Image dataset annotation and segmentation method based on deep learning for fault prediction

Abstract

The present invention discloses a method for labeling and segmenting image data sets based on deep learning for fault prediction. This method is a coal-rock fault prediction method. The method comprises the following steps: obtaining original three-dimensional seismic data within a work area to be measured, performing dimensionality reduction processing on the original three-dimensional seismic data, and cutting the original two-dimensional matrix along the inline direction. Through data dimensionality reduction, the three-dimensional seismic data is compressed into two-dimensional planar data. Then, by cutting along the inline direction, the three-dimensional seismic data is materialized into two-dimensional cross-sectional image data of Time and CDP in each inline direction. This method skips the traditional three-dimensional seismic data processing method of remaining focused on data calculation, directly materializes image features, and retains image features at the edge. This method reduces the amount of computation and computing time without losing edge information, thereby improving the overall network operating efficiency.

Description

Technical Field

[0001] The present invention relates to a coal rock fault prediction method, and more specifically, to an image data set annotation and segmentation method based on deep learning for fault prediction. Background Art

[0002] When rock is subjected to stress that exceeds its ultimate strength, it fractures, forming a fault structure. Faults are one of the most common geological phenomena. During coal mining, faults not only alter the burial conditions of coal seams, causing them to shift and displace significantly. This, on the one hand, disrupts the continuity and integrity of the coal seams, creating obstacles to mining; on the other hand, faults are prone to accidents such as gas outbursts, water seepage, and roof collapse, seriously impacting safe mining in the mining area. Therefore, predicting the distribution of faults in mining areas is crucial for safe coal mining. Currently, coalfield geological structures are typically detected using artificial seismic methods. Conventional fault interpretation methods determine the presence and magnitude of faults by visually identifying reflected wave event shifts and anomalies. Traditional fault interpretation methods are time-consuming, highly subjective, and difficult to interpret for small faults. Furthermore, they rely heavily on the interpreter's experience and prior geological knowledge, making it difficult to meet the high-precision, efficient, and fast fault information requirements for safe coal mining. The development and rise of deep learning research, particularly neural networks, has opened up the possibility of automated fault identification. Traditional machine learning detection methods extract features from earthquake attributes and combine them with classification algorithms to learn from seismic data. However, fault profiles often suffer from low contrast, noise interference, data imbalance, a small number of fault samples, and varying fault heights, making traditional machine learning methods challenging to detect. In recent years, deep learning models based on convolutional neural networks have played an important role in fault prediction. By autonomously learning fault features and simultaneously learning non-fault features around the fault points, the model is continuously refined to enable fault identification across the entire region. Therefore, research on methods for constructing and segmenting fault prediction datasets based on deep learning is of great significance. Summary of the Invention

[0003] One of the purposes of the present invention is to address the above-mentioned shortcomings and provide an image dataset annotation and segmentation method based on deep learning for fault prediction, in the hope of solving the problems in the existing technology that the traditional fault interpretation method is difficult to interpret small faults and relies on the experience of the interpreter and prior information of relevant geological knowledge; the fault identification method of deep learning of neural networks also has technical problems such as low contrast, noise interference, small number of layer samples, and different fault drops.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] The present invention provides an image dataset annotation and segmentation method based on deep learning for fault prediction, which includes the following steps:

[0006] Step A: Acquire original 3D seismic data in the work area to be measured. The original 3D seismic data records the amplitude information of each point in the work area to be measured in three dimensions: Inline, CDP, and Time. The data format is .segy.

[0007] Step B: Perform dimensionality reduction processing on the original three-dimensional seismic data to obtain the original two-dimensional matrix M by the following formula: tow-dim .

[0008] M tow-dim =Time×(Inline×CDP)

[0009] Step C: Inline the original two-dimensional matrix M tow-dim Perform cutting to obtain multiple cutting two-dimensional matrices M Inline , cutting the two-dimensional matrix M Inline The number is the number of Inlines.

[0010] Step D: for each of the cutting two-dimensional matrix M Inline Cut again to generate multiple two-dimensional small windows M of the current window patch Each of the two-dimensional small window M patch Time, Inline, and CDP information are recorded.

[0011] Step E: Through each two-dimensional small window M patch Recorded Inline, CDP, Time information, search for the corresponding Inline, CDP, Time information in the disclosed information of the work area to be tested, and determine the current two-dimensional small window M patch Is it the required data? If the judgment result is yes, then the current two-dimensional small window M patch Add the counter parameter as an index and put it into a new three-dimensional matrix; otherwise, continue to judge the current cutting two-dimensional matrix M Inline The next two-dimensional small window M in patch .

[0012] Step F: Based on the counter parameter, the fault and non-fault label matrices are simultaneously established, and the data set matrix and the label matrix are linked based on the counter parameter to complete the current cutting two-dimensional matrix M. Inline Repeat steps DF to obtain a three-dimensional fault data set, fault labels, a three-dimensional non-fault data set, and non-fault labels based on the exposed information of the work area to be measured.

[0013] Preferably, a further technical solution is: the original three-dimensional seismic data in the work area to be measured includes disclosed information of the work area to be measured, and the disclosed information of the work area to be measured includes disclosed fault information and disclosed non-fault information.

[0014] A further technical solution is: for each cutting two-dimensional matrix M Inline The cutting method is: use an n×n window and cut the two-dimensional matrix M Inline Move upward, determine the center point of the window, extend left by n / 2-1 points, right by n / 2 points, upward by n / 2-1 points, and downward by n / 2 points.

[0015] A further technical solution is that the original three-dimensional seismic data is obtained from Geoframe software.

[0016] A further technical solution is: the new three-dimensional matrix is ​​a small two-dimensional window M patch After adding the counter parameter as an index, the counter×n×n three-dimensional data matrix is ​​rebuilt and the two-dimensional matrix M is cut successively. Inlin After averaging, the data are automatically incremented according to the counter parameter and sequentially placed into the seismic data volume of the three-dimensional data matrix; the newly created three-dimensional matrix includes a fault three-dimensional matrix and a non-fault three-dimensional matrix.

[0017] A further technical solution is: the method also includes: dividing the three-dimensional fault data set and the three-dimensional non-fault data set into a training set and a test set based on the counter parameter, and at the same time forming training labels and test labels, and putting the training set, test set, training labels and test labels into a convolutional neural network for deep learning of fault prediction.

[0018] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0019] Through data dimensionality reduction, 3D seismic data is compressed into 2D planar data. Then, by slicing along the inline direction, the 3D seismic data is visualized as 2D cross-sectional image data showing time and CDP along each inline direction. This bypasses the traditional data calculations required for 3D seismic data processing and directly visualizes image features.

[0020] By using an n×n patch image segmentation method, each captured patch is indexed by its center point, establishing a connection between image features and data, thereby ensuring that image feature capture does not lose the actual data as its foundation. Furthermore, this segmentation method can better simulate the function of human vision to capture information from two-dimensional images, allowing computers to obtain more information about the differences between fault points and non-fault points when building recognition models. It can also quickly capture the inline, CDP, and time information of the center point. While capturing amplitude information, using inline, CDP, and time as index information can quickly establish a connection between the two-dimensional image data and the revealed fault point information, thus saving time and effort in creating dataset labels.

[0021] Through the patch image segmentation method of the center point, the image features of the edge part are retained, while the edge information is not lost, the amount of calculation and the calculation time are reduced, thereby improving the operating efficiency of the overall network.

[0022] By re-entering the truncated 2D matrix into a 3D matrix containing the counter parameter, the label information and the 2D matrix can be easily indexed. In the established convolutional neural network, it is also directly used as a 3D matrix input. The 3D matrix input can process each dimension of the 2D image simultaneously, thereby enabling parallel computing within the network, reducing computation time and improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart illustrating a method according to an embodiment of the present invention.

[0024] Figure 2 1 is used to illustrate an original two-dimensional matrix after dimensionality reduction processing in one embodiment of the present invention.

[0025] Figure 3 It is used to illustrate a two-dimensional cutting matrix after cutting along the Inline in one embodiment of the present invention.

[0026] Figure 4 25 random image patches are used to illustrate a single inline in one embodiment of the present invention.

[0027] Figure 5 Partial cross-sectional images of a single inline fault and non-fault data set are used to illustrate one embodiment of the present invention. DETAILED DESCRIPTION

[0028] As mentioned above, to meet the training model's requirements for the dataset, the image dataset annotation and segmentation method based on deep learning for fault prediction of the present invention is implemented as follows:

[0029] (1) Obtain the 3D seismic data of the work area to be measured. The data format is to store the amplitude information of each point in the form of Time, Inline, and CDP in the segy file S origin .

[0030] (2) Obtain the fault information of some exposed parts of the work area to be tested. The data format is a label file L containing Time, Inline, CDP, etc. Fault .

[0031] (3) Obtain the non-fault information of the part of the work area to be tested that has been exposed. The data format is a label file L containing Time, Inline, CDP, etc. non-Fault .

[0032] (4) Perform dimensionality reduction processing on the original 3D seismic data to obtain M tow-dim =Time×(Inline×CDP) two-dimensional matrix M tow-dim .

[0033] (5) For the two-dimensional matrix M tow-dim By cutting in the inline direction, we can get I (the number of inline numbers) Time × CDP two-dimensional matrices M Inline .

[0034] (6) On each cut Time×CDP two-dimensional matrix, a window (patch) of size n×n is moved on the data points of this two-dimensional matrix, and a small n×n two-dimensional window M of the patch is generated according to the center point of the patch (extending 7 points to the left, 8 points to the right, 7 points above, and 8 points below the patch). patch , the M patch The stored information is the amplitude information of n×n points centered at the point, and the Time, Inline, and CDP information are used as the matrix M patch Index label L patch .

[0035] (7) Given that in (5) Inline The Time, Inline, and CDP information of each point in the M Inline M intercepted by traversal patch The label L obtained in patch The index information is in L Fault and L non-Fault Traverse search.

[0036] (8) The search logic is label L patch The index information must be consistent with L Fault or L non-FaultThe information in L is exactly the same, that is, Time, Inline, and CDP are all the same. Fault or L non-Fault Find the same information in the matrix M patch Add another label L classification , store the M patch The center point of the matrix is ​​the fault point or the non-fault point, and the two-dimensional matrix M patch Stored in two fault and non-fault three-dimensional matrices M three-dim (counter×n×n), L classification Corresponding to M three-dim The counter in is recorded in vector V lable and proceed to the next one; if label L patch The index information is in L Fault or L non-Fault If no point is found, then the point is discarded and the next one is moved to the next one. Inline All M patch All intercepted and searched completed.

[0037] (9) After the inner loop of (8) is completed, M Inline The next Inline section is intercepted to form an outer loop. In this way, the inner and outer loops are nested to capture all the breakpoint information and label information. Finally, the fault three-dimensional matrix M is generated based on the exposed fault information. three-dim-fault (counter×n×n) and its label information V lable-fault (fault is marked as 1, non-fault is marked as 0); and the non-fault three-dimensional matrix M three-dim-nonfault (counter×n×n) and its label information V lable-nonfault .

[0038] (10) Furthermore, in the three-dimensional matrix M three-dim-nonfault and M three-dim-fault The training set and test set are divided into 7:3 ratios. Each two-dimensional image in this three-dimensional matrix is ​​controlled by a unique third-dimensional counter, so the counter can be used as an index to divide the label matrix into training set and test set. Thus, the training data set D is formed. train And the test dataset D test , and the accompanying training label vector L train and the test label vector L test .

[0039] (11) The initial work of making the data set has been completed. Now we can put the cut data set into the convolutional neural network to detect and identify the fault images, establish a recognition model, and then predict the fault distribution of the entire area, so as to achieve the purpose of fault prediction.

[0040] The present invention will be further described below with reference to the accompanying drawings.

[0041] refer to Figure 1 As shown, one embodiment of the present invention is the above-mentioned image dataset annotation and segmentation method based on deep learning for fault prediction, which is performed according to the following steps:

[0042] S1. First, we need to obtain raw 3D seismic data within the work area. This data, in the .segy format, records the amplitude information for each point within the work area, including the inline, CDP, and time domains. As 3D seismic data, it is difficult to directly process and segment it. Therefore, we need to reduce its dimensionality and convert it into a 2D image while preserving its 3D seismic characteristics.

[0043] S2, such as Figure 2 As shown in the figure, after dimensionality reduction, the Inline and CDP in the 3D seismic data are integrated into the same dimension, but the stratigraphic features and fault locations are still unclear.

[0044] S3. To obtain stratigraphic characteristics and determine fault locations, we need to slice the seismic data along inlines and generate a single-inline 2D seismic profile image on a particular inline. This 2D seismic profile image has time as the vertical axis and CDP as the horizontal axis.

[0045] like Figure 3 As shown, on this image, we can already perform traditional manual identification work, that is, we can judge the fault location based on experience and knowledge of interpretation work.

[0046] S4. In order to better capture the features and details near the breakpoint and put them into the neural network for training, it is necessary to segment the single inline 2D seismic profile image.

[0047] Preferably, the above-mentioned method for segmenting a single inline 2D seismic profile image is to use an n×n patch and move it across the target 2D image to be segmented. The logic for selecting points in the patch is to use a point as the center point, extend n / 2-1 points to the left, n / 2 points to the right, n / 2-1 points upward, and n / 2 points downward (n is usually an even number).

[0048] To prevent unnecessary features from interfering with the model's judgment, edge information points will not be used as center points. Instead, edge features will be captured together in the patch of their nearest neighbors.

[0049] In addition to the amplitude information of the n×n points, the intercepted patch also records the Inline, CDP, and Time information of the center point, such as Figure 4 Shown are 25 random patches of a single inline.

[0050] After a single inline two-dimensional cross-sectional image is segmented by patches, it is aggregated into a four-dimensional matrix. Since the inline is uniquely determined, the first two dimensions of this four-dimensional matrix are CDP and Time, respectively, and the last two dimensions store the patch data for easy access.

[0051] Based on the exposed work area information, the corresponding inline, CDP, and time information for each small 2D data matrix is ​​searched for the corresponding inline, CDP, and time information in the slice and non-slice data to determine whether the 2D data matrix is ​​the desired data. If it is slice or non-slice data, a counter parameter is added to the small 2D data matrix as an index and placed into a new 3D matrix.

[0052] According to the index parameter counter, the fault and non-fault label matrices are established at the same time, and the data set matrix and label are connected based on the counter. In this way, the segmentation of a single inline 2D seismic profile image and the production of data sets and labels are completed. Figure 5 Partial cross-sectional images of single inline fault and non-fault datasets.

[0053] S5. After completing the segmentation of a two-dimensional seismic profile image in the Inline direction, continue to complete the next two-dimensional seismic profile image in the Inline direction using the same cutting and searching method.

[0054] S6. During the process of segmenting and finding all datasets and labels, you can use a loop or parallel approach to cut. The loop inline approach sequentially cuts all inline 2D seismic profile images. The parallel approach is relatively fast but requires higher computer computing power.

[0055] S7. After cutting all the two-dimensional cross-sectional images, a three-dimensional fault data set, fault labels, a three-dimensional non-fault data set, and non-fault labels obtained based on the exposed information of the work area can be obtained.

[0056] Based on the above embodiments, it can be seen that the present invention has the following improvements over the prior art:

[0057] Through data dimensionality reduction, 3D seismic data is compressed into 2D planar data. Then, by slicing along the inline direction, the 3D seismic data is visualized as 2D cross-sectional image data showing time and CDP along each inline direction. This bypasses the traditional data calculations required for 3D seismic data processing and directly visualizes image features.

[0058] By using an n×n patch image segmentation method, each captured patch is indexed by its center point, establishing a correlation between image features and data. This allows image feature capture without losing the underlying data. Furthermore, this segmentation method better simulates human vision to capture information from two-dimensional images, allowing computers to better understand the differences between fault points and non-fault points when building recognition models.

[0059] Through the n×n patch image segmentation method, the inline, CDP, and time information of the center point can be quickly captured. While intercepting the amplitude information, using inline, CDP, and time as index information, the two-dimensional image data and the revealed breakpoint information can be quickly linked, thereby saving time and effort in producing dataset labels.

[0060] Through the patch image segmentation method of the center point, the image features of the edge part are retained, while the edge information is not lost, the amount of calculation and the calculation time are reduced, thereby improving the operating efficiency of the overall network.

[0061] By re-entering the truncated 2D matrix into a 3D matrix containing counters, we can easily index the label information and the 2D matrix. Furthermore, the 3D matrix is ​​directly used as input to the established convolutional neural network. The 3D matrix input allows for simultaneous processing of each dimension of the 2D image, enabling parallel computation within the network, reducing computation time and significantly improving efficiency.

[0062] In addition to the above, it should be noted that references to "one embodiment," "another embodiment," "an embodiment," and the like in this specification refer to specific features, structures, or characteristics described in conjunction with that embodiment as included in at least one embodiment generally described in this application. The appearance of the same expression in multiple places in the specification does not necessarily refer to the same embodiment. Furthermore, when a specific feature, structure, or characteristic is described in conjunction with any embodiment, it is intended that such feature, structure, or characteristic, when implemented in conjunction with other embodiments, also falls within the scope of the present invention.

[0063] Although the present invention has been described herein with reference to a number of illustrative embodiments thereof, it will be understood that numerous other modifications and implementations may be devised by those skilled in the art that fall within the scope and spirit of the principles disclosed herein. More specifically, within the scope of the present disclosure, the drawings, and the claims, numerous variations and modifications may be made to the components and / or layout of the subject combination arrangement. In addition to variations and modifications to the components and / or layout, other uses will also be apparent to those skilled in the art.

Claims

1. A method for annotating and segmenting image datasets based on deep learning for fault prediction, characterized in that The method comprises the following steps: Acquire original 3D seismic data in the work area to be measured, wherein the original 3D seismic data records the amplitude information of each point in the work area to be measured in three dimensions: Inline, CDP, and Time; The original three-dimensional seismic data is subjected to dimensionality reduction processing, and the original two-dimensional matrix M is obtained by the following formula: tow-dim ; M tow-dim =Time×(Inline×CDP) According to the direction of Inline, the original two-dimensional matrix M tow-dim Perform cutting to obtain multiple cutting two-dimensional matrices M Inline ; For each of the cutting two-dimensional matrix M Inline Cut again to generate multiple two-dimensional small windows M of the current window patch Each of the two-dimensional small window M patch Time, Inline, and CDP information are all recorded; Through each two-dimensional small window M patch Recorded Inline, CDP, Time information, search for the corresponding Inline, CDP, Time information in the disclosed information of the work area to be tested, and determine the current two-dimensional small window M patch Is it the required data? If the judgment result is yes, then the current two-dimensional small window M patch Add the counter parameter as an index and put it into a new three-dimensional matrix; otherwise, continue to judge the current cutting two-dimensional matrix M Inline The next two-dimensional small window M in patch ; According to the counter parameter, the fault and non-fault label matrices are established at the same time, and the data set matrix and the label matrix are connected according to the counter parameter to complete the current cutting two-dimensional matrix M Inline Cutting and dataset,labeling; When all cuts are made on the two-dimensional matrix M Inline After the cutting and data set and label production are completed, the three-dimensional fault data set, fault labels, three-dimensional non-fault data set, and non-fault labels obtained based on the exposed information of the work area to be measured are obtained.

2. The image dataset annotation and segmentation method based on deep learning for fault prediction according to claim 1, characterized in that: The original three-dimensional seismic data in the work area to be measured includes disclosed information of the work area to be measured, and the disclosed information of the work area to be measured includes disclosed fault information and disclosed non-fault information.

3. The image dataset annotation and segmentation method based on deep learning for fault prediction according to claim 1 or 2, characterized in that: For each cutting two-dimensional matrix M Inline The cutting method is: use an n×n window and cut the two-dimensional matrix M Inline Move upward, determine the center point of the window, extend left by n / 2-1 points, right by n / 2 points, upward by n / 2-1 points, and downward by n / 2 points.

4. The image dataset annotation and segmentation method based on deep learning for fault prediction according to claim 1 or 2, characterized in that: The original three-dimensional seismic data is obtained from Geoframe software.

5. The image dataset annotation and segmentation method based on deep learning for fault prediction according to claim 1, characterized in that: The new three-dimensional matrix is ​​a small two-dimensional window M patch After adding the counter parameter as an index, the counter×n×n three-dimensional data matrix is ​​rebuilt and the two-dimensional matrix M is cut successively. Inlin After averaging, the counter is incremented and placed into the seismic data volume of the three-dimensional data matrix in sequence; The newly created three-dimensional matrix includes a fault three-dimensional matrix and a non-fault three-dimensional matrix.

6. The image dataset annotation and segmentation method based on fault prediction deep learning according to claim 1 or 5, characterized in that The method further comprises: According to the counter parameter, the three-dimensional fault dataset and the three-dimensional non-fault dataset are divided into training sets and test sets, and training labels and test labels are formed at the same time. The training set, test set, training labels and test labels are put into the convolutional neural network for deep learning of fault prediction.

Images