knowledge of digital image processing, for two-dimensional data profiles, two morphological smoothing operations (opening operation and closing operation) are used to achieve smoothing within the profile; then, between two-dimensional data profiles, a mean smoothing operation is added to make each two-dimensional data profile be constrained by the prediction results of its adjacent two-dimensional data profiles to achieve smoothing between profiles, thus jointly constituting the three-dimensional smoothing operation.

[0044] 1.3.1 Smoothing within the profile

[0045] Mathematical morphology is a new method in the fields of image processing and pattern recognition. Its basic idea is: using a structural element with a certain morphology to measure and extract the corresponding shape in the image to achieve the purpose of image analysis and recognition. The opening and closing operations of the image are important operations in mathematical morphology, which are formed based on the combination of erosion and dilation operations and can be used for binary images or grayscale images. Erosion and dilation are image processing methods developed based on the set theory method of mathematical morphology and have been widely used in the fields of digital image processing and machine vision. Erosion is a process of eliminating boundary points and shrinking the boundary inward, which can be used to eliminate small and meaningless objects. Dilation is a process of merging all background points in contact with the object into the object and expanding the boundary outward, which can be used to fill holes in the object. Although erosion processing can separate adhered objects and dilation processing can connect disconnected objects, there are certain problems, that is, after erosion processing, the area of the object is smaller than the original area, and after dilation processing, the area of the object is larger than the original area. The opening operation and the closing operation are precisely proposed to solve this problem.

[0046] The process of first dilating and then eroding is called the closing operation, which is used to fill small holes in the object, connect adjacent objects, and smooth its boundary without significantly changing its area. The main function of the closing operation is similar to that of dilation. Compared with the dilation operation, it has the advantage of basically keeping the original size of the object unchanged. The process of first eroding and then dilating is called the opening operation, which is used to eliminate small objects, separate objects at thin points, and smooth the boundary of larger objects without significantly changing its area. The main function of the opening operation is similar to that of erosion. Compared with the erosion operation, it has the advantage of basically keeping the original size of the object unchanged. The present invention adopts the operation of first performing the closing operation and then the opening operation, which is closer to the true label.

[0047] The advantages of opening and closing operations mainly include effectively filtering out noise, retaining the original information in the image, and the algorithms being easily and effectively implemented by parallel processing methods. The edge information extraction based on mathematical morphology is superior to the edge extraction algorithm based on differential operations. The extracted edges are relatively smooth, and the extracted image contours are also relatively continuous. In the task of detecting geological anomalies, adding opening and closing operations can better correct the contours of geological anomalies and eliminate the outliers predicted by the network.

[0048] 1.3.2 Inter-profile smoothing

[0049] Between two-dimensional data profiles, the surrounding geological anomalies are inferred from the prediction results of adjacent two-dimensional data profiles, and the prediction result of the current two-dimensional data profile is smoothed to the mean of this prediction result and the prediction results of the adjacent two two-dimensional data profiles:

[0050]

[0051] where f k represents the prediction result of the k-th two-dimensional data profile after in-profile smoothing, represents the prediction result of f k after inter-profile smoothing, and the range of k is from 1 to n, where n represents the number of two-dimensional data profiles.

[0052] Since there is only one adjacent two-dimensional data profile for the first and last two-dimensional data profiles, only the average value is calculated with the adjacent two-dimensional data profile. By using three-dimensional smoothing operations, the prediction results output in each round will be smoothed to varying degrees, making the prediction results more and more in line with geological understanding. The smoothed prediction results are used as pseudo-labels to update the next round of the learning process until the network converges and the training is completed.

[0053] The present invention adopts a semi-supervised self-training method. In the first stage, a labeled dataset is first used to preliminarily train the network model, and the network is optimized by calculating the loss function and backpropagation until the loss function converges. In the second stage, all unlabeled data is input into the network model, and the output prediction results are passed to the three-dimensional smoothing operation module. Morphology-based smoothing operations are used within the two-dimensional data profiles, and mean-filtering-based smoothing operations are used between the two-dimensional data profiles to correct the prediction results, calculate the pseudo-labels of the unlabeled data, and form a pseudo-labeled dataset after assigning pseudo-labels to the unlabeled data in the unlabeled dataset. In the third stage, the pseudo-labeled dataset is used to perform supervised learning on the network model, and a smaller weight is assigned to the pseudo-labeled data during the calculation of the loss.

[0054] For a three-dimensional seismic data, in the first and second steps, a labeled dataset and an unlabeled dataset are obtained through expert annotation processing. In the third step, the labeled data is input into the network model to preliminarily train the network model. In the fourth and fifth steps, the unlabeled data is input into the network model to obtain preliminary prediction results, and then through two smoothing processes in the sixth and seventh steps, pseudo-labels are obtained. In the eighth step, the unlabeled data is labeled with pseudo-labels to obtain pseudo-labeled data. In the ninth step, supervised learning is performed using the pseudo-labeled dataset, and steps 3-9 are repeated, and iterative training is performed multiple times until the network model converges.

[0055] 2. Experimental verification

[0056] 2.1 Experimental preparation

[0057] 2.1.1 Data description

[0058] The three-dimensional seismic data used in the present invention is a three-dimensional geological body of 200×500×500, and 400 longitudinal profiles (Inline profiles) are annotated by experts. Among them, the geological anomaly bodies have longitudinal profiles, crossline profiles, and time slice profiles.

[0059] 2.1.2 Evaluation metrics

[0060] The present invention uses the evaluation metrics commonly used in the field of semantic segmentation to calculate the pixel accuracy (PA), class pixel accuracy (CPA), mean pixel accuracy (MPA), intersection over union (IoU), and F1 value of each image.

[0061] The pixel accuracy PA is the ratio of the correctly predicted pixels to the total pixels. In the present invention, the pixel accuracy PA is the sum of the number of correctly detected geological anomaly body pixels and the number of correctly detected non-geological anomaly body pixels divided by the number of all geological anomaly body and non-geological anomaly body pixels. The calculation formula of the pixel accuracy PA is as follows:

[0062]

[0063] Among them, TP (True Positive) represents the true positive, that is, the number of positive samples correctly identified as positive samples; TN (True Negative) represents the true negative, that is, the number of negative samples correctly identified as negative samples; FP (False Positive) represents the false positive, that is, the number of negative samples misidentified as positive samples; FN (False Negative) represents the false negative, that is, the number of positive samples misidentified as negative samples. Here, the samples refer to the pixel points in the image, the positive samples refer to the pixel points of geological anomalies, and the negative samples refer to the pixel points of non-geological anomalies. This will not be elaborated further below.

[0064] The class pixel accuracy CPA represents the probability of correct prediction for a certain class. In the present invention, the class pixel accuracy CPA is the proportion of true geological anomaly pixels among the geological anomaly pixels output by the network model. The calculation formula for the class pixel accuracy CPA is as follows:

[0065]

[0066] The mean per-class pixel accuracy MPA is a simple improvement of PA. It calculates the proportion of correctly classified pixels within each class and then takes the average of all classes. The calculation formula for the mean per-class pixel accuracy MPA is as follows:

[0067]

[0068] In the present invention, there are only two types of pixels, namely geological anomaly pixels and non-geological anomaly pixels, that is, P i represents the probability CPA1 of correct prediction for geological anomaly pixels and the probability CPA2 of correct prediction for non-geological anomaly pixels. sum(·) represents the summation function, MPA represents the average of the two, and C represents the total number of classes.

[0069] The intersection over union Iou is the ratio of the intersection of the predicted value and the true value to the union of the predicted value and the true value, that is, the ratio of the intersection and union of the predicted geological anomaly pixel region and the true geological anomaly pixel region. The calculation method of the intersection over union IoU is as follows:

[0070]

[0071] The F1 value is an index that comprehensively considers precision and recall. The calculation formula for the F1 value is as follows:

[0072]

[0073] Among the above indicators, IoU and F1 are more comprehensive. Therefore, they are two key indicators for judging the quality of the network model.

[0074] 2.1.3 Experimental Equipment

[0075] The network model of the present invention is implemented based on the Pytorch deep learning framework on the Windows 10 system, and a GPU NVIDIA GeForce RTX 3090 is used to accelerate the training. The entire network model training is completed in less than 10 hours, and the prediction time for a single seismic profile is less than 2 seconds.

[0076] 2.2 Experimental process

[0077] 2.2.1 Comprehensive performance display

[0078] Starting from the 3D seismic data, 200 longitudinal profile images of geological anomaly data are extracted and labeled, and the size of a single longitudinal profile image is 500×500. According to the number of labeled longitudinal profile images in the training data, the experiment is divided into supervised learning, simple self-training, and semi-supervised self-training in the present invention with different numbers of labels. Supervised learning only needs to use the labeled longitudinal profile images to train the network model, and then directly predict the remaining unlabeled longitudinal profile images; simple self-training is based on supervised learning. First, use the labeled longitudinal profile images to train the network model and obtain the prediction results of the unlabeled longitudinal profile images as pseudo-labels, and then add the pseudo-label images to retrain the model, and multiple iterations are used to achieve self-training; the semi-supervised self-training method in the present invention processes the pseudo-labels and then adds them to the training set after each network model training is completed, which is different from simple self-training.

[0079] The experimental results with different numbers of labels are shown in Table 1. Generally speaking, using the same labels for training, the experimental results of the semi-supervised self-training method in the present invention are significantly improved compared with supervised learning and simple self-training. The specific analysis is as follows:

[0080] The improvement in the number of labels can improve the performance of all methods as a whole. Therefore, the number of labels is the key factor determining the model performance.

[0081] From the comprehensive indicators IoU and F1, the effect of supervised learning is poor because pure supervised learning does not use unlabeled images to improve its generalization performance; in the case of insufficient labels, it is easy to overfit to the labeled images.

[0082] Simple self-training is a semi-supervised learning method. However, as can be seen from Table 1, except for the case where the number of labels is 5, the performance of simple self-training has decreased compared to supervised learning. This is because the proportion of pseudo-labels is relatively high and the pseudo-labels themselves are not accurate enough, resulting in the introduction of incorrect label information in the next round of iterative learning and triggering the avalanche effect. Therefore, simple self-training may lead to semi-supervised risk, that is, the semi-supervised performance is lower than that of supervised learning. In fact, the best results in simple self-training are shown in Table 1. It can be observed that after multiple rounds of iterative self-training, the performance of simple self-training will decrease significantly, far lower than that of supervised learning.

[0083] Compared with supervised learning, the semi-supervised self-training method in the present invention generally has an improvement of about 10% to 30% in terms of IoU and F1 metrics, verifying the effectiveness of the self-training method in the present invention. Among them, under the tasks with the number of labels being 5, 10, 15, and 20 respectively, the improvements in F1 are 23.9%, 16.2%, 12.3%, and 9.3% respectively, and the improvements in IoU are 32.7%, 25.1%, 20.4%, and 15.9% respectively. Obviously, even if semi-supervised learning is effective, its improvement effect compared to supervised learning decreases as the number of labels increases, which is due to a certain amount of information redundancy brought about by the increase in the number of labels.

[0084] With the increase in the number of labels, the performance of the semi-supervised self-training method in the present invention has a certain degree of improvement. Among them, when the number of labels increases from 5 to 10, from 10 to 15, and from 15 to 20, the improvements in F1 are 4.3%, 1%, and 0.6% respectively, and the improvements in IoU are 7.6%, 2%, and 0.9% respectively. It can be seen that as the number of labels increases, the improvement effect of the self-training method in the present invention also decreases. In the tasks of the present invention, increasing the number of labels from 5 to 10 can bring a relatively large improvement. Therefore, the experimental results with the number of labels being 10 will be further analyzed subsequently.

[0085] Table 1 Experimental results with different numbers of labels (the metrics are the averages on the test set)

[0086]

[0087]

[0088] 2.2.2 Longitudinal section detail display

[0089] In supervised learning, the prediction of 10 labeled longitudinal section images is very accurate, indicating that the training effect of supervised learning is good. However, the prediction performance metrics of the unlabeled longitudinal section images between two adjacent labeled longitudinal section images show a "U-shaped" distribution. This indicates that the unlabeled images closer to the labeled images have greater similarity and thus have better prediction effects; otherwise, the effects are worse.

[0090] The semi-supervised self-training method in the present invention can significantly suppress the above-mentioned "U-shaped" distribution problem of the indicators, and has good prediction effects on all unlabeled longitudinal profile images.

[0091] Furthermore, the 1st, 100th, 181st, and 201st images of the longitudinal profile are selected for visualization. Among them, the 1st and 201st longitudinal profile images are located at both ends of the longitudinal profile of the 3D data volume, the 100th longitudinal profile image is in the middle, and the 181st longitudinal profile image is the longitudinal profile image with the worst test effect in supervised learning. The following conclusions are obtained:

[0092] The effects of supervised learning and simple self-training are poor, and the smoothness of the prediction results is poor; the semi-supervised self-training method in the present invention has a good effect, and the prediction results have good smoothness; especially for the 100th and 181st longitudinal profile images, their prediction results are almost the same as the true values.

[0093] The prediction effect of the semi-supervised self-training method in the present invention on the longitudinal profile images at both ends of the longitudinal profile is relatively poor compared to the longitudinal profile images in the middle. This is because the middle profile can receive the smoothness constraints from both the front and back directions, while the profiles at both ends can only be constrained unidirectionally.

[0094] 2.2.3 Detail display of time slice profile

[0095] For the time slice profile images before the 300th and after the 500th, it is difficult for the performance indicators of supervised learning to reach 80%. Combining Table 2, it can be seen that its evaluation indicators PA, CPA, and MPA are generally good, but F1 and IoU are poor. This indicates that supervised learning is relatively accurate in predicting pixel points within the geological anomaly profile, but the overall effect is not good, especially the prediction performance of the images at both ends of the time slice profile gradually decreases.

[0096] The semi-supervised self-training method in the present invention can significantly suppress the problems existing in supervised learning, has good prediction effects on all profile images, and the improvement is obvious.

[0097] The prediction effect of the profile images near both ends of the time slice profile is relatively poor compared to the profile images in the middle. This is because the profiles in the middle of the geological anomaly are mostly connected and distributed in blocks, while the geological anomalies in the profiles at both ends are smaller and not connected.

[0098] Furthermore, the 200th, 250th, 500th, 700th, and 800th images of the time slice profile are selected for visualization in the present invention. Among them, the 200th, 250th, and 800th time slice profile images are located in the two end regions of the time slice profile of the 3D seismic data, and the prediction result indicators are poor. The 500th and 700th are located in the middle region, and the results are good:

[0099] The effect of supervised learning and simple self-training is poor, and the smoothness of the prediction results is poor; the self-training method in the present invention has a better effect, and the prediction results have better smoothness; especially for the 500th and 700th images, the prediction results of supervised learning and simple self-training are poor, while the prediction results of the method of the present invention are almost the same as the true values.

[0100] The geological anomaly body in the 200th time slice profile image is extremely small. The semi-supervised self-training result in the present invention is almost the same as the true value, while the results of supervised learning and simple self-training are poor.

[0101] Table 2 Prediction results of images from the 185th to 850th of the time slice profile (the index is the average value on the test set)

[0102] Method PA(%) CPA(%) MPA(%) F1(％) IoU(%) Supervised learning 95.8 78.0 81.4 68.8 57.7 Simple self-training 91.3 81.1 66.6 45.9 32.3 Self-training of the present invention <![CDATA 99.1 > <![CDATA 88.9 > <![CDATA 94.6 > <![CDATA 89.0 > <![CDATA 84.1 >

[0103] 2.2.4 Horizontal profile detail display

[0104] The conclusions obtained by the analysis are almost the same as those in Section 2.2.3, so they will not be elaborated here.

[0105] Since there is no small target problem in the time slice profile in the horizontal profile, its prediction effect is better. In addition, the performance index shows a "hump shape". The prediction performance index of the profile near the outside of the 3D seismic data is poor, but in fact its prediction results may not be bad. The prediction performance indexes of the 100th and 400th horizontal profile images can only reach about 80% and 60% respectively. However, the visualization results show that the effect is not bad. This is because the target is small and the number of pixel points is small. Therefore, even a small number of mispredictions may lead to a large decrease in the performance index.

[0106] Table 3 Prediction results of profile images from the 83rd to 460th of the horizontal profile (the index is the average value on the test set)

[0107] Method PA(%) CPA(%) MPA(%) F1(％) IoU(%) Supervised learning 95.1 74.6 81.1 67.2 56.5 Simple self-training 89.5 81.0 68.0 49.1 34.8 Self-training of the present invention <![CDATA 99.0 > <![CDATA 84.2 > <![CDATA 91.2 > <![CDATA 83.0 > <![CDATA 78.0 >

[0108] The present invention proposes a method for identifying three-dimensional geological anomalies based on a small number of cross-section annotations. Based on a semantic segmentation model and geological knowledge, a semi-supervised self-training network with three-dimensional smoothing operations is constructed, which can achieve high-precision prediction of geological anomalies in the entire area only by using a small amount of annotated data. The present invention further verifies the effectiveness of the proposed three-dimensional smoothing operation through experiments. The results show that constraining the prediction results using the physical structure characteristics of geological bodies can effectively improve the accuracy of geological anomaly identification. In addition, experiments are carried out with different numbers of labels, and general supervised learning and simple self-training methods are compared in the research. Different experimental results show that the present invention can significantly improve the accuracy and efficiency of seismic facies interpretation. The network model trained by the semi-supervised self-training method has great potential in seismic interpretation and geological anomaly detection.

[0109] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

[0110] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for interpreting a data body, characterized in that, Predict a specific structure in the three-dimensional data volume through the trained network model to obtain a prediction result; the three-dimensional data volume is composed of multiple two-dimensional data, and the training method of the network model is a semi-supervised self-training method, which specifically includes the following steps: Step 1: Label some of the two-dimensional data in the three-dimensional data volume to form a labeled data set, and the remaining two-dimensional data form an unlabeled data set; Step 2: Initially train the network model through the labeled data set, and optimize the network model by backpropagation through calculating the loss function until the loss function converges; Step 3: Input the unlabeled data set into the network model, perform a three-dimensional smoothing operation on the output prediction result, correct the prediction result of the network model, and calculate the pseudo-labels of the unlabeled data; the three-dimensional smoothing operation includes: adopting a morphology-based smoothing operation within the prediction result of the unlabeled data, and adopting a mean filter-based smoothing operation between the prediction results of adjacent unlabeled data; Step 4: Assign the corresponding pseudo-labels to the unlabeled data in the unlabeled data set to form a pseudo-labeled data set, and use the pseudo-labeled data set to perform supervised learning on the network model; Step 5: Repeat Step 2 to Step 4 until the network model converges; The three-dimensional data volume is three-dimensional seismic data; the specific structure is a geological anomaly in the three-dimensional seismic data; the two-dimensional data is a two-dimensional data profile in the three-dimensional seismic data.

2. The data body interpretation method according to claim 1, wherein: In Step 3, when adopting a morphology-based smoothing operation within the prediction result of the unlabeled data, perform a closing operation and an opening operation on the prediction result of the unlabeled data in sequence.

3. The data body interpretation method according to claim 1, characterized in that: In Step 3, when adopting a mean filter-based smoothing operation between the prediction results of adjacent unlabeled data, smooth the prediction result a of the current unlabeled data A to: the prediction result a and the mean of the prediction results of the two unlabeled data adjacent to the unlabeled data A; Among them, f k represents the prediction result of the k-th unlabeled data after the morphological-based smoothing operation, represents f k after the smoothing operation based on mean filtering. The range of k is from 1 to n, where n represents the number of unlabeled data.

4. The data body interpretation method according to claim 1, characterized in that: The network model is a U-Net network.

5. A data body interpretation system, characterized in that Predict a specific structure in the three-dimensional data volume through the trained network model to obtain a prediction result; Train the network model through a semi-supervised self-training system until the network model converges. The semi-supervised self-training system includes: A labeling module that labels some of the two-dimensional data in the three-dimensional data volume to form a labeled data set, and the remaining two-dimensional data form an unlabeled data set; An initial training module that initially trains the network model through the labeled data set, and optimizes the network model by backpropagation through calculating the loss function until the loss function converges; A three-dimensional smoothing operation module that inputs the unlabeled data set into the network model, performs a three-dimensional smoothing operation on the output prediction result, corrects the prediction result of the network model, and calculates the pseudo-labels of the unlabeled data; the three-dimensional smoothing operation includes: adopting a morphology-based smoothing operation within the prediction result of the unlabeled data, and adopting a mean filter-based smoothing operation between the prediction results of adjacent unlabeled data; A supervised learning module that assigns the corresponding pseudo-labels to the unlabeled data in the unlabeled data set to form a pseudo-labeled data set, and uses the pseudo-labeled data set to perform supervised learning on the network model; The three-dimensional data volume is three-dimensional seismic data; the specific structure is a geological anomaly in the three-dimensional seismic data; the two-dimensional data is a two-dimensional data profile in the three-dimensional seismic data.

Images