CN113902769A

CN113902769A - Seismic fault identification method based on deep learning semantic segmentation

Info

Publication number: CN113902769A
Application number: CN202110946552.1A
Authority: CN
Inventors: 胡广; 杨胜雄; 李沅衡; 田冬梅; 曹荆亚; 邓雨恬
Original assignee: Southern Marine Science and Engineering Guangdong Laboratory Guangzhou
Current assignee: Southern Marine Science and Engineering Guangdong Laboratory Guangzhou
Priority date: 2021-08-18
Filing date: 2021-08-18
Publication date: 2022-01-07
Anticipated expiration: 2041-08-18
Also published as: CN113902769B

Abstract

The invention discloses a seismic fault identification method based on deep learning semantic segmentation, which comprises the following steps: acquiring three-dimensional seismic data, and performing slice extraction to obtain a two-dimensional seismic amplitude image; inputting the obtained two-dimensional seismic amplitude image into a fault segmentation model, and inputting a fault binary image into the fault segmentation model; optimizing the fault binary image to obtain a fault identification result; the fault segmentation model is trained by fault training samples and a deep learning semantic segmentation model. The fault identification method is based on a deep learning semantic segmentation method, adopts the idea of pattern identification to learn the fault interpretation pattern by virtue of the strong fitting capability of a deep convolutional neural network, utilizes computer vision to pick up the fault in the seismic amplitude image, realizes the rapid and accurate fault identification in the three-dimensional seismic data, reduces the occurrence of human intervention and errors in the identification process and shortens the time required by fault interpretation.

Description

Seismic fault identification method based on deep learning semantic segmentation

Technical Field

The invention relates to a seismic data processing technology, in particular to a seismic fault identification method based on deep learning semantic segmentation.

Background

Fault interpretation is the basis of seismic data interpretation, accurate fault interpretation is directly related to the accuracy of a constructed map, and how to accurately identify fault construction from seismic data is a problem which is always troubling learners. At present, fault identification methods are various, and generally, the methods are roughly divided into a conventional seismic fault interpretation method, an identification method based on seismic attributes, an automatic fault identification method and a fault identification method based on an image processing technology. The conventional fault interpretation method comprises the steps of manually researching reflection wave event axis abnormality (dislocation, local change, sudden increase or disappearance and the like) in a seismic section, analyzing seismic attributes (coherence, variance, curvature and the like) and the like, and further judging the characteristics of the fault such as the trend, the dip angle, the extension and the like. However, with the continuous improvement of seismic exploration technology, the volume of seismic data is larger and larger, fault interpretation by using a conventional method is very trivial and time-consuming, difficult and incapable of repeated verification, and the complex interpretation flow has high requirements on the expertise of interpreters

Although the method of fault identification using a method of calculating seismic attributes characterizing a fault is continuously developed and advanced, there are some drawbacks because seismic attribute selection and calculation are not very simple. In addition to conventional fault interpretation methods, automatic and semi-automatic fault interpretation methods have been the focus of research by researchers, including ant tracing algorithms, BP neural networks, AFE methods, and the like. Although the methods can overcome errors caused by visual resolution of naked eyes in the process of identifying the artificial fault to a certain extent, the problem of fault interpretation is very complex and is influenced by various factors, so that any current fault automatic or semi-automatic identification method has certain limitations. In addition, an image processing technology is utilized to carry out fault identification, another idea is provided to visually express seismic data in an image form, and the expression capability of image ground information is improved, for example, an edge detection technology, a three-color mixing technology and the like are applied to fault interpretation.

The image processing technology can display important information in the image without generating new information, has strong multi-scale property, and reduces subjectivity in the process of explaining the fault layer. However, the detected object is not seismic data, but is an image which is processed to reflect the seismic data, so that certain requirements are made on the seismic data processing precision. Although not few studies on fault interpretation methods, fault interpretation techniques need to be improved continuously in terms of practicality, credibility, limitations and the like, and need to be summarized and innovated continuously in practice, so that it is also very important and necessary to study fault identification methods in combination with the latest development techniques.

In recent years, with the rapid development of artificial intelligence in various industries, more and more deep learning methods are always applied in seismic exploration, methods combined with convolutional neural networks are also used in seismic data interpretation, and many researchers also combine seismic fault recognition with the deep learning methods to obtain some results, but some problems exist more or less at the same time. For example, Axelle (2018) extracts accurate fault locations by building a synthetic seismic dataset with simple fault geometry, applying a classification strategy to the image, and performing a simple post-processing. The method constructs a network and shows a good interpretation result on synthetic data, and tests are carried out on The actual section of The Netherland offset F3 block, so that an encouraging fault recognition result is obtained. However, the network constructed by the method is very simple, and the actual seismic data is predicted by training the synthetic seismic data set, so that the actual effect is not ideal.

Guo Bowen (2018) and the like simulate and explain the working mode of personnel, and provide a method for automatically detecting a fault directly from a three-dimensional seismic amplitude image based on a convolutional neural network. Although the test result of the synthetic three-dimensional image shows that the convolutional neural network has higher accuracy in predicting the fault in a new seismic image, the method only carries out fault identification on the network construction, and does not further consider the optimization processing of the identification effect. Yaoxingsheng (2018) and the like invent a fault plane recognition method based on a full convolution neural network, and the patent comprises the steps of obtaining seismic amplitude fault data, constructing a full convolution neural network model, training the full convolution neural network training model and then recognizing the seismic amplitude fault data. However, this patent only uses a full convolution network for fault recognition, and the full convolution network is a semantic division mountain-opening operation, and can realize end-to-end image recognition. However, the full convolutional layer neural network does not sufficiently consider the semantic relationship between adjacent pixels, and thus the logical and consistency of context information is poor, and the obtained result is not fine enough. Although the skip structure has a rough problem for processing results, the up-sampling result has the defects of blurring and smoothing, and the result is not sensitive to understanding of detailed parts in the image.

Three orthogonal surfaces with one point in a synthetic seismic data body as a center are selected as a single training sample, the value of the center point in a corresponding label data body is a label, and a convolutional neural network structure transformed from an AlexNet network is constructed for automatic fault identification. Although the faults in the synthetic seismic data and the actual seismic data in the prediction set can be accurately identified, the network needs to consider the information of three orthogonal surfaces in order to predict whether the fault of a central point is the fault or not, which causes a large amount of repeated information calculation during training and is very inconvenient for predicting the fault of the central point. Wuxin (2019) regards the fault recognition problem as a binary image segmentation problem by marking the fault position on the three-dimensional seismic image as a value 1 and other non-fault positions as a value 0. And the image is effectively segmented by adopting a supervised full convolution neural network U-net. The fault in the three-dimensional seismic image can be predicted more accurately and effectively than the fault in the three-dimensional seismic image can be predicted by the traditional method. However, in the method, the synthetic seismic data is used for training to predict the actual seismic data, a large amount of synthetic seismic data and fault labels thereof need to be generated for training, and in an actual situation, because different actual seismic data of an acquisition mode and a processing method may have more or less difference, the fault recognition effect is unstable when the synthetic seismic data is used for training a model. Zhang Qie (2019) proposes a deep convolutional neural network model trained from synthetic data for automatic pick-up of three-dimensional faults. The network model divides the image center into whether a fault exists or not, and predicts the fault dip angle and the azimuth angle at the same time. Although in practical data applications it has been shown that such convolutional neural network models perform well in a variety of seismic images of different tectonic areas and produce consistent high quality fault picks. However, the network still performs fault judgment according to the image center, and the image center is used as the input and the output of the convolutional neural network to implement training and prediction processes, so that the storage cost is high, the calculation efficiency is low, the size of a sensing area is limited by the size of an image block, and the workload during prediction is increased. Through research on the background of the related art, at present, although many fault recognition researches are carried out by utilizing a convolutional neural network, problems still exist in the aspects of network structure design, sample selection and fault recognition results.

Disclosure of Invention

In order to solve at least one technical problem in the background art, the invention provides a seismic fault identification method based on deep learning semantic segmentation.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a seismic fault identification method based on deep learning semantic segmentation comprises the following steps:

acquiring three-dimensional seismic data, and performing slice extraction to obtain a two-dimensional seismic amplitude image;

inputting the obtained two-dimensional seismic amplitude image into a fault segmentation model, and inputting a fault binary image into the fault segmentation model;

optimizing the fault binary image to obtain a fault identification result;

the fault segmentation model is trained by fault training samples and a deep learning semantic segmentation model.

Further, the deep learning semantic segmentation model is a fault semantic segmentation deep learning model improved based on a VGG16 network, the network is a full convolution network, a convolution kernel in the front part of the whole network uses a size step of 3 x 3 to be 1, and a rear full connection layer is changed into a convolution layer with a convolution kernel size of 1 x 1 and a step of 1; the sizes of the filters of the pooling layers are all 2 multiplied by 2, and the step length of the Pool1 pooling layer is 2; step sizes of pool2 and pool3 pooling layers are 1, and activation functions are both ReLU functions; the 1 st layer and the 2 nd layer of the whole network adopt standard convolution, the 3 rd layer of the whole convolution layer adopts expansion convolution and mixed expansion convolution design, and the expansion rate of each layer is [1,3,5 ]; convolution kernels of

layers

4, 5 and 6 of the full connection layer use convolution with 1 × 1 step size and 1; four different expansion convolutions with expansion ratios of 6,12,18 and 24 were used at the FC4 layer, the FC5 and FC6 layers were connected, respectively, and then pixel-plus-pixel fusion was performed.

Further, the fault training sample is obtained by:

selecting a sample: firstly, extracting all two-dimensional slice results from processed three-dimensional seismic data according to Xline or Inline, selecting a proper amount of slice data, then interpreting and marking, and taking the rest other slices as data to be interpreted; all seismic amplitude images to be extracted are converted into single-channel gray images, and a plurality of blocks with proper sizes are randomly cut from each explained section and label thereof to serve as training samples of the model

And (3) expanding the sample: and carrying out mirror image transformation and rotation on the determined training sample image so as to increase the diversity of the sample.

Further, the deep learning network loss function of the fault segmentation model is as follows:

L_w＝-αylogy′-(1-α)(1-y)log(1-y′) (8)

wherein α is a ratio of non-fault in the entire data, and the weight of fault and non-fault in the loss function is controlled according to the ratio of fault and non-fault data;

y is the label of the true sample and y' is the predicted output via the softmax function.

Further, the optimizing the tomographic binary image includes:

and sequentially removing isolated small connected regions, skeletonizing, pruning optimization and disconnected line connection of non-connected faults from the fault binary image, and finally outputting an optimization result.

Further, the removing of the isolated small connected regions comprises:

firstly, acquiring contour information in a fault binary image; then, acquiring the area surrounded by the outline and setting an area threshold, and when the area is smaller than the threshold, considering the area as an isolated small area and needing to be removed; finally, removing the isolated small area.

Further, the skeletonization treatment comprises the following steps:

firstly, carrying out binarization processing on a fault identification framework result, and directly inputting the result which is identified by a deep learning semantic segmentation model into a fault and non-fault binary image;

secondly, performing fast Euclidean distance transformation calculation in the binary image so as to generate a distance transformation graph;

judging the distance value relationship between the current pixel and the eight neighborhood pixels thereof so as to generate a local central point;

fourthly, the coverage relation between the local central point and the boundary point of the object is considered in combination with the definition of the maximum circle so as to calculate the operation of generating a correlation matrix from the coverage relation between the local central point and the boundary point of the object;

fifthly, adopting a minimum coverage set method to convert the coverage test problem of the circle into row-column transformation of a correlation matrix;

removing the pseudo center point according to the result, and calculating the center point of the framework;

and connecting the central point of the skeleton and outputting the skeleton of the fault.

Further, the pruning optimization comprises:

for any point x on the skeleton, assuming that the importance measure s (x) is the pruning degree when the point is deleted, s (x) t (x), and t (x) is the moment when x is deleted, which is mainly a description of the pruning degree; for the pruning rate V (x), there are

While

Derived by two sides on t to obtain

Knowing that the pruning rate is in inverse proportion to the reciprocal of the importance measure; let R (x) be a function of the radius of the skeleton point, and a (x) be a function of the arc length of the skeleton branch at point x

On the basis of the above-mentioned information, a differential importance measure can be constructed, and the corrosion thickness differential importance measure is 1-R_aThe length-to-differential importance of the boundary and branch segments is measured as

Differential importance measure of corrosion area

Differential importance measure R (1-R) based on boundary smoothness_a)。

Further, the disconnected fault disconnection connection includes:

for the discontinuous part of the fault line after skeletonization, an eight-neighborhood end point detection method is adopted to search the end point of the discontinuous part, then the identified end point is judged whether to belong to a fault line according to the angle and the length of a connecting line, and the disconnected points of the same fault line are connected.

Furthermore, the invention also provides an electronic device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the seismic fault identification method based on deep learning semantic segmentation as described in any one of the above.

Compared with the prior art, the invention has the beneficial effects that:

(1) and (5) building a deep learning network. At present, most deep learning convolutional neural networks usually use information of an image block around a pixel to perform training and prediction processes by taking the image block and the pixel information as input and output of the convolutional neural network in order to classify a central point pixel of a seismic amplitude image. The method has the advantages of high storage overhead, low calculation efficiency, limitation of the size of the sensing area due to the size of the pixel block and increase of workload during prediction. The improved and simplified end-to-end deep learning semantic segmentation is adopted, image blocks with a certain size and all fault information of the image blocks are used as training samples, training time is short, good effects can be kept, and the seismic amplitude image with a certain size is used as input and the whole fault prediction result is output.

(2) And selecting fault training samples. In the training sample selection, a large number of manually marked fault samples are very difficult to implement and are relatively high in cost, so that a large number of synthetic fault data samples are generated to train a network, and although a large number of synthetic seismic data are easy to implement, the effect of predicting actual seismic data by using a synthetic seismic data training model is often unstable due to the difference of seismic data. In order to avoid the problem, the invention selects part of two-dimensional slice data in the three-dimensional slice data to be interpreted to manually interpret the slice as a training sample instead of generating a synthetic seismic record, randomly cuts a certain number of seismic amplitude images according to a certain size in the selected slice data and carries out slice marking, and also increases the diversity of learning samples by a data expansion method.

(3) And (5) performing optimization research on a loss function of the seismic fault recognition network model. In the case of the deep learning semantic segmentation method used for seismic fault recognition, since there are far more non-faults than faults in the actual seismic amplitude image, this causes a problem of sample imbalance. If the neural network is still trained using this loss function, the network easily converges to the wrong direction and zero-predicts everywhere, since zero-predicts is a good solution to this loss function in the fault segmentation problem. In order to solve the problem, the invention carries out weighting processing on the loss function, and the weights of the fault and the non-fault in the loss function are artificially controlled according to the ratio of fault data and non-fault data.

(4) And optimizing fault identification results. In most of the fault identification problems by using the convolutional neural network, the fault identification problem is often limited to only using a model for prediction to obtain a result. Although the accuracy is improved continuously with the increase of the number of iterations of the deep learning method, the fault continuity and the fault rationality are deficient. Therefore, the invention carries out the optimized post-processing on the images after the fault recognition of the model. Thinning the fault predicted by the learned model by a skeletonization method, removing redundant branches generated by skeletonization by adopting a pruning algorithm, removing noise in a prediction result by using an isolated small connected region removing method, and connecting discontinuous same faults by adopting an end point detection and connection method.

Drawings

FIG. 1 is a flowchart of a seismic fault identification method based on deep learning semantic segmentation according to an embodiment of the present invention;

FIG. 2 is a diagram of semantic segmentation models and their parameter settings constructed in accordance with the present invention;

FIG. 3 is a void space convolution pooling pyramid;

FIG. 4 is a sample generation process diagram;

FIG. 5 is a diagram of a sample expansion method;

FIG. 6 is a diagram of a fault identification result post-processing optimization process;

FIG. 7 is a partial enlarged view of a fault line break point connection process;

FIG. 8 is an angle view of two identified fault lines and their connecting lines;

FIG. 9 is a three-dimensional visualization of the resulting synthetic seismic data and its fault results;

FIG. 10 is a graph of the results of fault identification for synthetic seismic data, where (a) (d) (g) are the results of fault identification for three input seismic amplitude images (b) (e) (h) for their labels (c) (f) (i), respectively;

FIG. 11 is a graph of the accuracy and the average cross-over ratio of the model fault identification results trained with different training data volumes;

FIG. 12 is a comparison graph of model fault recognition results trained with different training data volumes;

FIG. 13 is a diagram of a full convolution structure for large dilated convolution rates;

FIG. 14 is a graph of two different network structure fault identification results;

FIG. 15 is a diagram of a post-processing optimization procedure for a noisy and discontinuous recognition result;

FIG. 16 is a diagram of a pruning optimization process;

FIG. 17 is a graph of the results of three-dimensional fault identification by Netherlands F3;

fig. 18 is a schematic composition diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

Example 1:

the fault identification is carried out quickly and accurately in the three-dimensional seismic data, human intervention and error are reduced in the identification process, and the time required for fault interpretation is shortened. The method is based on a deep learning semantic segmentation method, adopts the idea of pattern recognition to learn the pattern of explaining the fault by virtue of the strong fitting capability of a deep convolutional neural network, and utilizes computer vision to pick up the fault in the seismic amplitude image.

Referring to fig. 1, the seismic fault identification method based on deep learning semantic segmentation provided by this embodiment mainly includes the following steps:

101. acquiring three-dimensional seismic data, and performing slice extraction to obtain a two-dimensional seismic amplitude image;

102. inputting the obtained two-dimensional seismic amplitude image into a fault segmentation model, and inputting a fault binary image into the fault segmentation model;

103. and optimizing the fault binary image to obtain a fault identification result.

The fault segmentation model is obtained by training a fault training sample and a deep learning semantic segmentation model.

Specifically, the deep learning semantic segmentation model is a fault semantic segmentation deep learning model improved based on the VGG16 network, and the whole network architecture and parameter setting of the deep learning semantic segmentation model are shown in fig. 2. A plurality of improvements are mainly made:

1) full convolution model

The network structure is simplified by reducing the number of convolutional layers, only 5 convolutional layers are reserved finally, and the FC4, FC5 and FC6 of the network are converted into convolutional layers from fully-connected layers to change the network into a fully-convolutional network structure. Different from the classic CNN which uses the fully-connected layer to obtain the feature vector with the fixed length for classification after the convolutional layer, the FCN structure changes the fully-connected layer into the convolutional layer, and performs upsampling on the feature map of the last convolutional layer to restore the feature map to the same size as the input image, so that the pixel-by-pixel classification is performed on the upsampled feature map. By the method, each pixel point in the seismic amplitude image becomes a training sample, and the whole spatial information of the seismic amplitude image is correspondingly combined, so that the method can be suitable for the input seismic amplitude image with any scale.

2) Dilated convolution

The network introduces an expanding convolution operation in the three convolutional layers of Conv _3, which introduces a new parameter called "expansion rate" to the convolutional layers, which defines the spacing of the values at which the convolutional kernel processes the data, referred to herein as the number of intervals of the points of the convolutional kernel. The two-dimensional dilation convolution can be expressed as:

(F*_lk)(p)＝∑_s+lt＝pF(s)k(t) (1)

wherein_lFor hole convolution, p is its domain; f is the input image and S is its domain; compared with the ordinary convolution, the condition of the hole convolution is changed from s + t ═ p to s + lt ═ p, that is, each time the convolution kernel only operates with the element at the position of the multiple l in the image F.

3) Hybrid dilated convolution design

The hybrid dilation convolution design is adopted in the network design process. Using a set of different spreading convolution rates after downsampling can mitigate the grid effect while enlarging the field of view. When designing a hybrid deconvolution, consider an N number of K deconvolution, corresponding to a dilation rate of [ r [ r ] ]₁,…,r_i,…,r_n]The design goal is to let the final receiving field fully cover the whole area, and define the maximum distance between two non-zero points as:

M_i＝max[M_i+1-2r_i，M_i+1-2(M_i+1-2r_i)，r_i] (2)

wherein M is_n＝r_nThe design goal is to let M₂≤K。

The network structure can simultaneously meet the segmentation requirements of small objects and large objects. Therefore, in the network proposed in the present application, three dilation convolution dilation rates are set to [1,3,5] respectively in Conv _ 3.

4) Void space convolution pooling pyramid

As shown in fig. 3, the void space convolution pooling pyramid in the architecture designed in this embodiment uses 4 sets of dilation convolutions with dilation convolution rates of 6,12,18, and 24 on the FC4 layer, and then connects the FC5 and FC6 layers, respectively, and finally fuses the results obtained by the 4 sets of dilation convolution layers with different scales together by pixel addition. The different-scale expansion convolutional layer can be understood as a network which captures image context information in a plurality of proportions and finally performs fusion processing on the features.

That is, the network is first changed to a full convolutional network, the convolutional kernels in the front part of the whole network use 3 × 3 steps of size 1, and the following full connection layers become convolutional layers with convolutional kernel sizes of 1 × 1 steps of size 1. The pooling layer filters were all used in sizes of 2 x 2, with Pool1 pooling layer step sizes of 2. However, to enable the expanding convolution, the step size at pool2 and pool3 pooling layers is 1, and the activation functions are both ReLU functions. The 1 st and 2 nd layers of the whole network still adopt standard convolution, in order to obtain an image with a larger receptive field, the 3 rd layer of the whole convolution layer is designed by using expansion convolution and mixed expansion convolution, and the expansion rate of each layer is [1,3,5] respectively and is distributed in a sawtooth shape. In order to realize the full convolution network, convolution with the size step of 1 multiplied by 1 is used for convolution kernels of the 4 th layer, the 5 th layer and the 6 th layer of the original full connection layer. In order to realize multi-scale fusion to obtain more information, a hollow space pooling pyramid design is also used, four groups of different expansion convolutions with expansion rates of 6,12,18 and 24 are respectively used at an FC4 layer, and then the FC5 layer and the FC5 layer are respectively connected, and then pixel addition fusion is carried out. Finally, a prediction probability label graph is obtained by adding a softmax function, a good result can be obtained by directly using bilinear upsampling on the probability graph, and the segmentation of the fault is determined by a label class with the highest softmax probability at each pixel.

In addition, in the present application, because a fault identification network for each seismic data needs to be designed, a method of predicting actual seismic data by using synthetic seismic data training is abandoned, and a mode of interpreting and marking partial section faults in the actual seismic data to further generate samples as training data is adopted. Firstly, extracting all two-dimensional slice results from processed three-dimensional seismic data according to Xline or Inline, manually interpreting and marking after selecting a proper amount of slice data (the fault identification problem is a two-classification, the fault is 1, and the background of the non-fault is 0), and then taking the rest other slices as data to be interpreted. For computational convenience, one chooses to convert all of the seismic amplitude images extracted into a single-channel grayscale image (with values in the range of 0-255). And randomly cutting a certain number of picture blocks with proper sizes from each slice and the label thereof which are manually explained as training samples of the model, so that sample data can be properly increased. At the time of clipping, for the size H × W of the clipped sample and the original seismic image size H × W, the divisor of H and W, respectively, is designed. Fig. 4 shows the generation of training samples in a marked profile.

In addition, in order to increase the generalization capability of the model, the operation of sample expansion is required to further increase the pattern and the number of samples. In consideration of the actual condition of the stratum, only the mirror image transformation and rotation operations are carried out on the input sample image. As shown in fig. 4. For implementing the mirror image transformation process of the seismic amplitude image, only two transformation modes, namely vertical mirror image and horizontal mirror image, are performed in consideration of the actual situation. The horizontal mirror image is centered on the vertical centerline of the image, thereby swapping pixels of the image, i.e., swapping the left and right halves of the image. The vertical mirror image is formed by taking the horizontal center line of the image as an axis and exchanging the upper half part and the lower half part of the image, as shown in fig. 5.

For the mirror transform, assume that a point at a location within the seismic amplitude map is (x)₀，y₀) Then the value of (x, y) after horizontal mirroring is:

where W is the width of the seismic amplitude image, and the value of (x, y) after vertical mirror transformation is:

where H is the height of the seismic amplitude image.

For each pixel in the seismic amplitude input image x, the corresponding trained model output represents the estimated posterior probability that the pixel belongs to a fault. In the multi-classification task, the softmax cross entropy loss function as formula (5) is generally adopted for learning. The essence of the present model is to perform a pixel two classification, so the formula can become (6). Therefore, we use the binary cross entropy as shown in equation (5) to be used as the loss function, in the following equation, where T is the number of classes, y is the label of the real sample (1 plus 0 minus), and y' is the prediction output (value between 0 and 1) through the softmax function, as shown in equation (7).

In general the distribution of samples is more or less balanced and the above mentioned loss function can be used. However, seismic amplitude image fault identification is essentially a binary problem, and after interpreting the seismic data and then marking samples, the quantity of faults and non-faults can be found to be extremely unbalanced, the quantity of non-fault samples is far greater than that of fault samples, and the quantity of fault samples and non-fault samples is extremely unbalanced. If the neural network is still trained using the conventional cross-entropy loss function, the network easily converges to the wrong direction and zero-predicts everywhere, since zero-predicts is a good solution to this loss function in the fault segmentation problem. In order to solve the problem of sample imbalance, weighting processing may be performed on the loss function, and a weighted cross-entropy loss function is used, which may be expressed as:

L_w＝-αylogy′-(1-α)(1-y)log(1-y′) (8)

where α is the ratio of non-fault to total data, the weights in the loss function for fault and non-fault are artificially controlled based on the ratio of fault and non-fault data.

In the fault recognition result by using the semantic segmentation method, some false faults still exist, and the false faults are distributed on the predicted binary image by small-part non-connected noise. Meanwhile, because the deep learning method is based on data driving, some attributes representing faults are not completely learned, disconnected fault frameworks may exist, and the unconnected parts need to be connected. In addition to the skeletonization processed by the optimization process, the method of removing isolated small connected regions is also adopted to remove isolated false faults, and after the skeletonization, the method of connecting discontinuous fault lines based on an eight-neighborhood endpoint detection method and the branch pruning based on the importance measure are adopted to optimize the fault lines after the skeletonization. The whole optimization process is shown in fig. 6, and includes:

1) removal of isolated small connected regions

The method for removing the small connected region is mainly used for determining whether the small connected region is removed or not based on the area of the connected region pair, and the whole removing process comprises the following steps: firstly, acquiring contour information in a binary image; then, acquiring the area surrounded by the outline and setting an area threshold, and when the area is smaller than the threshold, considering the area as an isolated small area and needing to be removed; and finally, removing the isolated small regions, wherein the removing operation is simple, and the fault identification result is a binary image, so that only 0 needs to be filled in the regions. Since the fault line obtained after skeletonization is one pixel, the process of removing the small connected region is usually performed before skeletonization.

2) Skeletonization

The fault identified by the deep learning semantic segmentation method is a strip-shaped result, and the actual fault interpretation result is a fault line, so that the fault line capable of accurately determining the fault position needs to be extracted, and the fault line is analyzed and compared by combining various skeleton extraction methods, so that a skeleton of the fault line is calculated by adopting a distance transformation-based skeleton extraction algorithm. The specific steps of the algorithm are as follows:

firstly, carrying out binarization processing on a fault identification framework result, and identifying that the result is a fault and non-fault binary image by using a deep learning semantic segmentation model, so that the result can be directly input;

3) Broken wire connection

For the discontinuous part of the fault line after skeletonization, an eight-neighborhood end point detection method is adopted to search the end point of the discontinuous part, then the identified end point is judged whether to belong to a closed fault line or not according to the angle and the length of a connecting line, and the disconnected points of the same fault line are connected, so that the disconnection connection can be realized. In the fault line disconnection point detection, the skeletonized result is a binary image fault line of one pixel, and in the fault line of the fault discontinuity, the point of the disconnection position is expressed as a boundary point and the number of connections of the pixel is 1. It can be understood that it is necessary to detect the end points in the fault line binary image. For the end points in the binary image of the fault line, because of skeletonization, only one pixel with the gray level equal to 1 is arranged around the end points, so that the end points can be detected according to the characteristics that the sum of eight neighborhood pixels is 1, and the process of end point detection and connection is shown in fig. 7. When all the end points are detected, the angle between the two fault lines to be connected and the connecting line and the length of the connecting line are judgedWhether the two broken fault lines belong to a through fault. As shown in fig. 8, for two fault lines L₁And L₂The detected endpoints are respectively P₁And P₂,L₁And L₂And a connecting line P₁P₂Are respectively theta₁And theta₂If theta is greater than theta₁And theta₂And P₁P₂The lengths are respectively within the set threshold value range, and then belong to the same fault. Thus, the connected fault line can be obtained by performing the breakpoint connection.

4) Pruning

At present, various skeleton pruning methods have been proposed, and various branch importance measures have been proposed by scholars. An efficient pruning method should evaluate the importance of the branches to decide whether to prune or preserve the fragment. The pruning algorithm based on the importance measure is to optimize the existing skeleton branches by adopting a proper standard after the skeleton lines are obtained. Considering these conditions, Shaked et al summarize the importance measure method, proposing a normalized method rate pruning framework for pruning skeletons in continuous space.

Mathematically, it can be described that for any point x on the skeleton, it is assumed that the importance measure s (x) is the pruning degree when the point is deleted, s (x) t (x), and t (x) is the moment when x is deleted, which is mainly a depiction of the pruning degree. For the pruning rate V (x), there are

While

Derived by two sides on t to obtain

It is known that the pruning rate is inversely related to the inverse of the measure of importance. Let R (x) be a function of the radius of the skeleton point, and a (x) be a function of the arc length of the skeleton branch at point x

Differential importance measure of corrosion area

Differential importance measure R (1-R) based on boundary smoothness_a)。

After pruning, the resulting skeleton is the main form that may not completely reconstruct the original object, but is capable of reconstructing the object. In the case of the fault identification of the present application, the skeleton means that the target object does not have to be restored from the shape parameters as long as the skeleton contains information sufficient to distinguish the fault lines that may be encountered. Therefore, the requirement of optimizing the fault framework can be met, and in the pruning, the fault line branch removal after the framework is carried out according to the lengths of the boundary segment and the branch segment.

Therefore, the depth learning method is applied to seismic data interpretation in seismic exploration, the depth learning semantic segmentation method is used for identifying the fault in the seismic amplitude image, and the fault in the seismic amplitude image is interpreted from the perspective of a computer. On the basis of the existing mature deep learning semantic segmentation network, an improved VGG16 deep learning semantic segmentation network is provided for seismic amplitude image fault identification according to the characteristics of seismic fault interpretation, and a post-processing optimization method is researched according to the defects of fault identification of a deep learning semantic segmentation model, so that the fault interpretation rationality and effect are further improved. Has the following advantages, characteristics and positive effects:

(1) the network structure has fewer parameters and is more simplified

For the network structure, parameters and FLOPs (floating point operands, which can be understood as computational quantities) are calculated respectively to measure the complexity of the model, and are compared with other parameters and FLOPs commonly used at present, as shown in table 1. It can be seen that the parameter quantity is greatly reduced and the model complexity is not very high after simplification and improvement. The parameter quantity and complexity of the model are optimized, so that the training time is reduced, and the training efficiency is improved.

TABLE 1 common models and parameters and FLOPs for the network model of the present application

(2) The constructed network structure can realize end-to-end fault identification

The invention constructs an end-to-end deep learning semantic segmentation network, and the whole seismic amplitude image corresponds to the whole fault recognition result. In order to test the end-to-end fault identification mode of the proposed network, the method flow of creating the three-dimensional synthetic training data set proposed by the five-new invention is adopted to generate synthetic seismic fault data and tags thereof, wherein the size of the synthetic seismic fault data is 128 x 128, and the data is shown in fig. 9.

The experimental test is to select 40 slices randomly selected from the Inine direction directly as sample data, visually display the sample data by using a gray scale image, and obtain the size of the image as 128 x 128 by using the number of pixels as a unit, wherein the size is already small, and all operations of cutting are not performed, wherein 30 slices are used as a training set, and 10 slices are used as a verification set. In the model training, the learning rate is set to be 0.001, 10 slices are input for training each time, and 351 times of iterative training are performed in total to obtain the optimal result. After the model is trained, the slice of the whole INine can be predicted, and for the training of the whole synthetic seismic data, the model training needs about one hour, and only about 0.1 second is needed for predicting one piece. The end-to-end semantic segmentation model also ensures that a binary image of a corresponding fault identification result can be obtained only by directly inputting the seismic amplitude image. The recognition results of three slices are selected from the prediction results, the three slices are not involved in the training process, the prediction results are shown in fig. 10, and it can be seen that the positions of the faults and the shapes and the distributions of the faults almost completely conform to the corresponding labels, so that the excellent performance of the semantic segmentation neural network model provided by the invention in the aspect of fault recognition accuracy is verified.

(3) The network still has good fault identification effect in a small number of fault marking sections

The designed network not only can realize end-to-end fault identification, but also has good fault identification effect in a small quantity of fault marking sections. The invention tests the recognition effect of the proposed network structure in the training results under different slice data volumes to verify whether the network can obtain good time effect under less data volume. In the same test data, four groups of data of 5%, 10%, 15% and 20% of the whole number of the Inine slices are randomly selected as training samples, and in order to control variables and facilitate comparison results, the number of the slices added in each group is randomly selected on the basis of the original number of the slices. The four groups of data are respectively input into the network structure provided by the application for training, after the optimal model is trained, the whole slice is predicted, and the accuracy and the average cross-over ratio of the recognition results of the four models are respectively calculated, as shown in fig. 11. In addition, one seismic amplitude image is selected and input into four trained models for prediction, and the result is shown in fig. 12.

Through comparison of the accuracy rate and the average cross ratio of the four groups of data identification results, it can be found that the accuracy rate and the average cross ratio are improved along with the increase of the data volume, the variation difference of the accuracy rate is not very large, the average cross ratio is improved by about 3%, and the actual rule of deep learning is also met, because the more samples are, the better the result of the overall learning is. In the identified fault image, the rule is satisfied, four groups of data can be detected for main faults in the seismic amplitude image, and the overall difference is not very large and is only slightly different in details. The main fault can be identified for 5% of the data volume, which also shows that the invention designs the network to be able to satisfy the good results in fewer samples.

(4) The structural design of the network can obtain better recognition effect

The network structure designed by the invention has a good recognition result relative to other network structures. Different network design methods are used for comparison with the network proposed by the present invention. On the basis of the network provided by the application, the conventional convolution is used for the whole convolution layer of the layer 3, the expansion convolution and mixed expansion convolution design is removed, the network of the cavity space pooling pyramid for multi-scale resampling is removed, the expansion convolution (the expansion convolution rate is set to be 12) capable of increasing the large receptive field is used for the FC4 layer of the network to replace the original cavity space pooling pyramid, the last three-layer structure parameters of the model are shown in figure 11, and other parameters and structures are still unchanged.

In addition to the calculation of the common evaluation criteria, two fault recognition results are selected from the prediction results of the two methods for comparison, as shown in fig. 13. Compared with the identification result of a network with a large expansion convolution rate, the fault identification method has better continuity, generates less noise while preserving the spatial form and continuity of the fault identification method, and has the defects of poor continuity and obvious grid effect in the identification result of the network with the large expansion convolution rate.

(5) The optimized post-processing can further improve the recognition effect

Through the optimized post-processing flow provided by the invention, the fault identification effect can be further improved. In order to verify the optimized post-processing method, a semantic segmentation model trained by synthetic seismic data is used for identifying fault results, and post-processing optimization test is carried out. In the skeletonization process, if the identified fault line is relatively simple, branching may not occur, and it is not necessary to perform pruning. FIG. 14 is a post-processing optimization procedure for a noisy and discontinuous recognition result.

And optimizing the fault recognition result with branches after skeletonization by using a pruning method, wherein the pruning optimization process is shown in the following figure 15.

The defects of the fault interpretation result obtained by the semantic segmentation method can be seen, and the optimization post-processing method provided by the invention can make up for the defects. The identified false fault can be removed, meanwhile, the obvious non-connected fault line can be connected, and the branches caused in the skeletonization process can also be removed by a pruning method, so that an unreasonable explanation part is eliminated, and the fault identification result is more accurate and reasonable.

The invention is further described below with reference to a specific application scenario example:

in the specific implementation process, open source seismic data on Github is selected for testing. The data is Netherlands F3 seismic data, which was collected in the north sea offshore in the Netherlands. The data dimension size is n 1-100, n 2-400, and n 3-420. In fault recognition of the Netherlands F3 seismic data, the seismic data are decomposed into 420 seismic amplitude image gray-scale maps with the size of 100 x 400, and 20 section maps and an interpretation result binary map obtained by thinning the section maps through an optimal surface voting method are randomly selected as labels. 480 blocks with the size of 100 x 100 are finally generated as training samples in 20 cross-sectional views through random clipping and sample expansion, and 470 blocks are selected as a training set and 10 blocks are selected as a verification set. In the training process, 10 blocks are selected as a batch, and the learning rate is set to be 0.001. After 16 training sessions are performed on the whole training data, fault contours can be basically obtained, although continuous iteration can improve the precision, continuity is sacrificed, overfitting possibly influences the generalization capability of the model, and therefore, the training result is optimized by post-processing instead of continuous iteration. The method can be used for fault prediction after the model is trained, in the training process, image blocks with specific sizes are input, therefore, the obtained network can only carry out pixel classification prediction on images with the same size, and the size of the image blocks to be cut is set to be the divisor of the size of the slice data, so that the prediction is convenient. The size of a training image block of the Netherlands F3 seismic data can evenly divide an original slice image into four parts, so that the four image blocks of the seismic amplitude image to be predicted decomposition are directly input into a trained model to obtain a fault classification result, and then the four parts are combined to obtain a fault identification result of the whole slice. And after the prediction is finished, optimizing post-processing, performing skeletonization, removing isolated small connected regions, connecting broken lines and pruning, and combining slices to realize the identification result of the three-dimensional fault. The Netherlands F3 seismic data fault identification results are shown in fig. 16.

To sum up, the present invention mainly solves the following technical problems in the prior art:

Example 2:

referring to fig. 18, the electronic device provided in this embodiment includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an unmanned aerial vehicle aerial image defogging processing program based on image matching. The processor, when executing the computer program, implements the steps of embodiment 1 described above, such as the steps shown in fig. 1.

Illustratively, the computer program may be partitioned into one or more modules/units that are stored in the memory and executed by the processor to implement the invention. The one or more modules/units may be a series of computer program instruction segments capable of performing specific functions, which are used to describe the execution process of the computer program in the electronic device.

The electronic device may be a desktop computer, a notebook, a palm computer, a cloud server, or other computing device. The electronic device may include, but is not limited to, a processor, a memory. Those skilled in the art will appreciate that fig. 18 is merely an example of an electronic device and does not constitute a limitation of an electronic device, and may include more or fewer components than shown, or combine certain components, or different components, for example, the electronic device may also include input output devices, network access devices, buses, etc.

The Processor may be a Central Processing Unit (CPU), other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf programmable Gate Array (FPGA) or other programmable logic device, discrete Gate or transistor logic, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory may be an internal storage element of the electronic device, such as a hard disk or a memory of the electronic device. The memory may also be an external storage device of the electronic device, such as a plug-in hard disk, a Smart Memory Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like provided on the electronic device. Further, the memory may also include both an internal storage unit and an external storage device of the electronic device. The memory is used for storing the computer program and other programs and data required by the electronic device. The memory may also be used to temporarily store data that has been output or is to be output.

The above embodiments are only for illustrating the technical concept and features of the present invention, and the purpose thereof is to enable those skilled in the art to understand the contents of the present invention and implement the present invention accordingly, and not to limit the protection scope of the present invention accordingly. All equivalent changes or modifications made in accordance with the spirit of the present disclosure are intended to be covered by the scope of the present disclosure.

Claims

1. A seismic fault identification method based on deep learning semantic segmentation is characterized by comprising the following steps:

optimizing the fault binary image to obtain a fault identification result;

2. The seismic fault identification method based on deep learning semantic segmentation as claimed in claim 1, wherein the deep learning semantic segmentation model is a fault semantic segmentation deep learning model improved based on a VGG16 network, the network is a full convolution network, the front part of convolution kernel in the whole network uses a size step of 3 x 3 to be 1, and the rear full connection layer becomes a convolution layer with a convolution kernel size of 1 x 1 and a step of 1; the sizes of the filters of the pooling layers are all 2 multiplied by 2, and the step length of the Pool1 pooling layer is 2; step sizes of pool2 and pool3 pooling layers are 1, and activation functions are both ReLU functions; the 1 st layer and the 2 nd layer of the whole network adopt standard convolution, the 3 rd layer of the whole convolution layer adopts expansion convolution and mixed expansion convolution design, and the expansion rate of each layer is [1,3,5 ]; convolution kernels of layers 4, 5 and 6 of the full connection layer use convolution with 1 × 1 step size and 1; four different expansion convolutions with expansion ratios of 6,12,18 and 24 were used at the FC4 layer, the FC5 and FC6 layers were connected, respectively, and then pixel-plus-pixel fusion was performed.

3. The seismic fault recognition method based on deep learning semantic segmentation as claimed in claim 1, wherein the fault training samples are obtained by:

selecting a sample: firstly, extracting all two-dimensional slice results from processed three-dimensional seismic data according to Xline or Inline, selecting a proper amount of slice data, then interpreting and marking, and taking the rest other slices as data to be interpreted; all the extracted seismic amplitude images are converted into single-channel gray images, and a plurality of blocks with proper sizes are randomly cut from each explained slice and labels thereof to serve as training samples of the model;

4. The seismic fault identification method based on deep learning semantic segmentation of claim 1, wherein the deep learning network loss function of the fault segmentation model is as follows:

L_w＝-αylogy'-(1-α)(1-y)log(1-y') (8)

5. The seismic fault identification method based on deep learning semantic segmentation as claimed in claim 1, wherein the optimizing the fault binary image comprises:

6. The depth-learning semantic segmentation-based seismic fault identification method of claim 5, wherein the removal of the isolated small connected regions comprises:

7. The method of deep learning semantic segmentation-based seismic fault recognition of claim 5, wherein the skeletonization process comprises the steps of:

8. The method of deep learning semantic segmentation-based seismic fault identification according to claim 7, wherein the pruning optimization comprises:

for any point x on the skeleton, assume its importance measure S (x) is to delete the point xThe pruning degree at the time of point, s (x) ═ t (x), and t (x) is the moment when x is deleted, and is mainly a depiction of the pruning degree; for the pruning rate V (x), there are

While

Derived by two sides on t to obtain

Differential importance measure of corrosion area

Differential importance measure R (1-R) based on boundary smoothness_a)。

9. The depth-learning semantic segmentation-based seismic fault identification method of claim 8, wherein the unconnected fault line breaks connection comprises:

for the discontinuous part of the fault line after skeletonization, an eight-neighborhood end point detection method is adopted to search the end point of the discontinuous part, then the identified end point is used for judging whether the fault line belongs to a closed fault line or not according to the angle and the length of a connecting line, and the disconnected points of the same fault line are connected.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the method for seismic fault recognition based on deep learning semantic segmentation of any one of claims 1 to 9.