Seismic facies prediction method based on deep learning semantic segmentation technology

Abstract

The application provides a seismic facies prediction method based on a deep learning semantic segmentation technology, which comprises the following steps: step 1, constructing a seismic facies recognition network model based on deep learning semantic segmentation; step 2, carrying out seismic facies recognition data preprocessing and data enhancement; step 3, carrying out network training to obtain a seismic facies recognition training model; and step 4, using the seismic facies recognition training model to carry out seismic facies prediction. The seismic facies prediction method based on the deep learning semantic segmentation technology improves the U-Net network for the seismic facies prediction problem, designs a double-target function based on the mean square error of the regression problem and the cross-entropy of the classification problem, and ensures that the model generalization result is closer to the actual situation; the method can well solve the geological problem of seismic facies recognition, and compared with artificial recognition, the recognition efficiency is greatly improved, and the recognition result is consistent with the geological understanding.

Description

Technical Field

[0001] This invention relates to the field of petroleum geophysical exploration technology, and in particular to a seismic facies prediction method based on deep learning semantic segmentation technology. Background Technology

[0002] Reflection seismic exploration is an important method for oil and gas resource exploration. It involves artificially generating seismic waves at the surface, which propagate underground. Upon encountering interfaces between rock strata with different properties, the seismic waves are reflected and refracted, and then received by geophones at the surface. Seismic data is then processed to form seismic profiles. Therefore, seismic profiles contain rich information about the subsurface medium, most notably structural information such as faults and anticlines. However, structural information alone is far from sufficient for oil and gas exploration. It is also necessary to extract the sedimentary facies of the subsurface strata from the seismic profile, which is crucial for further determining the location and extent of oil and gas reservoirs. However, due to the complexity of the subsurface medium, the seismic phase axes on the seismic profile are intricately interwoven, and there is no effective traditional technique to directly identify seismic phases from the profile. Relying on manual methods, it is even more difficult to identify seismic phases based on changes in amplitude, phase, and frequency; moreover, manually interpreting seismic phases is virtually impossible.

[0003] Seismic facies identification is a crucial research area in exploration seismic stratigraphy. The purpose of studying seismic facies is to analyze sedimentary facies, assign clear geological meaning to seismic profiles, and help determine the sedimentary facies and distribution range of reservoirs, further identifying favorable targets and hydrocarbon enrichment areas. However, traditional seismic facies identification methods can only predict the planar distribution of seismic facies, making it difficult to identify the spatial distribution of seismic facies on the profile, and it is also difficult to incorporate the prior geological knowledge of interpretation experts. Existing methods relying on manual interpretation of seismic facies on profiles have low interpretation efficiency, depend heavily on manual interpretation experience, and are difficult to apply on a large scale.

[0004] Deep learning, a core technology in artificial intelligence and machine learning, has evolved alongside neural networks over more than half a century and possesses a complete theoretical foundation. From the concept of a single neuron in the 1940s, to the perceptron in the 1970s, to the backpropagation (BP) network in the 1980s, and the rapid development of the internet in the 1990s, the theoretical and data foundations for deep learning have been invaluable. In 2006, Hinton proposed deep belief networks, a deep network model. Unlike traditional shallow models, deep learning emphasizes the depth of the model structure and the importance of feature learning. Through layer-by-layer feature transformation, it transforms the feature representation of the sample metadata space to a new feature space, making classification or prediction easier. Therefore, using deep learning technology to solve classification and prediction problems no longer requires manual intervention in characterizing the features of the target. This makes its application to complex problems such as seismic facies identification possible.

[0005] With the development of deep learning technology, computer vision has made leaps and bounds. Image segmentation is a crucial research and application area in computer vision, involving dividing pixels in an image into different parts and assigning them different labels according to certain rules. Image segmentation is further divided into image classification, semantic segmentation, instance segmentation, and panoptic segmentation, depending on the task. Semantic segmentation evolved from fully convolutional networks (FCNs) and has formed many frameworks, such as SegNet, DeepLab, and U-net. FCNs replace fully connected layers with convolutions, allowing the network input to images of any scale. The general process involves the image being encoded and decoded to restore it to its original size. Therefore, semantic segmentation based on FCNs is also called full-pixel semantic segmentation, which involves classifying each pixel in an image. This invention borrows techniques from semantic segmentation to solve image segmentation problems and applies them to the identification and processing of seismic phases.

[0006] Chinese patent application CN202011463003.0 discloses an intelligent seismic strata identification method, comprising: constructing an improved U-Net model; obtaining seismic slices from a 3D seismic dataset, converting them into grayscale images and adjusting them to an m×m size; training the model; and enabling the trained U-Net model to identify strata in the work area and obtain pixel-level target strata segmentation results. The improved U-Net model is characterized by using a 1×1 kernel, a stride of 1, and 8 feature layers at the network output, outputting the training results through Keras softmax, and mapping the 8 output categories to seismic profiles of the same size with probability values ​​between 0 and 1. The seismic dataset preprocessing involves converting the seismic slice dataset to grayscale, padding the grayscale images around the profiles with zero values ​​to obtain the slice dataset, cropping the seismic slice dataset, and using an overlapping sliding window decomposition to obtain image patches, resulting in the preprocessed seismic dataset. The training model uses a multi-class cross-entropy loss function. However, this application does not address the problem of seismic facies identification by improving the convolutional and pooling layers in the downsampling and upsampling stages based on the traditional U-net network structure, nor does it employ a dual objective function to train the deep learning model. Therefore, it cannot solve the problems of overly smooth results caused by using the mean squared error objective function of a single regression problem, especially where the predicted results at the edges of some seismic facies data volumes differ significantly from the actual situation. It also cannot solve the problems of overly fragmented prediction results caused by using traditional semantic segmentation as the objective function of a single classification problem, which may even result in voids within some seismic facies data volumes.

[0007] Chinese patent application CN202310521011.3 discloses a method for intelligent seismic facies identification and detection based on the TransUNet deep learning network, characterized by the following steps: A. Planarizing a 3D sample set using a sliding window strategy; B. Constructing an intelligent seismic facies determination model based on TransUNet; C. Training the intelligent seismic facies determination model based on TransUNet. The sample set consists of original 2D seismic data images obtained by slicing the original 3D seismic data, and an encoder structure using a convolutional neural network and Transformer to encode the original 2D seismic data images, resulting in a seismic data encoding feature map. Training the intelligent seismic facies determination model requires dividing the original seismic data into a 3:7 ratio for model training. However, this application does not use the U-Net network, which is improved for seismic facies prediction, for seismic facies training and prediction to ensure that the model's generalization results are closer to reality; nor does it address the issue of using a preprocessed and data-augmented small 3D data volume as the training dataset. This fails to adequately address the geological problems of seismic facies identification from actual 3D data and prevents the identification results from accurately matching geological understanding.

[0008] In a non-patented paper titled "Intelligent Seismic Physics Recognition Based on Improved Deep Learning Methods," a deep learning method for seismic phasing classification and recognition is proposed. This method adds a pyramid pooling module (PPM module) to the end of the U-Net model to improve its ability to acquire global information. It employs an objective function that integrates cross-entropy and the Dice index to improve the characterization of minority seismic phasing boundaries in imbalanced data. The concept of "prediction information entropy" is proposed to evaluate the uncertainty of seismic phasing prediction results. However, the improvements to the U-Net network in this paper do not involve increasing the number of network layers or network outputs, nor do they incorporate techniques based on a dual objective function that addresses both regression and classification problems. Therefore, it is difficult to guarantee that the model's generalization results are closer to reality.

[0009] In a non-patent literature paper titled "Research on Key Technologies for Left Ventricular Myocardial Segmentation in Cardiac MRI Images Based on Fully Convolutional Networks," batch normalization technology was employed, an optimizer with adaptively decreasing learning rate and a dual-channel three-label loss function were proposed to train the network, and different upsampling methods were used according to the characteristics of the endocardium and epicardium to achieve left ventricular segmentation in cardiac MR images. Specifically, different attention mechanisms were used for the endocardium and epicardium for dual-channel network training. For the endocardium network, a self-attention mechanism was introduced at the bottleneck layer of the U-Net network, and a cross-attention mechanism was introduced at the skip connections; for the epicardium network, a gated attention mechanism was introduced at the skip connections. To expand the receptive field of the U-Net network to the left ventricle, five parallel dilated convolutions with different dilation rates were subsequently introduced into both the endocardium and epicardium networks. Although the aforementioned literature also improves upon the traditional U-Net network to enhance image segmentation accuracy, the actual problems it addresses are different from those in this application, and the improvements to the U-Net network are also different.

[0010] The existing technologies described above are significantly different from this invention. A search reveals no literature in the XY category, indicating that this invention possesses novelty. Since existing technologies lack solutions to the technical problem we aim to address, we have invented a novel seismic facies prediction method based on deep learning semantic segmentation technology. Summary of the Invention

[0011] The purpose of this invention is to provide a seismic phase prediction method based on deep learning semantic segmentation technology that creatively improves the U-Net network based on practical technical problems, thereby increasing the recognition speed and accuracy of the recognition model.

[0012] The objective of this invention can be achieved through the following technical measures: a seismic phase prediction method based on deep learning semantic segmentation technology, which includes: step 1, constructing a seismic phase recognition network model based on deep learning semantic segmentation; step 2, performing seismic phase recognition data preprocessing and data augmentation; step 3, performing network training to obtain a seismic phase recognition training model; and step 4, using the seismic phase recognition training model to perform seismic phase prediction.

[0013] The objective of this invention can also be achieved through the following technical measures:

[0014] In step 1, the seismic phase recognition network model is a deep learning semantic segmentation network structure model, which includes two parts: encoding and decoding, and has the following characteristics: it adopts fully convolutional layers; it uses pooling layers for downsampling and deconvolution operations for upsampling; it adopts skip connection layers; it uses BN layers and residual blocks; the network adopts single input and dual output, where the output corresponding to the one-hot encoded label requires softmax operation.

[0015] In step 1, the deep learning semantic segmentation network structure model is based on the traditional U-Net network, and a targeted improvement design is made for the seismic phase problem. It adopts a completely symmetrical network structure, that is, the downsampling encoding part on the left half of the network structure and the upsampling decoding part on the right half of the network structure are symmetrical. It uses skip chain layers to directly pass the features extracted in the downsampling stage to the upsampling stage and merge them with the data in the upsampling stage into feature data with a larger number of channels. However, compared with the traditional U-Net network, it adds two convolutional layers and one pooling layer in the downsampling stage, symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage, increases the number of skip chains, and adds Dropout design to the max pooling in the downsampling stage.

[0016] In step 1, the application of BN layers in the seismic phase recognition network model refers to adding BN layers to all convolutional and deconvolutional layers to improve training stability, training speed, and generalization ability. The application of residual blocks in the seismic phase recognition network model refers to using a residual block structure based on the idea of ​​local skip links to replace general convolutional layers for computation in order to eliminate the risk of gradient vanishing that may occur in backpropagation speed calculation.

[0017] In step 1, the dual output of the seismic facies identification network model refers to the fact that, unlike the traditional U-Net network, the seismic facies identification network model adopts a dual-output design, which corresponds to two types of objective functions. Among them, network output 1 is a 3D convolution with 1 channel followed by BN and Sigmoid activation functions, and network output 2 is a 3D convolution with the number of channels equal to the number of seismic facies categories, followed by BN, ReLU activation functions, and Softmax operation.

[0018] In step 1, the objective function of the seismic facies identification network model adopts a dual objective function of mean square error (MSE) and cross-entropy (CE), and the two objective functions are weighted using the following formula to obtain the final objective function:

[0019] loss = w1 × loss mse +w2×loss ce

[0020] In the formula, loss mse The objective function is the mean squared error, with a weight of w1, corresponding to a network output of 1, and the corresponding label not being one-hot encoded; loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

[0021] In step 1, the objective function based on cross-entropy refers to the multi-class weighted focal objective function derived from the cross-entropy problem of binary classification, and its formula is:

[0022] loss ce = -w×(1-p pred ×p true ) γ ×log(p pred )

[0023] In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The outputs for each model category are represented by γ, which is the adjustment factor.

[0024] In step 2, seismic facies identification data preprocessing refers to edge padding and resampling of the data. Edge padding means padding the actual number of sample points in a single channel of seismic data with zero values ​​according to the number of sample points that are powers of 2, and adding a new category of zero to the corresponding label category. Resampling means coarsening and resampling the expanded data according to actual needs.

[0025] In step 2, seismic facies identification data enhancement refers to extracting small three-dimensional data volumes as samples from the seismic data of the training area in the line and xline directions according to certain rules, and rotating them horizontally at multiple angles to obtain more training samples.

[0026] In step 3, the training data obtained after preprocessing and data augmentation in step 2 is fed into the seismic phase recognition network model provided in step 1, and appropriate parameters are selected for training to obtain the seismic phase recognition training model.

[0027] In step 4, the model trained in step 3 is invoked to make multiple seismic phase predictions on the seismic data of the prediction area. The final output category of each numerical point in the data space is the category with the highest probability of occurrence in the multiple prediction results.

[0028] In step 4, multiple prediction of seismic phases refers to the use of a processing method similar to data augmentation in step 2, which extracts a large number of three-dimensional small data volumes from the test data, so that each seismic trace in the training area corresponds to multiple prediction results, and then the category with the highest probability is obtained for each trace and each sampling point as the final prediction result.

[0029] The objective of this invention can also be achieved through the following technical measures: a seismic facies prediction system based on deep learning semantic segmentation technology, which uses a seismic facies prediction method based on deep learning semantic segmentation technology to predict seismic facies in complex lithological zones in the field of petroleum geophysical exploration.

[0030] The seismic facies prediction method based on deep learning semantic segmentation technology in this invention is applicable to the seismic facies division of complex lithological oil and gas blocks. It directly identifies seismic facies on seismic profiles, obtains a sufficient number of 3D data labels through data augmentation, utilizes semantic segmentation technology from the field of image recognition, and trains a modified U-Net neural network. The trained deep learning model is then applied to the data prediction area. Results show that using semantic segmentation technology in deep learning is a feasible and efficient method for seismic facies identification. Compared with existing technologies, the specific advantages and features of this invention are reflected in the following two aspects:

[0031] First, in terms of network structure and deep learning algorithms, semantic segmentation technology from the field of computer vision was introduced and applied. The U-Net network was improved for the problem of earthquake phase prediction, and a dual objective function based on the mean square error of regression and the cross-entropy of classification was designed to ensure that the generalization results of the model are closer to the actual situation.

[0032] Secondly, the successful application of deep learning technology benefits from a large amount of training data, making data preprocessing and augmentation particularly important. Data augmentation requires ensuring a sufficiently large amount of data while avoiding redundancy and distribution issues. Therefore, the seismic facies prediction method provided by this invention can effectively solve the geological problems of seismic facies identification, and compared with manual identification, it significantly improves identification efficiency, with the identification results consistent with geological understanding. Attached Figure Description

[0033] Figure 1 This is a flowchart of a specific embodiment of the seismic facies prediction method based on deep learning semantic segmentation technology of the present invention;

[0034] Figure 2 This is a schematic diagram of the U-Net network model structure based on the seismic phase problem used in a specific embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the planar range of SEG seismic phase zone data used in a specific embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of a seismic profile (solid line portion of survey line A) and its corresponding label from training data in a specific embodiment of the present invention.

[0037] Figure 5 This is a seismic profile at location A of survey line in a specific embodiment of the present invention;

[0038] Figure 6 This is a schematic diagram of the label and prediction results corresponding to the seismic profile at location A of the survey line in a specific embodiment of the present invention;

[0039] Figure 7This is a schematic diagram of the planar range of the SEG seismic phase zone data training area and prediction area used in a specific embodiment of the present invention;

[0040] Figure 8 This is a schematic diagram of the prediction results obtained from profile A of survey line when performing seismic phase prediction on the entire work area in a specific embodiment of the present invention.

[0041] Figure 9 This is a schematic diagram of a seismic profile (Line 1966) and its corresponding label from the training data in the Parihaka data used in a specific embodiment of the present invention.

[0042] Figure 10 This is a schematic diagram of a seismic phase prediction profile (Line2008) in a specific embodiment of the present invention. Detailed Implementation

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, and / or combinations thereof.

[0045] like Figure 1 As shown, Figure 1 This is a flowchart of the seismic facies prediction method based on deep learning semantic segmentation technology according to the present invention. The seismic facies prediction method based on deep learning semantic segmentation technology includes:

[0046] Step S1: Construct a seismic facies recognition network model based on deep learning semantic segmentation;

[0047] Step S2: Conduct seismic facies identification data preprocessing and data augmentation;

[0048] Step S3: Conduct network training to obtain a seismic facies recognition training model;

[0049] Step S4: Use the seismic facies identification training model to carry out seismic facies prediction.

[0050] The objective of this invention can also be achieved through the following technical measures:

[0051] In step S1, the seismic phase recognition network model is based on a deep learning semantic segmentation network structure model, which includes two parts: encoding and decoding, and has the following characteristics: it adopts fully convolutional layers; it uses pooling layers for downsampling and deconvolution operations for upsampling; it adopts skip connection layers; it uses BN layers and residual blocks; the network adopts single input and dual output, where the output corresponding to the one-hot encoded label requires softmax operation.

[0052] Furthermore, the deep learning semantic segmentation network structure model is based on the traditional U-Net network and has been specifically improved for the seismic phase problem. It adopts a completely symmetrical network structure (i.e., the downsampling encoding part on the left half of the network structure and the upsampling decoding part on the right half of the network structure are symmetrical), and uses skip chain layers to directly pass the features extracted in the downsampling stage to the upsampling stage and merge them with the data in the upsampling stage into feature data with a larger number of channels. However, compared with the traditional U-Net network, it adds two convolutional layers and one pooling layer in the downsampling stage, symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage, increases the number of skip chains to 5, and adds Dropout design to the 5 max pooling operations in the downsampling stage.

[0053] Furthermore, the application of BN layer features in the earthquake phase recognition network model refers to adding BN (Batch Normalization) layers to all convolutional and deconvolutional layers to improve training stability, training speed, and generalization ability.

[0054] Furthermore, the application of residual block characteristics in the seismic phase identification network model refers to the use of a residual block structure based on the idea of ​​local skip links instead of general convolutional layers for computation in order to eliminate the risk of gradient vanishing that may occur when backpropagation is used to obtain the velocity.

[0055] Furthermore, the dual-output characteristic of the seismic facies identification network model refers to the fact that, unlike the traditional U-Net network, the seismic facies identification network model adopts a dual-output design, corresponding to two types of objective functions. Among them, network output 1 is a three-dimensional convolution with 1 channel followed by BN and Sigmoid activation functions, and network output 2 is a three-dimensional convolution with the number of channels equal to the number of seismic facies categories, followed by BN, ReLU activation functions, and Softmax operation.

[0056] In step S1, the objective function of the seismic facies identification network model adopts a dual objective function of mean square error (MSE) and cross entropy (CE), and the two objective functions are weighted using the following formula to obtain the final objective function:

[0057] loss = w1 × loss mse +w2×loss ce

[0058] In the formula, loss mse The objective function is the mean squared error, with a weight of w1, corresponding to a network output of 1, and the corresponding label not being one-hot encoded; loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

[0059] Furthermore, the objective function based on cross-entropy refers to the multi-class weighted focal objective function derived from the cross-entropy of binary classification problems, with the following formula:

[0060] loss ce = -w×(1-p pred ×p true ) γ ×log(p pred )

[0061] In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The output is for each type of model, and γ is the adjustment factor, which is usually set to 2.

[0062] In step S2, the seismic phase identification data preprocessing refers to edge padding and resampling of the data. Edge padding means padding the actual number of sample points in a single channel of seismic data with zero values ​​according to the number of sample points that are powers of 2, and adding a category of 0 to the corresponding label category. Resampling means coarsening and resampling the expanded data according to actual needs.

[0063] In step S2, the seismic facies identification data enhancement refers to extracting three-dimensional small data volumes as samples from the seismic data of the training area in the line and xline directions according to certain rules, and rotating them horizontally by 90 degrees, 180 degrees and 270 degrees to obtain more training samples.

[0064] In step S3, the training data obtained after preprocessing and data augmentation in step S2 is fed into the seismic phase recognition network model provided in step S1, and appropriate parameters are selected for training to obtain the seismic phase recognition training model.

[0065] In step S4, the model trained in step S3 is called to make multiple seismic phase predictions on the seismic data of the prediction area. The final output category of each numerical point in the data space is the category with the highest probability of occurrence in the multiple prediction results.

[0066] Furthermore, the multiple prediction of seismic phases refers to the use of a processing method similar to the data augmentation in step S2, which extracts a large number of three-dimensional small data volumes from the test data, so that each seismic trace in the training area corresponds to multiple prediction results, and then the category with the highest probability is obtained for each trace and each sampling point as the final prediction result.

[0067] This invention addresses the problem of seismic facies identification by proposing a deep learning solution utilizing an improved U-net model. In terms of network structure and deep learning algorithms, semantic segmentation techniques from computer vision are introduced and applied. A dual-objective function based on regression and classification problems is designed to ensure that the model's generalization results more closely approximate reality. The successful application of deep learning technology benefits from a large amount of training data, making data preprocessing and augmentation particularly important. Data augmentation must ensure a sufficiently large amount of data while avoiding redundancy and distribution issues. Actual network training and prediction results in the test area demonstrate that the seismic facies identification deep learning solution of this invention is feasible, significantly improving identification efficiency compared to manual identification, and the identification results are consistent with geological understanding.

[0068] The following are several specific embodiments of the application of the present invention.

[0069] Example 1:

[0070] Based on SEG public test data, a sufficient number of 3D data labels were obtained through data preprocessing and data augmentation. Semantic segmentation techniques from the field of image recognition were used, combined with an improved U-Net deep learning neural network model for training. The trained deep learning model was then applied to the test area of ​​the public test data. Results show that using semantic segmentation techniques in deep learning is a feasible and efficient method for seismic facies identification. This seismic facies prediction method based on deep learning semantic segmentation technology includes the following steps:

[0071] Step S1: Construct a seismic phase recognition network model based on deep learning semantic segmentation.

[0072] A seismic facies identification network model is constructed using a deep learning-based semantic segmentation network structure model, comprising encoding and decoding parts. Its main features include:

[0073] (1) Using fully convolutional layers

[0074] Most semantic segmentation models use fully convolutional layers. Using fully convolutional layers not only makes it possible to input any image, but also makes it possible to perform semantic segmentation in a pixel-like sense. Furthermore, the use of fully convolutional layers utilizes weight sharing, which greatly reduces the number of weights in the model and simplifies network training.

[0075] (2) Downsampling uses pooling layers, and upsampling uses deconvolution.

[0076] (3) Use jump link layer

[0077] Pooling operations during encoding can cause a decrease in resolution and loss of information. Therefore, combining the information from downsampling directly during upsampling to form new feature data can help reduce information loss and improve prediction resolution.

[0078] (4) Using BN layers and residual blocks

[0079] Convolutional operations are typically accompanied by Batch Normalization (BN) layers to improve training stability, speed, and generalization ability. If the network has many layers, backpropagation for calculating gradients risks vanishing, so residual blocks are often used instead of regular convolutional layers. In fact, residual blocks are also local skip connections.

[0080] (5) The network output requires softmax operation.

[0081] The network adopts a single-input dual-output approach. The output corresponding to the one-hot encoded label needs to be subjected to a softmax operation in the pixel sense before it is used to calculate the objective function together with the encoded label, so as to ensure that each pixel obtains the maximum probability in its class.

[0082] Furthermore, in step S1, the deep learning semantic segmentation network structure model, based on the traditional U-Net network, has undergone targeted improvements to address the seismic phase problem. Similar to the traditional U-Net network, it adopts a completely symmetrical network structure (i.e., the downsampling encoding part on the left half of the network structure and the upsampling decoding part on the right half of the network structure are symmetrical), and uses skip-chain layers to directly pass the features extracted in the downsampling stage to the upsampling stage and merge them with the upsampling stage data into feature data with a larger channel volume. This ensures that the information lost in the downsampling stage is supplemented in the upsampling process.

[0083] However, the traditional U-Net network structure is relatively simple and insufficient to solve the problem of seismic facies identification. Therefore, this invention has been improved by adopting, for example... Figure 2 Compared to the traditional U-Net, this network structure adds two convolutional layers and one max pooling layer in the downsampling stage, and symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage. Consequently, the number of hop links increases from 4 in the traditional U-Net to 5.

[0084] To enhance the network's generalization ability and training stability, and reduce training difficulty, Batch Normalization (BN) layers were added to all convolutional and deconvolutional layers. Using BN layers allows the network to use a larger learning rate, accelerating training. Unlike the traditional U-Net, Dropout is added during the downsampling phase with five max pooling operations to prevent overfitting and further enhance generalization ability.

[0085] On another front, the U-Net network model based on the seismic facies problem described in this invention differs from the traditional U-Net network. It adopts a single-input, dual-output design, corresponding to two types of objective functions. Output 1 is a three-dimensional convolution (with 1 channel), followed by BN and Sigmoid activation functions. Output 2 is a three-dimensional convolution (with the number of channels equal to the number of seismic facies categories), followed by BN and ReLU activation functions, and then a Softmax operation for each numerical point.

[0086] The choice of objective function is particularly important in deep learning and is closely related to the solution of practical problems. Seismic facies identification can be solved as a regression problem or as a multi-class classification problem. Regression problems often use mean squared error (MSE) as the objective function, while classification problems often use cross-entropy (CE) as the objective function.

[0087] The problem of seismic facies identification, which uses only mean squared error as the objective function, has yielded overly smooth results, especially at the edges of some seismic facies data volumes, where the predicted results deviate significantly from the actual situation. However, if semantic segmentation is used as the objective function, the predicted results are too fragmented, even containing gaps within some seismic facies data volumes. Considering these factors, a dual objective function is adopted to train the deep learning model. That is, the seismic facies identification problem is simultaneously treated as a regression and classification problem. The two objective functions are weighted using the following formula to obtain the final objective function (see Formula 1).

[0088] loss = w1 × loss mse +w2×loss ce (

[0090] Formula 1)

[0091] In the formula, loss mse The objective function is the mean squared error, with a weight of w1, corresponding to a network output of 1, and the corresponding label not being one-hot encoded; loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

[0092] In Equation 1, the objective function loss based on cross-entropy is... ce This is the multi-class weighted focal objective function. The focal objective function is derived from the cross-entropy problem in binary classification. It mainly addresses the class imbalance problem in classification. The formula for the multi-class weighted focal objective function is shown in Equation 2:

[0093] loss ce = -w×(1-p pred ×p true )γ×log(p pred (Equation 2)

[0094] In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The model output for each category is γ, which is an adjustment factor that reduces the contribution of easily classified samples to the objective function and focuses more on difficult-to-classify categories. Generally, a value of 2 is sufficient.

[0095] Step S2: Conduct seismic facies identification data preprocessing and data augmentation.

[0096] Sufficient high-quality labeled data is crucial for the success of deep learning. Based on the SEG seismic facies survey data, data preprocessing and data augmentation operations were performed. First, edge padding and resampling were performed on the data. Since the SEG seismic survey data time axis has 1006 sampling points, edge padding was performed on sampling points 1007 to 1024 to ensure a regular network structure. The padding value was 0, corresponding to a label category of 0, thus increasing the number of seismic facies identification categories from 6 to 7. This invention uses three-dimensional data samples for model training, with 1024 sampling points in the time dimension. Considering the actual computer hardware conditions, the sampling was limited to 256 sampling points.

[0097] Data augmentation needs to consider the distribution of training data, obtain as many seismic phase patterns as possible, and also consider the capacity of computer hardware. Therefore, in this embodiment, a three-dimensional small data volume is extracted from the training area every ten grids in the line and xline directions as a sample. The size of the three-dimensional small data volume is 64*64*256 (sample points), and it is rotated by 90 degrees, 180 degrees and 270 degrees to obtain more training samples.

[0098] Figure 3 This is a planar range map of the SEG seismic phase zone data used in this embodiment, where the lower part is the training area and the upper part is the test area. Figure 4 For a seismic profile of the training data ( Figure 3 The solid line portion of test line A (i.e., the training region) and its corresponding label, where, Figure 4 'a' represents the location of the training area (the survey line is...). Figure 3 Training data profile of the solid line portion of the central measurement line A. Figure 4 b represents a seismic profile (the solid line portion of survey line A) and its corresponding labels from the training data. It is divided into 6 seismic facies, labeled as numbers 1-6, with the following geological meanings: 1: basement, 2: low-velocity mudstone A, 3: massive transport sedimentary layer, 4: high-velocity mudstone B, 5: river valley, 6: submarine canyon.

[0099] Step S3: Conduct network training to obtain a seismic phase recognition training model.

[0100] The training data obtained after preprocessing and data augmentation in step S2 is fed into the seismic facies recognition network model provided in step S1 for training, and a seismic facies recognition training model is obtained. The training parameters are as follows: learning rate is 0.001, optimization method is Adam[4], batch size is 15, number of iterations is 5, and dropout is 0.5.

[0101] Step S4: Use the seismic facies identification training model to conduct seismic facies prediction.

[0102] Call the model trained in step S3, in Figure 3 Seismic facies prediction was performed on the test area, employing multiple predictions. For each numerical point in the data space, the output category was the category with the highest probability among the multiple predictions. Specifically, similar to the data augmentation method used in the training area, a large number of small 3D data volumes were extracted from the test area. Thus, a single seismic trace in the test area corresponded to multiple prediction results. Then, the category with the highest probability was obtained for each trace and each sampling point as the final prediction result. The accuracy of this prediction reached 0.913.

[0103] by Figure 3 The location of survey line A in the image shows the seismic facies prediction results from deep learning. Figure 5 This is the seismic profile of survey line A. Figure 6 for Figure 5 The labels and prediction results corresponding to the seismic profiles. Figure 6 The left side of the dividing line represents known labels, and the right side represents the predicted seismic facies. Comparing the seismic profiles and the predicted results, it is evident that the predicted results accurately depict the seismic facies of the test area, and there is no clear boundary between the label data and the predicted results, indicating that the predicted results are reliable.

[0104] Example 2

[0105] In a specific embodiment 2 of the present invention, based on SEG public test data, an improved U-Net deep learning neural network model is trained, and the trained deep learning model is applied to the entire work area for comparative prediction. This seismic facies prediction method based on deep learning semantic segmentation technology includes the following steps:

[0106] Step S1: Construct a seismic phase recognition network model based on deep learning semantic segmentation.

[0107] A seismic phase recognition network model is constructed using a deep learning-based semantic segmentation network structure model. It consists of two parts: encoding and decoding, and has the following characteristics: it uses fully convolutional layers; it uses pooling layers for downsampling and deconvolution operations for upsampling; it uses skip connection layers; it uses BN layers and residual blocks; the network adopts a single input and dual output, where the output corresponding to the one-hot encoded label requires softmax operation.

[0108] Furthermore, in step S1, the deep learning semantic segmentation network structure model, compared to the traditional U-Net, adds two convolutional layers and one max-pooling layer in the downsampling stage, and symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage. Consequently, the number of skip links increases from four in the traditional U-Net to five. Batch normalization (BN) layers are added to all convolutional and deconvolutional layers, and Dropout is added to the five max-pooling operations in the downsampling stage to prevent overfitting and further enhance generalization ability.

[0109] On another front, the U-Net network model based on the seismic facies problem described in this invention adopts a single-input dual-output design, corresponding to two types of objective functions. Output 1 is a three-dimensional convolution (with 1 channel), followed by BN and Sigmoid activation functions. Output 2 is a three-dimensional convolution (with the number of channels equal to the number of seismic facies categories), followed by BN and ReLU activation functions, and then a Softmax operation for each numerical point.

[0110] Regarding the selection of the objective function, a dual objective function of mean squared error (MSE) and cross-entropy (CE) is adopted, and the two objective functions are weighted using the following formula to obtain the final objective function:

[0111] loss = w1 × loss mse +w2×loss ce (

[0113] Equation 3)

[0114] In the formula, loss mse The objective function is the mean squared error, with a weight of w1, corresponding to a network output of 1, and the corresponding label not being one-hot encoded; loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

[0115] In Equation 1, the objective function loss based on cross-entropy is... ceThis is the multi-class weighted focal objective function. The focal objective function is derived from the cross-entropy problem in binary classification. It mainly addresses the class imbalance problem in classification. The formula for the multi-class weighted focal objective function is shown in Equation 2:

[0116] loss ce = -w×(1-p pred ×p true )γ×log(p pred (Equation 4)

[0117] In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The output is for each type of model, and γ is an adjustment factor, which is set to 2 in this embodiment.

[0118] Step S2: Conduct seismic facies identification data preprocessing and data augmentation.

[0119] In this embodiment, the data is first padded and resampled. The padded value is 0, and the corresponding label category is 0. The total number of seismic facies identification categories is 7, which are labeled with numbers 1-6 except for the padded 0. Their geological meanings are: 1: basement, 2: low-velocity mudstone A, 3: massive transport sedimentary layer, 4: high-velocity mudstone B, 5: river valley, 6: submarine canyon. Three-dimensional data samples are used for model training, with the time dimension coarsened to 256 sample points. In this embodiment, three-dimensional small data volumes are extracted as samples in the training area. The size of the three-dimensional small data volume is 64*64*256 (sample points), and it is rotated by 90 degrees, 180 degrees, and 270 degrees to obtain more training samples.

[0120] Figure 7 This is a schematic diagram of the planar range of the training and prediction areas of the SEG seismic phase zone data used in this embodiment, where the lower part is the training area and the entire area is the prediction area.

[0121] Step S3: Conduct network training to obtain a seismic phase recognition training model.

[0122] The training data obtained after preprocessing and data augmentation in step S2 is fed into the seismic phase recognition network model provided in step S1 for training, resulting in a seismic phase recognition training model.

[0123] Step S4: Use the seismic facies identification training model to conduct seismic facies prediction.

[0124] The model trained in step S3 is called to perform seismic facies prediction in the entire region. Seismic facies prediction is carried out by multiple predictions. The output category of each numerical point in the data space is the category with the highest probability of occurrence in the multiple prediction results.

[0125] The obtained seismic facies recognition training model was used to predict seismic facies for the entire work area. The profile corresponding to the predicted line A is shown in the figure below. Figure 8 As shown, compared with the known labels in the training area, the trained model effectively represents the seismic facies characteristics of the prediction area.

[0126] Example 3

[0127] In a specific embodiment 3 of the present invention, the application was carried out based on Parihaka public test data. The results showed that the seismic phase identification method using semantic segmentation technology in deep learning achieved good results. The steps of the seismic phase identification method based on semantic segmentation technology in deep learning of the present invention are as follows:

[0128] Step S1: Construct a seismic phase recognition network model based on deep learning semantic segmentation, which includes two parts: encoding and decoding, and has the following characteristics: it adopts fully convolutional layers; it uses pooling layers for downsampling and deconvolution operations for upsampling; it adopts skip connection layers; it uses BN layers and residual blocks; the network adopts single input and dual output, where the output corresponding to the one-hot encoded label requires softmax operation.

[0129] This deep learning semantic segmentation network architecture model is based on the traditional U-Net network and features targeted improvements for the seismic phase problem. It adopts a completely symmetrical network structure, meaning that the downsampling encoding part on the left side of the network structure and the upsampling decoding part on the right side of the network structure are symmetrical. It also uses skip-chain layers to directly pass the features extracted in the downsampling stage to the upsampling stage and merge them with the upsampling stage data into feature data with a larger channel volume. Compared with the traditional U-Net network, it adds two convolutional layers and one pooling layer in the downsampling stage, symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage, increases the number of skip links, and adds Dropout design to the max pooling in the downsampling stage.

[0130] To enhance the network's generalization ability and training stability, and reduce training difficulty, batch normalization (BN) layers were added to all convolutional and deconvolutional layers. Using BN layers allows for a larger learning rate during network training, accelerating the training process. Unlike the traditional U-Net, Dropout was added during the downsampling phase with five max pooling operations to prevent overfitting and further enhance generalization ability.

[0131] The U-Net network model improved based on the seismic facies problem described in this invention differs from the traditional U-Net network. It adopts a single-input, dual-output design, corresponding to two types of objective functions. Output 1 is a 3D convolution (1 channel), followed by BN and Sigmoid activation functions. Output 2 is a 3D convolution (the number of channels equals the number of seismic facies categories), followed by BN and ReLU activation functions, and then a Softmax operation for each numerical point.

[0132] The objective function of the seismic facies identification network model adopts a dual objective function of mean square error (MSE) and cross-entropy (CE), and the two objective functions are weighted using the following formula to obtain the final objective function:

[0133] loss = w1 × loss mse +w2×loss ce (Equation 5)

[0134] In the formula, loss mse The objective function is the mean squared error, with a weight of w1, corresponding to a network output of 1, and the corresponding label not being one-hot encoded; loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

[0135] In Equation 5, the objective function loss based on cross-entropy is... ce This is the multi-class weighted focal objective function. The focal objective function is derived from the cross-entropy problem in binary classification. It mainly addresses the class imbalance problem in classification. The formula for the multi-class weighted focal objective function is shown in Equation 6:

[0136] loss ce = -w×(1-p pred ×p true ) γ ×log(p pred (Equation 6)

[0137] In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The output for each model category is γ, which is an adjustment factor. In this embodiment, there are 6 seismic facies, labeled as numbers 1-6, with the following geological meanings: 1: basement, 2: mudstone A, 3: massive transported sedimentary layer, 4: mudstone B, 5: river valley, 6: submarine canyon. The corresponding weights w for each category are 6.6, 2.4, 24.2, 3.5, 100.0, and 11.3, respectively.

[0138] Step S2: Conduct seismic facies identification data preprocessing and data augmentation.

[0139] In this embodiment, the data is first padded and resampled, reducing the number of time axis sampling points from 1501 to 2048 with a padding value of 0, corresponding to a label category of 7, and changing the seismic facies category from 6 to 7. Three-dimensional data samples are used for model training, with the time dimension sampling points thinned to 512 points. Small three-dimensional data points are extracted as training samples, with a sample size of 64*64*512 (sample points), and rotated by 90 degrees, 180 degrees, and 270 degrees to obtain more training samples.

[0140] Figure 9 This example uses a seismic profile (Line 1966) and its corresponding label from the training data in the Parihaka dataset. Figure 9 'a' represents the training data profile of Line 1966 in the training area. Figure 9 b is the label corresponding to this survey line, which is divided into 6 seismic phases except for the edge patching.

[0141] Step S3: Conduct network training to obtain a seismic phase recognition training model.

[0142] The training data obtained after preprocessing and data augmentation in step S2 is fed into the seismic phase recognition network model provided in step S1 for training, resulting in a seismic phase recognition training model.

[0143] Step S4: Use the seismic facies identification training model to conduct seismic facies prediction.

[0144] Call the model trained in step S3, in Figure 3 Seismic facies prediction is performed in the test area. The seismic facies prediction is carried out in multiple predictions. A processing method similar to the data augmentation in step S2 is used to extract a large number of three-dimensional small data volumes from the test data, so that each seismic trace in the training area corresponds to multiple prediction results. Then, the category with the highest probability is obtained for each trace and each sampling point as the final prediction result.

[0145] Figure 10 The seismic phase prediction results for Line2008 survey line in this embodiment are shown. It can be seen that the prediction results have well described the seismic phase prediction problem in this embodiment, and the prediction effect is good.

[0146] This invention addresses the problem of seismic facies identification. Building upon the traditional U-net network structure, it adds two convolutional layers and one max-pooling layer during the downsampling stage, and symmetrically adds two convolutional layers and one upsampling layer during the upsampling stage, increasing the number of skip links from four to five. Batch normalization (BN) layers are added to all convolutional and deconvolutional layers, and Dropout is incorporated into the five max-pooling operations during the downsampling stage. Furthermore, the input is single-input, and the output employs a dual-output design, corresponding to two types of objective functions. The deep learning model is trained using dual objective functions, treating the seismic facies identification problem simultaneously as a regression and classification problem. Weights are applied to the two objective functions. Using dual objective functions to train the deep learning model effectively solves the problems of overly smooth results (especially significant discrepancies between predicted results and actual conditions at the edges of seismic facies data volumes) resulting from a single regression objective function, and the fragmented prediction results (even voids within some seismic facies data volumes) resulting from traditional semantic segmentation as a single classification objective function. This approach better addresses the technical problems targeted by this invention. This invention uses an improved U-Net network for seismic facies prediction to train and predict seismic facies. This ensures that the model's generalization results are closer to reality. Furthermore, the training dataset is a preprocessed and augmented small 3D data volume, which better addresses the geological problems of seismic facies identification from actual 3D data, and the identification results show good agreement with geological understanding. The improved U-Net network in this invention mainly involves increasing the number of network layers and network output, and designing a dual objective function based on regression and classification problems to ensure that the model's generalization results are closer to reality, thereby improving the identification speed and accuracy of the model.

[0147] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0148] Except for the technical features described in the specification, all other technologies are known to those skilled in the art.

Claims

1. A seismic facies prediction method based on deep learning semantic segmentation technology, characterized in that, This seismic facies prediction method based on deep learning semantic segmentation technology includes: Step 1: Construct a seismic facies recognition network model based on deep learning semantic segmentation; Step 2: Conduct seismic facies identification data preprocessing and data augmentation; Step 3: Conduct network training to obtain a seismic facies recognition training model; Step 4: Use the seismic phase identification training model to carry out seismic phase prediction.

2. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 1, characterized in that, In step 1, the seismic phase recognition network model is a deep learning semantic segmentation network structure model, which includes two parts: encoding and decoding, and has the following characteristics: it adopts fully convolutional layers; it uses pooling layers for downsampling and deconvolution operations for upsampling; it adopts skip connection layers; it uses BN layers and residual blocks; the network adopts single input and dual output, where the output corresponding to the one-hot encoded label requires softmax operation.

3. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 2, characterized in that, In step 1, the deep learning semantic segmentation network structure model is based on the traditional U-Net network, and a targeted improvement design is made for the seismic phase problem. It adopts a completely symmetrical network structure, that is, the downsampling encoding part on the left half of the network structure and the upsampling decoding part on the right half of the network structure are symmetrical. It uses skip chain layers to directly pass the features extracted in the downsampling stage to the upsampling stage and merge them with the data in the upsampling stage into feature data with a larger number of channels. However, compared with the traditional U-Net network, it adds two convolutional layers and one pooling layer in the downsampling stage, symmetrically adds two convolutional layers and one upsampling layer in the upsampling stage, increases the number of skip chains, and adds Dropout design to the max pooling in the downsampling stage.

4. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 2, characterized in that, In step 1, the application of BN layers in the seismic phase recognition network model refers to adding BN layers to all convolutional and deconvolutional layers to improve training stability, training speed, and generalization ability. The application of residual blocks in the seismic phase recognition network model refers to using a residual block structure based on the idea of ​​local skip links to replace general convolutional layers for computation in order to eliminate the risk of gradient vanishing that may occur in backpropagation speed calculation.

5. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 2, characterized in that, In step 1, the dual output of the seismic facies identification network model refers to the fact that, unlike the traditional U-Net network, the seismic facies identification network model adopts a dual-output design, which corresponds to two types of objective functions. Among them, network output 1 is a 3D convolution with 1 channel followed by BN and Sigmoid activation functions, and network output 2 is a 3D convolution with the number of channels equal to the number of seismic facies categories, followed by BN, ReLU activation functions, and Softmax operation.

6. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 5, characterized in that, In step 1, the objective function of the seismic facies identification network model adopts a dual objective function of mean square error (MSE) and cross-entropy (CE), and the two objective functions are weighted using the following formula to obtain the final objective function: loss=w1×loss mse +w2×loss ce In the formula, loss mse The objective function is the mean squared error, with a corresponding weight of w1, a network output of 1, and a label that is not one-hot encoded. loss ce The objective function is based on cross-entropy, with a corresponding weight of w2, a network output of 2, and a label that is a one-hot encoded label.

7. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 6, characterized in that, In step 1, the objective function based on cross-entropy refers to the multi-class weighted focal objective function derived from the cross-entropy problem of binary classification, and its formula is: loss ce =-w×(1-p pred ×p true ) γ ×log(p pred ) In the formula, w represents the weight of each category, and p true p represents the probability of each category on the true label. pred The outputs for each model category are represented by γ, which is the adjustment factor.

8. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 1, characterized in that, In step 2, seismic facies identification data preprocessing refers to edge padding and resampling of the data. Edge padding means padding the actual number of sample points in a single channel of seismic data with zero values ​​according to the number of sample points that are powers of 2, and adding a new category of zero to the corresponding label category. Resampling means coarsening and resampling the expanded data according to actual needs.

9. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 1, characterized in that, In step 2, seismic facies identification data enhancement refers to extracting small three-dimensional data volumes as samples from the seismic data of the training area in the line and xline directions according to certain rules, and rotating them horizontally at multiple angles to obtain more training samples.

10. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 1, characterized in that, In step 3, the training data obtained after preprocessing and data augmentation in step 2 is fed into the seismic phase recognition network model provided in step 1, and appropriate parameters are selected for training to obtain the seismic phase recognition training model.

11. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 1, characterized in that, In step 4, the model trained in step 3 is invoked to make multiple seismic phase predictions on the seismic data of the prediction area. The final output category of each numerical point in the data space is the category with the highest probability of occurrence in the multiple prediction results.

12. The seismic facies prediction method based on deep learning semantic segmentation technology according to claim 11, characterized in that, In step 4, multiple prediction of seismic phases refers to the use of a processing method similar to data augmentation in step 2, which extracts a large number of three-dimensional small data volumes from the test data, so that each seismic trace in the training area corresponds to multiple prediction results, and then the category with the highest probability is obtained for each trace and each sampling point as the final prediction result.

13. A seismic facies prediction system based on deep learning semantic segmentation technology, characterized in that, This seismic facies prediction system based on deep learning semantic segmentation technology uses the seismic facies prediction method based on deep learning semantic segmentation technology as described in any one of claims 1-12 to predict seismic facies in complex lithological zones in the field of petroleum geophysical exploration.

Images