CN111815680A - An automatic horizon tracking method for deep convolutional neural networks based on identity shortcut mapping - Google Patents

Abstract

The invention discloses an automatic horizon tracking method of a deep convolutional neural network based on constant fast mapping, which comprises the following steps: s1, preprocessing the seismic section image; s2, extracting, compressing and refining the seismic reflection features by using an identity shortcut connection convolution neural network; s3, further collecting local features and global features for the information extracted in the step S2; and S4, performing classification calculation on the collected features. The method takes the deep convolutional neural network as a backbone to build a model, takes the constant and quick connection as a core, changes the gradient flow of the traditional network, more effectively extracts the characteristics in seismic reflection, realizes the deep cascade polymerization of seismic phase characteristics, and improves the accuracy of horizon resolution.

Description

An automatic horizon tracking method for deep convolutional neural networks based on identity shortcut mapping

technical field

The invention relates to a deep convolutional neural network automatic layer tracking method based on identity shortcut mapping.

Background technique

There is a very important correlation between seismic horizon interpretation and sequence stratigraphic division. Accurate horizon information is the basic work of seismic interpretation, and horizon identification and tracking are an important part of seismic horizon interpretation. Resolve automatic tracking of the event axis. Seismic data are seismic waves generated by explosives or seismic vehicles on the ground through explosions or heavy blows. When they encounter the subsurface, they will emit reflected waves and receive signals by the geophones on the surface. Finally, reflection data such as seismic amplitudes are obtained. . Seismic interpreters generally consider peaks or troughs in a profile of seismic data to be locations where the event axis passes. The conventional artificial horizon marking is to determine the position of the event axis through the peaks and troughs of the seismic section and connect them, and finally connect the horizons marked on each section to form the event axis surface. Due to the large cost and time required for manual marking, researchers have developed methods for automatic horizon tracking. Based on the above principles, researchers find similar points within a certain range by calculating the waveform correlation coefficient or coherence, and use them as the points that the event axis passes through, and finally delineate the event axis. However, when faults or bifurcations occur, calculation errors are prone to cause problems with the final marking. The 3D seismic data can roughly reflect the complex strata under the surface of the detection area. Since most of the underground structures are distributed in layers, there are different strata produced in different ages at different depths underground, and the density of the stratum itself is different due to the different nature. The difference causes the seismic waves to travel at different speeds here, which ultimately enables people to obtain different seismic reflection data. The density of the formation is also called the density of the rock, which together with the propagation speed of the sound wave on the rock is called the petrophysical property, or lithological parameter for short. The lithological parameters of different formations are very different, and the wave impedance is the product of the above two, which means that the wave impedance will also change significantly in different formations. Due to the difference in lithological parameters, the sound wave will reflect when it propagates to the junction of the two formations, and finally such reflection is received by the sensors on the surface to generate various waveforms, and the waveforms with strong changes and consistent and adjacent to each other are It is called the event axis, that is, the horizon. Therefore, researchers believe that horizons can express the subsurface sequence division and the subsurface lithology distribution structure. A horizon is generally composed of multiple points, and each point has a corresponding line number and time (depth). Generally, there are multiple horizons in the 3D seismic body obtained from the exploration area, but it is necessary to find the corresponding horizons for all horizons. The level to which it belongs, which reflects the importance and significance of level tracking.

At present, there are mainly the following layer tracking methods:

(1) Automatic horizon tracking based on correlation

It is a traditional automatic horizon tracking technology, mainly using the coherent algorithm proposed in the past. The stability and resistance to noise are good, but the demand for computing power is quite large, and when the waveforms of the two adjacent layers are very similar, it is easy to produce a situation of cascading layers, which may lead to insufficient results. precise.

(2) Image-based automatic tracking method

Hale and Naeini et al. used structure tensor in the field of horizon tracking and fault identification. Through some directional filters, the orientation of specific features was obtained, and finally the main orientation of horizon extension and breeding was obtained. The calculation efficiency of this method is high, but if the stratum with complex structure is encountered, the accuracy and stability of the tracking cannot be guaranteed.

(3) Automatic tracking method based on neural network

Huang, Lu et al. published an unsupervised automatic horizon tracking method using a self-organizing map network, but the success or failure of this scheme depends on the initialization of the self-organizing map network, which is directly related to the quality of the results. Alberts, Huang et al. used artificial neural networks to track horizons. This method requires more training data and requires data representativeness, but the results are good. Due to the rapid development of artificial intelligence itself and deep learning, computing power-related technologies such as data warehouses and graphics processors have also developed rapidly. The convolutional neural network has been proposed to be applied to common applications such as image recognition, NLP, and regression. , in which Alex et al. applied CNN to image classification and achieved good results, further detonating the popularity of deep learning.

(4) Coherence-based automatic tracking method of event axis

At present, the field of seismic exploration has developed to the third-generation coherent technology, which calculates a certain intrinsic physical connection, or similarity, between the seismic trace and adjacent traces in a certain way, which is called coherence. The calculation of the degree of coherence can represent the lateral change in the formation. If the value is large, it means that the lateral change here is small, the similarity between adjacent traces is large, and the coherence is strong. Where the value is small, it indicates that there may be faults or abnormal bodies, the lateral variation of the rock formation is large, and the coherence is weak in this place. Coherence-based automatic horizon tracking operates on seismic reflections using coherent techniques and ultimately results in a coherent volume that can account for changes in lithological parameters. By extracting features from a specific area of the coherent volume, and finally using interpolation technology to obtain a complete horizon, this is the currently more popular commercial coherence-based automatic horizon tracking technology.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a model built with a deep convolutional neural network as the backbone, with constant and fast connections as the core, changing the gradient flow of the traditional network, and more effectively extracting the The deep convolutional neural network automatic horizon tracking method based on identity shortcut mapping can more accurately track the target horizon.

The object of the present invention is to be achieved through the following technical solutions: a deep convolutional neural network automatic horizon tracking method based on identity shortcut mapping, comprising the following steps:

S1. Preprocess the seismic profile image;

S2. Extract, compress and refine seismic reflection features by using the identity shortcut connection convolutional neural network;

S3, further collect local features and global features for the information extracted in step S2;

S4, classify and calculate the collected features.

Further, the step S1 specifically includes the following sub-steps:

S11. Divide the entire seismic section into a plurality of small image blocks, and each small image block is 32×32 in size;

S12, assigning a label to each small image block, the specific implementation method is: judging whether the center of each small image block is located on the horizon, if so, marking the image block as 1, otherwise, marking it as 0;

S13, perform two-class training on the image block data to determine whether the center position of the image block is on the horizon;

S14. Process the non-horizontal data: set the loss probability to reduce the number of obtained non-horizontal image blocks, and set the loss probability to 0.5, that is, randomly select 50% of the non-horizontal data to discard; Set the deviation value. If the distance between the center point of a small image and a certain horizon is within the deviation value range, even if the data is manually marked as non-horizontal, the data will be regarded as horizon data. is 1;

S15. Extract 20% from the horizon data set as verification data, which is used to adjust the model hyperparameters; the rest of the data is used as training data;

S16. Normalize the data: use the Min-Max normalization method to normalize the data scale to the range of 0-1; the Min-Max normalization method is realized by the following formula:

where X _nor is the normalized data, X is the original seismic data, and X _max and X _min are the maximum and minimum values of the seismic amplitude, respectively.

The beneficial effects of the present invention are:

1. The present invention takes the data mining feature classification as the starting point, and proposes a deep convolutional neural network automatic horizon tracking method based on constant shortcut connections. The deep convolutional neural network is used as the backbone to build a model, with constant and fast connections as the core, which changes the gradient flow of the traditional network, extracts the features of seismic reflections more effectively, and realizes the deep cascade aggregation of seismic phase features. The invention has a good tracking effect on the high dip angle horizon and the weak seismic amplitude reflection horizon in the complex multi-fault zone, and improves the accuracy of horizon resolution.

2. The present invention proposes a new idea of data pre-processing and post-processing in order to solve the problem of data imbalance by using the data-driven method to solve the problem of horizon tracking. Track the target horizon.

3. Compared with the ordinary convolutional neural network model, the model of the present invention has stronger generalization performance and better feature expression ability when the data distribution is more different than the original data distribution, and is more suitable for tracking Complex target horizons.

Description of drawings

Figure 1 is a comparison of neural network connection methods;

Figure 2 is a schematic diagram of the convolution process;

Figure 3 is a schematic diagram of the residual function F(x);

Fig. 4 is the model frame structure proposed by the present invention;

5 is a flowchart of the deep convolutional neural network automatic horizon tracking method based on identity shortcut mapping of the present invention;

Figure 6 is a volume image of 3D seismic data in a work area in southwest China;

7 is a schematic diagram of the division of small image blocks;

Fig. 8 is a schematic diagram of horizon data and non-horizon data;

Figure 9 is a two-dimensional cross-sectional view of inline1760 in a work area in the southwest;

Figure 10 is a two-dimensional cross-sectional view of inline1762 in a work area in the southwest;

Figure 11 is the probability distribution map of the inline1762 horizon in a work area in the southwest;

Figure 12 is a comparison diagram of the three methods in a work area in the southwest of inline1762;

Fig. 13 is an enlarged view of the block portion in Fig. 12;

Figure 14 is a two-dimensional cross-sectional view of inline1775 in a work area in the southwest;

Figure 15 is the probability distribution map of the inline1775 horizon in a work area in the southwest;

Figure 16 is a comparison diagram of the three methods in a work area in the southwest of inline1775;

FIG. 17 is an enlarged view of the block portion in FIG. 16 .

Detailed ways

The technical principle relevant to the present invention:

Convolutional Neural Network

A convolutional neural network is a special kind of deep feed-forward neural network that consists of neurons with weights and biases. By adopting the method of partial connection, the redundancy problem caused by the full connection is effectively avoided, and the parameter quantity of the network is effectively reduced, thereby reducing the dependence of the model on the data itself. For example, each neuron is only connected to 10×10 pixel values. The neural network connection mode is shown in Figure 1, (a) is the original connection (full connection), (b) is the new connection mode (partial connection):

A convolutional neural network generally consists of a convolutional layer, a Relu nonlinear activation layer, a pooling layer and a fully connected layer. Among them, the pooling layer can be selected according to the situation of the task itself. The default input data format of the convolutional neural network is depth × height × width. If the input is an image, the depth represents the number of channels of the image, and the latter two represent the resolution of the image. In general, the channels used for color images are RGB three-channel by default, and the default channel for gray-scale images is 1. Each layer of the convolutional neural network is locally connected to the previous layer.

计算得出，其中W为输入的原始图像的尺寸，F为卷积核尺寸，S为步长，表示卷积核每次在图像上滑动经过的像素个数。P为填充数，表示当经过卷积操作后的特征图大小与卷积核大小不满足点积运算计算要求时，将特征图外部用0进行填充补齐大小的一种方式。在卷积神经网络中，卷积的数学表达式和常规的二维卷积有些许不同，其表达如下：Convolutional layers extract and collect features through convolution operations, so this type of layer undertakes most of the computation of the entire neural network. In a normal convolutional neural network, the convolutional layer is usually placed after the input layer or the pooling layer. During forward propagation, use a fixed-size convolution kernel to walk on the input image with a set step size and perform convolution calculations at the same time. output. In order to further illustrate the working process of the convolutional layer, an example is described. As shown in Figure 2, the size of the original image input by the network is 5 × 5, and a 3 × 3 convolution kernel is used to extract image features. Output of size 3x3. The left side and the top two lines in the figure are the original images of the input, the lower right part represents a 3×3 convolution kernel, and the right image represents the network output after filtering by the convolution kernel. The size of the output value is determined by the step size of the network and the size of the padding value. Available formulas

It is calculated, where W is the size of the input original image, F is the size of the convolution kernel, and S is the step size, which represents the number of pixels that the convolution kernel slides on the image each time. P is the padding number, which means that when the size of the feature map after the convolution operation and the size of the convolution kernel do not meet the calculation requirements of the dot product operation, the outside of the feature map is filled with 0 to fill the size. In the convolutional neural network, the mathematical expression of convolution is slightly different from the conventional two-dimensional convolution, which is expressed as follows:

s(i, j)=X*W(i, j)=∑ _m ∑ _n x(i+m, j+n)w(m, n) (2)

When the multi-channel convolution kernel is input, the function expression is as follows:

where inchannel is the number of input channels, X _k represents the k-th input matrix, and W _k represents the k-th sub-convolution kernel matrix of the convolution kernel. s(i, j) is the corresponding element value of the output matrix corresponding to the convolution kernel W.

If the feature size obtained after the convolutional layer operation is still too large, which is not conducive to subsequent network construction, it can be input to the pooling layer for downsampling compression. Such an operation can avoid the skyrocketing amount of parameters caused by excessive resolution, and at the same time can reduce the degree of overfitting to a certain extent. The usual practice is to select an appropriate window size, set an appropriate step size, and take the maximum value, average value or random sampling of the data in the window. These three schemes represent maximum pooling, average pooling and random pooling respectively. . Generally speaking, max pooling is the most common pooling method. This method can usually retain the original features better while diluting the data resolution, and at the same time reduces the dimension, facilitates calculation, and enhances the generalization ability. .

Fully connected layer and random loss: After going through the convolutional layer and the pooling layer, the final feature needs to be extracted through the fully connected layer, and the final dimension is integrated according to the required task. This layer compresses features into vectors, so this layer is also called a feature vector layer. Since it is different from the local connection of the convolutional layer, the full connection operation is used, which leads to an increase in the amount of parameters. The dropout method can be used. By setting a loss probability, the neurons are randomly selected and lost. This method can reduce overfitting. At the same time, it reduces the amount of network parameters, reduces the amount of calculation and improves the training efficiency. Dropout can also be used in other layers with weights, such as convolutional layers.

Residual network: The deep residual network was proposed by He et al. to solve the problem of gradual degradation of network performance due to the deepening of the network depth. According to the universal approximation theorem, given sufficient capacity, any function can be expressed by a single-layer feedforward neural network. However, this premise may lead to unimaginably large dimensions of such layers, increasing the probability of overfitting. Therefore, the academic community generally believes that the multi-layer network architecture is superior to the single-layer high-capacity network architecture. AlexNet has achieved good results in image recognition tasks, and since then the state-of-the-art network architecture has gradually deepened, in contrast, AlexNet has only 5 convolutional layers. In its development, the follow-up VGG network and GoogleNet (codenamed Inception_v1); there are 19 layers and 22 layers respectively. However, under the limitation of vanishing gradients, the number of network layers cannot be easily increased by stacking, and for the above reasons, such a network is extremely difficult to train. When backpropagation is caused by gradient confusion, the gradient passed to the previous layer may be close to 0 under continuous multiplication. As a result, after the number of network layers is deepened to a certain limit, the network performance reaches a bottleneck or even begins to decline. The core idea of ResNet is to introduce a so-called identity shortcut connection, which directly skips one or more layers, as shown in Figure 3.

The residual learning unit establishes a direct correlation channel between input and output through the introduction of identity mapping, so that the powerful parameterized layer can concentrate on learning the residual between input and output. Generally, F(x) is used to represent the residual mapping, then the output is: y=F(x)+x. When the number of input and output channels is the same, you can directly use x for addition. When the number of channels between them is different, it is necessary to consider establishing an effective identity mapping function so that the processed input x and the output y have the same number of channels, ie y=F(x)+W*x. The emergence of residual networks makes it possible for people to build deeper networks. At the same time, the use of "jump-layer" connections such as constant shortcut connections achieves the effect of increasing feature diversity and speeding up training.

The network proposed by the invention is composed of two parts: the first half uses the identity and shortcut connection convolutional neural network to extract, compress and refine the seismic data; obtain a 256*8 data body; the second half uses the 256*8 data The body becomes 4 local structures of 256*4 and a global structure of 256*8. As input, the latter part further collects local features (referring to the features extracted from the original input data as a whole) and global features ( Extract the features of small image patches whose scale is relatively small relative to the original input), and use the collected final features for classification calculations. The module structure in Figure 3 is used in the first half of the network. The first half includes a convolution layer and five module structures. Each module structure contains a convolution layer and a Relu operation, and is accompanied by an identity shortcut connection; every two The module makes the convolution step size 2 (referring to the moving step size of the convolution kernel, that is to say, the step size of the convolution kernel is changed to 2 in the calculation of every two modules) and performs a dimensionality reduction operation to halve the dimension; In the second half, 1/4 of the global structure (an overall data of the original input) is understood as a local structure, and the modular operation shown in 3 is performed on each local structure and the global structure to obtain four 256*4 local structures. and a 256*8 global structure; then average pooling is performed to obtain five vectors with a dimension of 1×1×256, representing four local features and one global feature respectively, and the above features are connected. For the feature vector of classification, input the fully connected layer to further reduce the dimension, and use the softmax function for binary classification. The model frame structure proposed by the present invention is shown in FIG. 4 .

Table 1 is a preview of the internal architecture of the entire network.

Table 1 Deep convolutional layer tracking network architecture based on identity shortcut connections

The input data dimension is time×xline, followed by the comma; the kernel is the size of the convolution kernel, and the block uses square brackets to indicate two consecutive convolution layers; Stride is 2 to indicate the block's first layer convolution step size is 2, the rest is 1; block6_local* means extracting the local information of the corresponding part, block6_global extracting the overall information; avg_pool_concat means that all data bodies are connected after average pooling; fc means the fully connected layer; softmax maps the output to [0,1 ] interval; the table is not completely represented in the order of the information flow, and the information flow follows that shown in Figure 5.

As mentioned earlier, each block inside can be expressed by the following formula:

y=F(x, {W _i })+x (4)

Here x and y are the input and output vectors of the layer under consideration. The function F(x, {W _i }) represents the residual map to be learned. Wherein F=W ₂ σ(W ₁ ), σ(W ₁ ) represents the activation function, the present invention uses ReLU as the activation function, and for the use of abbreviated symbols, the deviation is omitted here. F(x, {W _i })+x is realized by identity shortcut connection, and the corresponding addition of elements is used at the same time. The quadratic nonlinear operation σ(y) is performed after the corresponding addition. The constant shortcut connection in formula (4) does not introduce additional parameters nor computational complexity, but the dimensions of x and F must be unified. If they are not unified (for example, when changing the number of channels), a linear mapping can be used. to match the number of channels:

y=F(x, {W _i })+W _s x (5)

In the present invention, 1×1 convolution is used to achieve the effect of W _s . Since the size of the 1×1 convolution kernel is only 1×1, it does not need to consider the relationship between pixels and surrounding pixels. It is mainly used to adjust the number of channels. The pixels on different channels are linearly combined and then non-linearized to complete the functions of dimension raising and dimension reduction.

In the network structure design of the present invention, this function is often only required in some blocks. It can be noted that the residual functions in Equation 4 and Equation 5 above are expressed by a fully connected layer for simplicity, and a convolutional layer can be used instead in actual use. In practice, multiple convolutional layers can also be used to express F(x, {W _i }), and the element-by-channel addition is performed on the two feature maps.

Comparing the back-propagation mechanism of the model of the present invention with the conventional model, it is assumed that when an ordinary block is executed, the back-propagation of the loss to the weight is:

And when the module uses a constant shortcut connection, the back-propagation of the gradient is:

L represents the loss function, l ₁ represents the first layer network, l ₂ represents the second layer network, o represents the output of the overall network, o ₁ represents the output of the first layer of the network, and W represents the weight matrix. This formula is a simple two-layer network formula, mainly to compare the difference between using a constant connection and not using it. Compared with the module without constant connection, there is an extra 1 at W, which transfers the gradient back without decay, which enables the network to build deeper seismic features and extract subtle features that conventional convolutional neural networks are not sensitive enough to. , which makes the classification effect more accurate.

The technical solutions of the present invention are further described below with reference to the accompanying drawings and specific embodiments. In the following, the abbreviation ISGL-Net of the English Identity Shortcut Global Local Convolutional Neural Network will be used to represent the network proposed by the present invention.

As shown in Figure 5, the deep convolutional neural network automatic layer tracking method based on identity shortcut mapping of the present invention includes the following steps:

S1. Preprocess the seismic profile image;

This example uses a 3D seismic volume in a work area in the southwest for the experiment. The seismic interpretation profile (inline) of the main survey line of the data volume has a total of 401 line numbers from line number 1760 to line number 2160, and the tie line seismic interpretation profile (xline) is represented by line number. There are 440 line numbers from 1780 to line number 2220, the sampling interval of this sample is 2ms, and the time range is 0ms to 1300ms. The three-dimensional seismic volume is shown in FIG. 6 .

This step specifically includes the following sub-steps:

S11. The entire seismic section contains too much information to be directly used as an input image. The entire seismic section is divided into several small image blocks, which are convenient for calculation and analysis; each small image block is 32×32 in size; Figure 7 shows the workflow for generating training data. The manually marked horizons are in the middle of the left image, and a window (the small window shown in the upper left corner) is slid across the entire seismic section to create a 2D image patch of size 32×32. Each 32x32 image has a width and height of 0-31 and a sampling interval of 2ms, which means that each image block used is 32 channels, 64ms and one color channel.

S12. After extracting the image blocks, a label needs to be assigned to each small image block, so that the image patches with the same label share specific features; the specific implementation method is: judging the center of each small image block (if each small image block It is placed in the two-dimensional coordinate axis, and one of its vertices is located at the origin, then the coordinate point of (15, 15) is used as the center of the picture) whether it is located on the horizon, if so, the image block is marked as 1, otherwise it is marked as 0 ; As shown in Figure 8, the small circle in the figure represents the center of the image block;

S13, perform two-class training on the image block data to determine whether the center position of the image block is on the horizon;

S14. In the process of extracting data, due to the characteristics of the horizon data itself, the phenomenon that the positive and negative samples are seriously unbalanced will occur. For the number of horizon image blocks, set the loss probability to 0.5, that is, randomly select 50% of the non-horizontal data to discard; the loss probability is used to alleviate the phenomenon of data imbalance, that is, to discard half of the non-horizon image block data, Alleviates the situation that the center of too many data blocks is not on the horizon.

At the same time, the deviation value is set on the horizon. If the distance between the center point of a small image and a certain horizon is within the deviation value range, even if the data is manually marked as non-horizontal, the data will be considered as Horizon data, marked as 1, the deviation value used in this embodiment is 10 data points;

In classification problems, when the sample sizes of different classes vary greatly, that is, the class distribution is unbalanced, it is easy to affect the classification results. Therefore correction is required. The specific method is to weight each category in the training set, so that the model pays more attention to the category with a small number of samples. The weight method is used, that is, the class with a large number of samples has a low weight, and vice versa, the loss is allocated. The weight used in the present invention is the inverse of the number of samples corresponding to the class; the class here is the horizon data marked as 1 and the non-horizontal data marked as 0. For example, if the number of data marked 0 is 90, its weight is 1/90; the number of data marked 1 is 10, its weight is 1/10.

In subsequent experiments, it was found that the above scheme can greatly reduce the imbalance of positive and negative samples, and can greatly improve the accuracy of the model.

S15. Extract 20% from the horizon data set as the validation data to adjust the model hyperparameters (according to the validation loss (validation error obtained when validating with the 20% of the training data) to adjust the parameters, or in When training, keep the model with the best performance in the checkpoint setting); the rest of the data is used as training data;

S16, normalize the data: data normalization is an essential step in the process of machine learning data preparation. Normalization refers to normalizing the input data in all dimensions within a fixed scale. The amplitude property of 3D seismic data ranges from -32767 to 32767, however if the ReLU activation function receives any negative input it returns 0 so that no learning will take place in this zero value region. Once the units become negative, they play no role in differentiating the input and are basically useless. If the input value is too large, it will affect the convergence of the neural network. Therefore, the input data needs to be normalized to improve the convergence speed and training efficiency. This embodiment uses the Min-Max normalization method to normalize the data scale to the range of 0-1; the Min-Max normalization method is implemented by the following formula:

where X _nor is the normalized data, X is the original seismic data, and X _max and X _min are the maximum and minimum values of the seismic amplitude, respectively.

S2. Extract, compress and refine the seismic reflection features by using the constant shortcut connection convolutional neural network; input the experimental data after preprocessing into the model, and the experimental training processor used in the present invention is NVIDIA GeForce GTX1050 (4g), which is similar to Using a graphics processor can speed up model training involving convolutional neural networks rather than a central processing unit.

S3, further collect local features and global features for the information extracted in step S2;

S4, classify and calculate the collected features.

The following is tested with the prepared data and analyzed with the results.

In this example, a horizon that is seriously disturbed by faults and has a large dip angle in a work area in the southwest is selected, and this horizon is considered to be typical. Figure 9 shows the reflection pattern of the target horizon, and Figure 9 shows a work area in the southwest of inline1760 II Dimensional section, respectively marking the original horizon, fault and high dip.

First, the horizons of the inline1762 section, which is closer to the training data, are used as shown in Figure 10. Figure 10 is a two-dimensional section of inline1762 in a work area in the southwest, with the original horizons, faults and large dips marked respectively. And compared with the ordinary deep convolutional neural network and the traditional automatic horizon tracking method, it can be seen that the horizon is very similar to the profile inline1760 due to the close proximity of the profile.

Figure 10 shows the horizon probability distribution of the two methods of CNN and ISGL-net on inline1762. It can be seen that both of them handle two faults and large dip angles well, at the beginning of the leftmost horizon and the first At the fault, the conventional CNN has some defects (Fig. 11(a)), and ISGL-net (Fig. 11(b)) is generally clearer and cleaner than the conventional convolutional neural network. Although the conventional CNN has some of the above-mentioned flaws, the experimental results are still in an acceptable state, and such results are predictable, because the physical locations of the training and test profiles are close, which makes the training data and test data indistinguishable. distribution is relatively close.

Next, compare the results of the two thresholds and remove the deviation with the results of the traditional horizon tracking method. The threshold set in this embodiment is 0.9, as shown in FIG. 12 . Conventional represents the traditional method; Manual represents manual marking. It can be seen that CNN/ISGL-net and the traditional waveform similarity-based event tracking method have good performance in the front area with little longitudinal change. In the region of large dip in the last part, several methods have good performance. Enlarge the block in Figure 12, as shown in Figure 13.

It can be seen from Figure 13 that the traditional horizon tracking method has a serious misjudgment at the first fault, traces the adjacent horizon, and is still misjudged at the second fault, and in this area, CNN and ISGL -net still maintains good consistency with human labeling;

Next, each method uses the artificially marked horizon as the standard to quantify the error. This paper uses the absolute error (Absolute Error) and the mean absolute error (Mean Absolute Error) to measure the distance error. The formula is as follows:

where m is the number of sampling points in the test set. Each method is based on the manually marked horizons. The judgment error at the fault is given to each judgment point with a penalty value of 20. The errors for the inline1762 horizon are shown in Table 2.

Table 2 Comparison of inline1762 horizon errors in a work area in southwest China

AE MAE CNN 1002 3.13 ISGL-net 968 3.03 conventional 1149 359

Combining Figure 12 and Table 1, it can be seen that the effect of ISGL-net is better than the other two, and the effect of CNN is close to that of ISGL-CNN. However, the actual performance and quantization results of traditional horizon tracking methods are not satisfactory. Since the physical distance between inline1762 and the training data inline1760 is relatively close, the data distribution of the two is relatively close. In order to test the generalization performance of the model, this embodiment selects inline1775, which is farther from the actual physical location of the original training data, for testing. Inline1775 and horizon calibration are shown in Figure 14. It can be seen that the profile data distribution is quite different from inline1760.

Figure 15 is the inline1775 horizon probability distribution image of a work area in the southwest, where (a) is the CNN processing result, and (b) is the ISGL-net processing result. It can be seen that CNN has made a large error prediction at the two faults marked by the box in Figure 15, while the ISGL-net proposed by the present invention has performed a relatively successful fault detection on the first fault. The second fault has a prediction error, but it is still better than CNN in terms of resolution. In the slight longitudinal change and the large inclination part, the resolution effect of the two is relatively good.

In the same way, the result is still set the threshold value and de-biased value operation, and compared with the traditional horizon tracking method, the result is shown in Figure 16. In the front area with little longitudinal change, CNN/ISGL-net and traditional horizon tracking methods have good performance. In the region of large dip in the last part, all three methods perform well.

Next, enlarge the marked part of the large box. As shown in Figure 17, it can be observed that within the range of the box mark, the traditional horizon tracking method has continuous tracking errors at two faults, both of which are traced to the adjacent horizons, and there are four traces including the fault. Error; within the range marked by the first blue box, it can be observed that CNN has a serious prediction error at the first transverse fault, while ISGL-net performs well here; within the range marked by the second blue box , in the second fault, both CNN and ISGL-net have different degrees of tracking errors, but overall, ISGL-net is still better than CNN. Next, for each method, the inline 1775 artificially marked horizon is used as the standard to quantify the tracking error, and the errors are shown in Table 3.

Table 3 Comparison of horizon error of inline1775 in a work area in southwest China

AE MAE CNN 1325 4.14 SGL-net 1162 3.63 conventional 1496 468

To sum up, ISGL-net has better generalization performance than CNN when the physical distance is far from the original training data and the data distribution gap is large. Compared with CNN and traditional waveform-similar-based horizon tracking methods, ISGL-net has better tracking effect.

Those of ordinary skill in the art will appreciate that the embodiments described herein are intended to assist readers in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations without departing from the essence of the present invention according to the technical teachings disclosed in the present invention, and these modifications and combinations still fall within the protection scope of the present invention.

Claims (2)

1. The deep convolutional neural network automatic layer tracking method based on identity shortcut mapping, is characterized in that, comprises the following steps: S1. Preprocess the seismic profile image; S2. Extract, compress and refine seismic reflection features by using the identity shortcut connection convolutional neural network; S3, further collect local features and global features for the information extracted in step S2; S4, classify and calculate the collected features. 2. the deep convolutional neural network automatic horizon tracking method based on identity shortcut mapping according to claim 1, is characterized in that, described step S1 specifically comprises following substep: S11. Divide the entire seismic section into a plurality of small image blocks, and each small image block is 32×32 in size; S12, assigning a label to each small image block, the specific implementation method is: judging whether the center of each small image block is located on the horizon, if so, marking the image block as 1, otherwise, marking it as 0; S13, perform two-class training on the image block data to determine whether the center position of the image block is on the horizon; S14. Process the non-horizontal data: set the loss probability to reduce the number of obtained non-horizontal image blocks, and set the loss probability to 0.5, that is, randomly select 50% of the non-horizontal data to discard; Set the deviation value. If the distance between the center point of a small image and a certain horizon is within the deviation value range, even if the data is manually marked as non-horizontal, the data will be regarded as horizon data. is 1; S15. Extract 20% from the horizon data set as verification data, which is used to adjust the model hyperparameters; the rest of the data is used as training data; S16, normalize the data: use the Min-Max normalization method to normalize the data scale to the range of 0-1; the Min-Max normalization method is realized by the following formula:

where X _nor is the normalized data, X is the original seismic data, and X _max and X _min are the maximum and minimum values of the seismic amplitude, respectively.

Images