Karst cave bead-like abnormal reflection detection method and system and storage medium

Abstract

The invention relates to the technical field of carbonate reservoir geophysical exploration, in particular to a karst cave bead-like abnormal reflection detection method and system and a storage medium. Comprising the following steps: S1, acquiring an actual seismic section, framing bead-like reflection characteristics on the seismic section, and forming a training sample according to the bead-like reflection characteristics; s2, establishing a Yolox network model, and performing model training according to the training sample formed in the step S1; and S3, predicting the bead-like abnormal reflection of the karst cave by using the trained Yolox network model. According to the method, a target detection technology in deep learning is applied to seismic data bead-like recognition, and compared with the prior art, algorithm parameters do not need to be manually selected in the recognition process; the calculation efficiency is higher than that of a traditional seismic attribute extraction method; the influence of noise in the recognition process is small, and the recognition result is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical exploration of carbonate reservoirs, and in particular to a method, system and storage medium for detecting abnormal reflections of beaded structures in karst caves. Background Art

[0002] my country boasts abundant carbonate oil and gas reserves, offering significant exploration potential. These reservoirs are primarily fracture-cavity reservoirs, formed by fractures and caves. These reservoirs are concentrated within tectonic fractures generated by tectonic deformation and pores, caves, and cracks formed by karstification. Large caves are the primary reservoir spaces. Due to the deep burial depths, strong heterogeneity, and complex genesis and distribution patterns of these cave-type reservoirs, identifying caves presents significant challenges.

[0003] Current cave identification primarily relies on a series of traditional interpretation methods, including seismic wave reflection feature analysis, cave attribute analysis, and well logging response feature analysis. For example, prior art CN109655904A discloses a multi-scale cave identification method and system, which includes: performing forward modeling analysis on caves of different scales to establish the relationship between the seismic dominant frequency, cave scale, and root mean square amplitude attributes; extracting the root mean square amplitude attribute from the actual frequency-divided data volume, analyzing the relationship between the seismic dominant frequency, cave scale, and root mean square amplitude attributes, and verifying consistency with the forward modeling conclusions; and utilizing the relationship between the seismic dominant frequency, cave scale, and root mean square amplitude attributes to identify caves of different scales.

[0004] These methods have high data requirements, and each interpretation method can only be applied to a specific scenario, making them limited in generality. Furthermore, these methods still require manual data interpretation and identification, which is inefficient and subject to human subjective factors. Furthermore, traditional deep learning methods primarily study large-scale structures such as salt domes, faults, and strata, but are prone to misjudging small-scale structures such as caves.

[0005] Therefore, in order to solve the technical problem that small-scale cave structures are easily misjudged in the existing technology, it is urgent to provide a method, system and storage medium for detecting abnormal reflections of beaded caves. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems existing in the prior art, the present invention provides a method, system and storage medium for detecting abnormal beaded reflections in caves, so as to realize the rapid and accurate detection of "beaded" reflection features and caves. According to previous studies, caves of different sizes show different reflection features on seismic profiles, and there is a certain mapping relationship between caves and "beaded" reflection features. Therefore, relatively large-scale "beaded" reflection features can be first identified on the seismic profile, and then the identification from "beaded" to caves can be completed based on the mapping relationship between "beaded" and caves. The identification of "beaded" is equivalent to a target detection problem. As a new detection algorithm, the Yolo (You only look once) network is based on end-to-end target detection of deep learning, integrating target area prediction and target category into a single neural network, which can realize accurate and rapid target detection and recognition. As the latest algorithm, Yolox has faster detection speed, higher accuracy and more flexible end-to-end deployment.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for detecting abnormal bead-like reflections in a karst cave, comprising:

[0009] S1. Obtain an actual seismic profile, frame a "beaded" reflection feature on the seismic profile, and form a training sample based on the "beaded" reflection feature;

[0010] S2. Establish a Yolox network model and perform model training based on the training samples formed in step S1;

[0011] S3. Use the trained Yolox network model to predict the abnormal reflection of cave beads.

[0012] Furthermore, in step S1, a training sample is formed according to the "beaded" reflection feature, specifically including: according to the "beaded" reflection feature, the position and size of the "beaded" reflection frame are used as training labels, the root mean square amplitude and instantaneous amplitude are introduced as constraints, and superimposed with the original seismic data to form a training sample.

[0013] Furthermore, before superposition, the right and bottom sides of the original seismic data, RMS amplitude data, and instantaneous amplitude data are padded with zeros to obtain data with a pixel size of 640×1280. Then, the original seismic data, RMS amplitude data, and instantaneous amplitude data are superimposed to obtain a training sample with a size of 640×1280×3.

[0014] Furthermore, the Yolox network model includes feature extraction, feature enhancement and decoupling heads, and the feature extraction network is a cross-stage local network.

[0015] Furthermore, during the Yolox network model training process, the feature extraction part extracts shallow features, middle features, and deep features. The shallow feature data size is 80×160×256, the middle feature data size is 40×80×512, and the deep feature data size is 20×40×1024.

[0016] Furthermore, the feature enhancement network includes: a bottom-up feature extraction process, and a top-down and lateral connection fusion process, and finally outputs three feature layers.

[0017] Furthermore, the model training in step S2 is iterated 50 times in total, wherein during the first 25 iterations, the feature extraction network is frozen with an initial learning rate of 0.001; during the next 25 iterations, the feature extraction network is unfrozen with an initial learning rate of 0.0001.

[0018] Furthermore, after obtaining the prediction results, decoding is performed to regress the output coordinate positions to the actual coordinates on the actual seismic section; the decoding process is performed according to the following algorithm:

[0019] (x+x_c,y+y_c,w,h)×strides

[0020] Among them, x and y represent the bias of the center of the prediction box relative to the upper left corner of the network, x_c and y_c represent the position of the output grid, w and h represent the parameters of the prediction box width and height relative to the grid, and strides represents the downsampling multiple.

[0021] Furthermore, after decoding is completed, the trust score is calculated, and the prediction boxes whose trust scores are greater than the confidence level are screened out. Then, the prediction box with the largest score is screened out from the prediction boxes whose trust scores are greater than the confidence level as the final prediction box.

[0022] Furthermore, the trust score is calculated as follows:

[0023] P = pr × IoU

[0024] Among them, P represents the trust value score, IoU represents the intersection over union ratio, and the calculation formula of the intersection over union ratio is:

[0025]

[0026] B gt represents the ground-truth box, B pred Represents the prediction box;

[0027] When the detection box contains the target object, pr is 1, and when the detection box does not contain the target object, pr is 0.

[0028] Furthermore, the confidence value is 0.85.

[0029] The present invention further provides a system for detecting abnormal beaded reflections in karst caves, which adopts the above-mentioned method for detecting abnormal beaded reflections in karst caves, and comprises:

[0030] A data processing unit is used to obtain an actual seismic profile, frame the "beaded" reflection features on the seismic profile, and form training samples based on the "beaded" reflection features;

[0031] Model training unit, used to establish Yolox network model and perform model training based on training samples;

[0032] The prediction unit is used to predict the abnormal reflection of cave beads using the trained Yolox network model.

[0033] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned method for detecting abnormal bead-like reflections in a karst cave is implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention provides a method and system for detecting abnormal beaded reflections in karst caves. Training samples are generated based on actual seismic profiles, and a Yolox network model is used for training. The trained Yolox network model is then used to predict abnormal beaded reflections in karst caves, thereby determining the actual location of the beads on the seismic profile. Based on the autonomous learning data characteristics of neural networks, the present invention incorporates target detection techniques from deep learning into the identification of "beaded" reflection characteristics in seismic data. This enables rapid and efficient detection, providing a basis for cave identification. Furthermore, it reduces subjectivity, eliminates the need for manual selection of algorithm parameters, and achieves high computational efficiency and more accurate identification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flow chart of the method of the present invention.

[0037] Figure 2a The original seismic data.

[0038] Figure 2b for Figure 2a The data graph is obtained after zero padding on the right and bottom sides. The borders in the figure are used to distinguish the image size before and after zero padding.

[0039] Figure 2c This is the data diagram after the original seismic data, root mean square data, and instantaneous data are superimposed.

[0040] Figure 2d This is the earthquake data map after adding training labels.

[0041] Figure 3 It is a structural diagram of the Yolox network model of the present invention.

[0042] Figure 4 This is a training error curve diagram of the present invention, where the horizontal axis epoch represents the number of iterations in which all data are used once, and the vertical axis loss represents the error value.

[0043] Figure 5a This is the original cross-section of the actual work area provided by the present invention.

[0044] Figure 5b This is the result after using instantaneous amplitude on the seismic section.

[0045] Figure 5c This is the result after applying the root mean square amplitude to the seismic section.

[0046] Figure 5d This is the identification result of the seismic profile by the present invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] It should be noted that, unless otherwise specifically stated, the relative arrangements of components and steps, and numerical expressions set forth in these embodiments should not be construed as limiting the scope of the present invention.

[0049] The following description of exemplary embodiments is merely illustrative and is not intended to limit the present invention, its application, or use in any sense. Technologies, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but to the extent applicable, such technologies, methods, and apparatuses should be considered part of this specification.

[0050] Example 1

[0051] The present invention provides a method for detecting abnormal reflections of beaded caves. Figure 1 As shown, including:

[0052] S1. Obtain an actual seismic profile, frame a "beaded" reflection feature on the seismic profile, and form a training sample based on the "beaded" reflection feature;

[0053] First, a typical seismic profile is selected from actual seismic data. Based on the seismic reflection characteristics of karst caves, a "beaded" reflection feature is framed on the seismic profile. The position and size of the "beaded" reflection frame are obtained as training labels. The root mean square amplitude and instantaneous amplitude are introduced as constraints and superimposed with the original seismic data as the three-channel input. Due to the varying sizes of the seismic data, zero padding is applied to the bottom and right sides to make the input data consistent. Figure 2a The original seismic data is 601×1206 in size. After zero padding on the right and bottom sides, the resulting data size is 640×1280. Figure 2b Similarly, the RMS amplitude and instantaneous amplitude are padded with zeros on the right and bottom sides, and then the original seismic data after zero padding, instantaneous amplitude, and RMS amplitude are superimposed in this order to obtain a training sample of 640×1280×3 pixels as output, as shown in Figure 2c The training labels are the positions and sizes of the corresponding “beaded” reflection features on the seismic profile, as shown in Figure 2d shown.

[0054] S2. Establish a Yolox network model and perform model training based on the training samples formed in step S1;

[0055] Conduct Yolox network model training, including feature extraction, feature enhancement and decoupling head three parts, such as Figure 3 As shown in the figure, the feature extraction network is a cross-stage local network (CSPDarknet). The feature extraction network part is divided into multiple residual convolutions and maximum pooling, which makes the feature extraction more sufficient. It is worth noting that the spatial pyramid pooling network is introduced before the input feature enhancement network. By pooling with pooling kernels of different sizes, the network's receptive field is improved and richer feature information is extracted.

[0056] The input data undergoes multiple residual convolutions and activation functions to complete feature extraction. In the feature extraction phase, the Yolox network extracts three different levels of feature data: shallow, medium, and deep. This fusion process facilitates the extraction of effective features. Given a text input data size of (640, 1280, 3), the feature extraction network generates three different levels of features: shallow features F1 (80, 160, 256), medium features F2 (40, 80, 512), and deep features F3 (20, 40, 1024). These features are then fed into the feature enhancement network for enhanced performance.

[0057] The Feature Enhancement Network (FPN) employs a feature pyramid structure, consisting of two main components: a bottom-up feature extraction process and a top-down and lateral connection fusion process. This involves upsampling high-level features and top-down concatenating low-level features. Each layer undergoes a simple network connection change, significantly improving small object detection performance with minimal increase in the computational overhead of the original model. The bottom-up process is completed during feature extraction. The feature maps are typically reduced in size after being processed by convolution kernels. Considering the insufficient feature extraction and low semantics of the first two layers, the last three layers (F1, F2, and F3) are selected as the input features of the FPN. The bottom-up process upsamples the more abstract and semantically rich high-level feature maps, while the lateral connection fuses the upsampled results with the bottom-up generated feature maps of the same size. The two lateral connection feature maps have the same spatial size, allowing for the utilization of detailed localization information from the lower layers. The low-resolution feature maps are upsampled by a factor of two. The upsampled maps are then merged with the corresponding bottom-up maps through element-wise addition. This process is iterative until the final resolution map is generated. The final output is three feature-enhanced feature layers: P3_out (80, 60, 256), P4_out (40, 80, 512), and P5_out (20, 40, 1024).

[0058] Yolox uses a decoupled head structure to separate the classification problem from the regression problem. The classification task network enables it to recognize more multi-scale fine-grained local information for different images, facilitating classification. For the regression task, more overall information is recognized by adding a coordinate convolutional network. This can effectively reduce the number of parameters and computational complexity, enhance the generalization and robustness of the model, and thus improve detection capabilities. Figure 3 As shown in the figure, the three enhanced features obtained after the FPN network undergo dimensionality reduction and channel unification through a 1×1 convolution, and are then input into the Yolo head. Subsequent work can be divided into two branches: classification and regression. The regression branch also has two parallel branches. Each branch is followed by two 3×3 convolutions and then a 1×1 convolution to adjust the channels. The number of channels in the Cls branch corresponds to the number of categories, which is 1 in this article. The size of the Reg branch is 4, and regression is used to obtain the coordinate parameters of the predicted box. The number of channels in the IoU branch is 1, indicating whether the object is present, that is, 0 or 1. The prediction results of the three branches are stacked to obtain the results of each feature layer. In this article, the three enhanced feature layers are processed by the decoupling head and the stacked prediction results are (20, 40, 6), (40, 80, 6), and (80, 160, 6), respectively.

[0059] The present invention iterates 50 times in total. The first 25 iterations freeze the feature extraction network, the parameters change slightly, the training speed is accelerated, and the initial learning rate is 0.001. The feature extraction network is unfrozen in the last 25 iterations, the parameters change with the network iteration, and the entire network is optimized. At this stage, the initial learning rate is 0.0001. As the number of iterations increases, the network parameters are gradually optimized, and the model loss function value becomes smaller and smaller. When the loss function value is reduced to a smaller value and basically no longer changes, it means that the model has reached convergence and obtained the optimal weight. The training error curve is as follows: Figure 4 shown.

[0060] S3. Use the trained Yolox network model to predict the abnormal reflection of the beaded cave.

[0061] The trained Yolox network model will be used to predict the "beaded" abnormal reflections in carbonate reservoir caves. After the Yolox network model is trained and the prediction results are obtained through the above-mentioned feature extraction, feature enhancement, and decoupling, decoding is required to regress the output coordinate positions relative to the network to the actual coordinates on the actual seismic profile. The decoding process can be expressed as follows:

[0062] (x+x_c,y+y_c,w,h)×strides

[0063] Among them, x and y represent the offset of the center of the prediction box relative to the upper left corner of the network, x_c and y_c represent the position of the output grid, w and h represent the parameters of the prediction box width and height relative to the grid, and strides represents the downsampling multiple. Therefore, the decoding process can be understood as the offset of the prediction box relative to the upper left corner of the grid plus the position of the grid, multiplied by the downsampling multiple, and mapped to the original seismic profile position. After decoding is completed, the trust score is calculated, and the prediction box with a trust score greater than the confidence level is selected. Then, the prediction box with the largest score is selected from the prediction boxes with a trust score greater than the confidence level as the final prediction box. In this embodiment, the confidence value is 0.85.

[0064] The trust score indicates the accuracy of the predicted box. The score is calculated based on two factors: the probability that the detection box contains the target object; and the degree of overlap between the detected bounding box and the bounding box of the real object. The trust score is calculated as follows:

[0065] P = pr × IoU

[0066] Among them, P represents the trust value score, IoU represents the intersection over union ratio, and the calculation formula of the intersection over union ratio is:

[0067]

[0068] B gtrepresents the real box (manually calibrated data in the training sample), B pred Represents the prediction box;

[0069] When the detection box contains the target object (beaded features), pr is 1, and when the detection box does not contain the target object, pr is 0.

[0070] Finally, the position of the "beaded" detection frame on the seismic section is obtained. Using the seismic section as the output of the network model, the coordinate position of the "beaded" abnormal reflection feature on the seismic section can be obtained.

[0071] In one embodiment of the present invention, seismic data from a region in northwest China is used to identify fractures and holes. Figure 5a The original cross-section of the actual work area provided for a certain area in the northwest. Figure 5b This is the result of applying instantaneous amplitude to the seismic profile. Figure 5c is the result of applying the root mean square amplitude to the seismic section. Figure 5d This is the identification result of the seismic profile by the present invention. For the traditional root mean square amplitude and instantaneous amplitude methods, which are based on amplitude, it is difficult to identify the abnormal reflection feature of the beaded shape when the beaded shape is on or near a strong phase axis. The high value part obtained at the same time cannot reflect the actual beaded range. If two beaded shape abnormal reflections are very close, they are easily identified as one beaded shape abnormality. The recognition accuracy is not high, and there are errors or omissions. For the method of the present invention, the detection frame obtained can reflect the position and actual size of the beads and regress to obtain the specific coordinates of the beads, and is not affected by the phase axis energy of the seismic data. It provides a basis for subsequent cave identification.

[0072] Example 2

[0073] The present invention further provides a system for detecting abnormal beaded reflections in a karst cave, which adopts the method for detecting abnormal beaded reflections in a karst cave provided in Example 1, comprising:

[0074] A data processing unit is used to obtain an actual seismic profile, frame the "beaded" reflection features on the seismic profile, and form training samples based on the "beaded" reflection features;

[0075] Model training unit, used to establish Yolox network model and perform model training based on training samples;

[0076] The prediction unit is used to predict the abnormal reflection of cave beads using the trained Yolox network model.

[0077] Example 3

[0078] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for detecting abnormal beaded reflections in a karst cave provided in Example 1 is implemented.

[0079] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the scope of the technical solutions of the present invention, and all of these should be included in the scope of the claims of the present invention.

Claims

1. A method for detecting abnormal reflections of beaded caves, characterized in that: include: S1. Obtain an actual seismic profile, frame a "beaded" reflection feature on the seismic profile, and form a training sample based on the "beaded" reflection feature; S2. Establish a Yolox network model and perform model training based on the training samples formed in step S1; S3. Use the trained Yolox network model to predict the abnormal reflection of cave beads.

2. The method for detecting abnormal reflections of karst cave beads according to claim 1, characterized in that: In step S1, a training sample is formed based on the "beaded" reflection feature, specifically including: based on the "beaded" reflection feature, the position and size of the "beaded" reflection frame are used as training labels, the root mean square amplitude and instantaneous amplitude are introduced as constraints, and they are superimposed on the original seismic data to form a training sample.

3. The method for detecting abnormal reflections of karst cave beads according to claim 2, characterized in that: Before stacking, the right and bottom sides of the original seismic data, RMS amplitude data, and instantaneous amplitude data were padded with zeros to obtain data with a pixel size of 640×1280. Then, the original seismic data, RMS amplitude data, and instantaneous amplitude data were stacked to obtain a training sample with a size of 640×1280×3.

4. The method for detecting abnormal reflections of karst cave beads according to claim 1, characterized in that: The Yolox network model includes feature extraction, feature enhancement and decoupling heads, and the feature extraction network is a cross-stage local network.

5. The method for detecting abnormal reflections of karst cave beads according to claim 4, characterized in that: During the Yolox network model training process, the feature extraction part extracts shallow features, middle features, and deep features. The shallow feature data size is 80×160×256, the middle feature data size is 40×80×512, and the deep feature data size is 20×40×1024.

6. The method for detecting abnormal beaded reflections in a karst cave according to claim 5, characterized in that: The feature enhancement network includes: a bottom-up feature extraction process, and a top-down and horizontal connection fusion process, and finally outputs three feature layers.

7. The method for detecting abnormal beaded reflections in a karst cave according to claim 4, wherein: The model training in step S2 is iterated 50 times. During the first 25 iterations, the feature extraction network is frozen with an initial learning rate of 0.001; during the next 25 iterations, the feature extraction network is unfrozen with an initial learning rate of 0.0001.

8. The method for detecting abnormal reflections of karst cave beads according to claim 1, characterized in that: After obtaining the prediction results, decoding is performed to regress the output coordinate positions to the actual coordinates on the actual seismic section; the decoding process is performed according to the following algorithm: (x+x_x,y+y_c,w,h)×strides Among them, x and y represent the bias of the center of the prediction box relative to the upper left corner of the network, x_c and y_c represent the position of the output grid, w and h represent the parameters of the prediction box width and height relative to the grid, and strides represents the downsampling multiple.

9. The method for detecting abnormal beaded reflections in a karst cave according to claim 8, wherein: After decoding is completed, the trust score is calculated, and the prediction boxes whose trust scores are greater than the confidence level are filtered out. Then, the prediction box with the largest score is filtered out from the prediction boxes whose trust scores are greater than the confidence level as the final prediction box.

10. The method for detecting abnormal reflections of beaded caves according to claim 9, characterized in that: The trust score is calculated as follows: P = pr × IoU Among them, P represents the trust value score, IoU represents the intersection over union ratio, and the calculation formula of the intersection over union ratio is: B gt represents the ground-truth box, B pred Represents the prediction box; When the detection box contains the target object, pr is 1, and when the detection box does not contain the target object, pr is 0.

11. The method for detecting abnormal reflections of karst cave beads according to claim 9, characterized in that: The confidence level is 0.

85.

12. A system for detecting abnormal beaded reflections in a karst cave, using the method for detecting abnormal beaded reflections in a karst cave according to any one of claims 1 to 11, characterized in that: include: A data processing unit is used to obtain an actual seismic profile, frame the "beaded" reflection features on the seismic profile, and form training samples based on the "beaded" reflection features; Model training unit, used to establish Yolox network model and perform model training based on training samples; The prediction unit is used to predict the abnormal reflection of cave beads using the trained Yolox network model.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting abnormal beaded reflections in a cave as described in any one of claims 1 to 11 is implemented.

Images