Earthquake large model construction method, device and equipment based on computer vision algorithm

Abstract

The invention discloses an earthquake large model construction method, device and equipment based on a computer vision algorithm. The method comprises the following steps: acquiring label seismic data; extracting profile data in the tag seismic data to obtain multiple groups of profile data; randomly selecting at least one preset mode to mask the plurality of groups of profile data, and constructing a plurality of groups of mask profile samples; based on self-supervised learning, respectively inputting the plurality of groups of mask profile samples into a plurality of pre-constructed computer vision algorithm models for training to obtain a plurality of base models; and fusing the plurality of base models based on an attention mechanism, and constructing a large earthquake model. According to the method, features of different dimensions in seismic data are extracted by using different computer vision algorithms, and the features are fused based on an attention mechanism, so that the constructed large seismic model can obtain more comprehensive and accurate feature representation, and the large seismic model with high calculation precision and strong generalization ability is constructed. And the accuracy and timeliness of intelligent identification of the seismic section are improved.

Description

Technical Field

[0001] The present invention relates to a method, a device and equipment for constructing a large earthquake model based on a computer vision algorithm. Background Art

[0002] Since the Transformer model architecture was launched in 2017, it has attracted widespread attention in the field of natural language processing as a powerful sequence modeling method, and has spawned a series of important pre-training models, such as BERT and T5. These models are based on the idea of ​​self-supervised learning, which uses large-scale text data for pre-training to learn the representation ability of language models, and applies the feature vectors obtained by unsupervised training to various downstream tasks, and fine-tunes these downstream tasks with limited labeled data, effectively improving task efficiency and accuracy.

[0003] In the field of computer vision, similar to Transformer, a series of important large model architectures have emerged in recent years, such as ViT and MAE. The emergence of these models is also based on a similar idea, that is, learning the ability to represent visual information through pre-training large-scale data sets. The ViT model divides the image into fixed-size tiles and then converts them into sequence inputs, so that the Transformer model can be directly used for training and reasoning, and the traditional convolutional neural network can be applied to visual tasks; the MAE model jointly models visual information with other modalities (such as text and sound), and learns multimodal representations through self-encoding, thereby achieving good results in multimodal tasks. The introduction of these two models further expands the application field of large models, making them not only limited to natural language processing, but also including fields such as computer vision.

[0004] Although the research and application of large models have achieved remarkable results in fields such as natural language processing and computer vision, relevant work has not yet been widely carried out in the field of oil and gas exploration. Therefore, before the industrial development model is applied to other fields, it is still necessary to fully study and verify the adaptability and effect of large models in specific fields. Summary of the invention

[0005] In order to better realize the construction of a large earthquake model, an embodiment of the present invention provides a method, device and equipment for constructing a large earthquake model based on a computer vision algorithm.

[0006] In a first aspect, an embodiment of the present invention provides a method for constructing a large earthquake model based on a computer vision algorithm, the method comprising:

[0007] Get labeled earthquake data;

[0008] Extracting profile data from the labeled seismic data to obtain multiple groups of profile data;

[0009] Randomly selecting at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples;

[0010] Based on self-supervised learning, the plurality of groups of mask profile samples are respectively input into a plurality of pre-built computer vision algorithm models for training to obtain a plurality of base models;

[0011] The multiple base models are fused based on the attention mechanism to construct a large earthquake model.

[0012] In an optional implementation of the embodiment of the present invention, before labeling the seismic data to obtain the labeled seismic data, the method further includes:

[0013] Perform data statistics on multiple sets of seismic data obtained, clean abnormal data based on the normal distribution method, and obtain cleaned seismic data;

[0014] By data normalization and standardization, the cleaned seismic data is mapped into a preset range to obtain preprocessed seismic data;

[0015] The preprocessed seismic data are labeled to obtain corresponding labeled seismic data.

[0016] In an optional implementation of the embodiment of the present invention, the seismic data includes pre-stack seismic data and post-stack seismic data; labeling the pre-processed seismic data to obtain corresponding labeled seismic data includes:

[0017] For the prestack seismic data, high-quality reservoir position analysis is performed according to the relationship between amplitude and offset distance to obtain labeled prestack seismic data;

[0018] For the post-stack seismic data, the structural characteristics of the target layer are obtained according to the lateral and vertical changes of the amplitude, and reservoir prediction is performed to obtain labeled post-stack seismic data.

[0019] In an optional implementation of the embodiment of the present invention, extracting the profile data from the labeled seismic data to obtain multiple groups of profile data includes:

[0020] The profile data are extracted along different inline, crossline and tangent directions of the labeled pre-stack seismic data and / or the labeled post-stack seismic data to obtain multiple groups of profile data.

[0021] In an optional implementation of the embodiment of the present invention, the randomly selecting at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples includes:

[0022] Randomly selecting different preset masking ratios, masking the profile data, and constructing multiple groups of mask profile samples;

[0023] The profile data is randomly divided into a plurality of regions, and for the data in each region, different preset masking area blocks are randomly selected for masking, so as to construct a plurality of groups of mask profile samples;

[0024] Different values ​​are randomly selected from the profile data, and the values ​​in the profile data are randomly masked with the randomly selected different values ​​to construct multiple groups of mask profile samples.

[0025] In an optional implementation of the embodiment of the present invention, based on the self-supervised learning, the multiple groups of mask profile samples are respectively input into multiple pre-built computer vision algorithm models for training to obtain multiple base models, including:

[0026] Inputting the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training respectively, to obtain a plurality of groups of reconstructed profiles; the computer vision algorithms include Unet, FCN and Deeplab computer vision algorithms;

[0027] Comparing the plurality of groups of mask profile samples with the corresponding reconstructed profiles respectively to obtain reconstruction errors;

[0028] Based on the reconstruction error, adjusting the weight of the corresponding computer vision algorithm model;

[0029] The above process is repeated until the reconstruction error meets the preset conditions, and multiple base models corresponding to multiple computer vision algorithm models are obtained.

[0030] In an optional implementation of the embodiment of the present invention, the fusing of the multiple base models based on the attention mechanism to construct the earthquake large model further includes:

[0031] A fusion layer is constructed based on the attention mechanism to fuse the features of different dimensions of the multiple base models to construct a large earthquake model.

[0032] In a second aspect, an embodiment of the present invention provides a large earthquake model construction device based on a computer vision algorithm, the device comprising:

[0033] An acquisition module, used to acquire labeled earthquake data;

[0034] An extraction module, used for extracting profile data from the labeled seismic data to obtain multiple groups of profile data;

[0035] A masking module, used for randomly selecting at least one preset method to mask the multiple groups of profile data, and constructing multiple groups of masked profile samples;

[0036] A self-supervised learning module, for inputting the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training based on self-supervised learning to obtain a plurality of base models;

[0037] The model building module is used to fuse the multiple base models based on the attention mechanism to build a large earthquake model.

[0038] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for constructing a large earthquake model based on a computer vision algorithm.

[0039] In a fourth aspect, an embodiment of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for constructing a large earthquake model based on a computer vision algorithm as described above is implemented.

[0040] In a fifth aspect, an embodiment of the present invention provides a computer program product comprising instructions, which, when executed on a computer device, enables the computer device to execute the above-mentioned method for constructing a large earthquake model based on a computer vision algorithm.

[0041] In a sixth aspect, an embodiment of the present invention provides a chip, the chip comprising a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run a computer program or instruction to implement the above-mentioned method for constructing a large earthquake model based on a computer vision algorithm.

[0042] The beneficial effects of the above technical solution provided by the embodiment of the present invention include at least:

[0043] The method for constructing a large earthquake model based on a computer vision algorithm provided by an embodiment of the present invention improves the diversity of training data by sampling earthquake data and extracting random sections along different inlines, crosslines and arbitrary tangent directions; masks earthquake data by drawing on the image processing method, and masks the multiple groups of profile data by randomly selecting at least one preset method to construct multiple groups of masked profile samples; uses different computer vision algorithms to extract features of different dimensions in earthquake data, and fuses the features based on the attention mechanism, so that the constructed large earthquake model can obtain a more comprehensive and accurate feature representation. This method constructs a large earthquake model with high calculation precision, good accuracy, strong generalization ability, stronger robustness and strong interpretability based on computer vision algorithms, improves the accuracy and timeliness of intelligent identification of earthquake profiles, and breaks through a series of problems such as the fragmentation of current intelligent applications, model customization and AI threshold.

[0044] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0045] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0047] Figure 1 A schematic diagram of the steps of a method for constructing a large earthquake model based on a computer vision algorithm provided in an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of the structure of a unet neural network provided by an embodiment of the present invention;

[0049] Figure 3 A schematic diagram of the structure of an FCN neural network provided by an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of the structure of a Deeplab neural network provided in an embodiment of the present invention;

[0051] Figure 5 A schematic diagram of a process for constructing a self-supervised learning task provided by an embodiment of the present invention;

[0052] Figure 6 A schematic diagram of the structure of a decoding network in a prediction model provided by an embodiment of the present invention;

[0053] Figure 7 A schematic diagram of the structure of an upsampling module provided in an embodiment of the present invention;

[0054] Figure 8 A training flow chart of a prediction model provided by an embodiment of the present invention;

[0055] Fig. 9 A schematic diagram of a profile of original seismic data of the Tarim Oilfield provided by an embodiment of the present invention;

[0056] Fig.10 A schematic diagram of a fracture-cavity body reflection boundary prediction result and a regression wave impedance prediction result obtained by predicting the original seismic data of the Tarim Oilfield provided in an embodiment of the present invention;

[0057] Fig.11A schematic diagram of the original seismic data profile and the predicted fracture-cavity reflection boundary position in the Tarim Oilfield provided by an embodiment of the present invention;

[0058] Fig.12 A schematic diagram of the structure of a large earthquake model building device based on a computer vision algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0060] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0061] It should also be understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0062] As used in the present specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]", depending on the context.

[0063] In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0064] References to "one embodiment" or "some embodiments" etc. described in the present specification mean that one or more embodiments of the present invention include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0065] It should be understood that the order of execution of the steps in the following embodiments does not imply a precedence of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0066] In order to illustrate the technical solution of the present invention, specific embodiments are provided below for illustration.

[0067] The inventors found that in the prior art, in the field of oil and gas exploration, compared with natural language processing and computer vision, the research and application of large artificial intelligence models are still relatively small. The long development cycle of large models, the need for separate optimization, difficult iteration, low efficiency and poor model generalization ability limit the application of artificial intelligence models in the field of oil and gas exploration. In addition, in the field of seismic exploration, although a large amount of data has been accumulated, due to the high manpower and time costs of labeling, most samples are usually unlabeled and cannot be fully utilized. This does not meet the inventor's expectations. Based on this, the inventors have made the present invention after further research and development, providing a method, device and equipment for constructing a large earthquake model based on a computer vision algorithm.

[0068] Embodiment 1

[0069] The embodiment of the present invention provides a method for constructing a large earthquake model based on a computer vision algorithm, referring to Figure 1 As shown, the method includes:

[0070] S101: Obtain labeled earthquake data.

[0071] In the embodiment of the present invention, before the seismic data is labeled to obtain the labeled seismic data, the method further includes:

[0072] Perform data statistics on multiple sets of seismic data obtained, clean abnormal data based on the normal distribution method, and obtain cleaned seismic data;

[0073] By data normalization and standardization, the cleaned seismic data is mapped into a preset range to obtain preprocessed seismic data;

[0074] The preprocessed seismic data are labeled to obtain corresponding labeled seismic data.

[0075] In an embodiment of the present invention, pre-stack seismic data and post-stack seismic data generated during seismic exploration are collected and sorted, wherein pre-stack seismic data mainly refers to CRP gather data, and post-stack seismic data includes attribute data such as seismic amplitude data and wave impedance. After the seismic data is collected, it is necessary to pre-process the seismic data, including operations such as abnormal data cleaning, data normalization and standardization. Among them, abnormal data cleaning is specifically to perform data statistics on multiple groups of seismic data and use the normal distribution method to clean out some abnormal data; data normalization and standardization specifically map the cleared seismic data to a preset range, so that the processed seismic data can better perform data analysis and model training.

[0076] In addition to preprocessing the collected seismic data in the above manner, the seismic data can also be denoised by filtering the seismic data through low-pass filtering, median filtering and high-pass filtering to eliminate noise; after converting the signal to the frequency domain through Fourier transform, the frequency component can be adjusted to achieve the purpose of denoising.

[0077] After preprocessing the seismic data, various statistical methods and machine learning techniques are used to extract and analyze the features of the processed data. These features include the magnitude of the seismic amplitude, the magnitude of the seismic wave impedance, the shape of the seismic waveform, and the seismic amplitude frequency, which help to understand and describe the content of the seismic data.

[0078] According to the results of the feature analysis of different dimensions of the pre-processed seismic data, the inventors chose to use Unet, Fully Convolution Network (FCN) and deeplab computer vision algorithms to build corresponding base models to extract corresponding features, wherein Unet, FCN and deeplab neural network algorithms are all algorithms suitable for computer vision. In the architecture of Unet, FCN and deeplab neural networks, multiple branches and multiple convolutional layers are used to extract features of different scales, and shallow features and deep features are connected through jump connections, so that the network can use low-level features and high-level features for prediction at the same time, effectively improving the expression ability of the model.

[0079] In a specific embodiment, the network structure of the Unet neural network can be as follows: Figure 2 As shown, the network structure of the FCN neural network can be as follows Figure 3 As shown, the network structure of the deeplab neural network can be as follows Figure 4As shown in the figure, the settings of each neural network hyperparameter can be set according to the actual situation.

[0080] In the embodiment of the present invention, the seismic data includes pre-stack seismic data and post-stack seismic data; labeling the pre-processed seismic data to obtain corresponding labeled seismic data specifically includes:

[0081] For the prestack seismic data, high-quality reservoir position analysis is performed according to the relationship between amplitude and offset distance to obtain labeled prestack seismic data;

[0082] For the post-stack seismic data, the structural characteristics of the target layer are obtained according to the lateral and vertical changes of the amplitude, and reservoir prediction is performed to obtain labeled post-stack seismic data.

[0083] In an embodiment of the present invention, for preprocessed seismic data, part of the preprocessed seismic data is labeled by sampling to obtain labeled seismic data. For prestack seismic data, the relationship between the amplitude and the offset distance (incident angle) in the prestack seismic data is processed. The amplitude at the same position is inconsistent due to the different offset distances (incident angles). According to this change, the P-wave velocity, S-wave velocity and density solved by the simplified Zoeppritz equation can calculate various elastic parameters such as the longitudinal and transverse wave velocity ratio, Poisson's ratio and Lame coefficient. According to rock physics analysis, elastic parameters sensitive to reservoir fluids in this area are selected. At the same time, combined with geological knowledge such as structure and reservoir formation, high-quality reservoir position analysis is performed to obtain labeled prestack seismic data.

[0084] For post-stack seismic data, the structural characteristics of the target layer and reservoir prediction are obtained through the horizontal and vertical changes of amplitude, and labeled post-stack seismic data are obtained. The main application features include the strength of the amplitude, the shape of the waveform, the lateral continuity and discontinuity of the waveform, etc.

[0085] S102: extracting profile data from the labeled seismic data to obtain multiple groups of profile data.

[0086] In the embodiment of the present invention, extracting the profile data from the labeled seismic data to obtain multiple groups of profile data specifically includes:

[0087] The profile data are extracted along different inline, crossline and tangent directions of the labeled pre-stack seismic data and / or the labeled post-stack seismic data to obtain multiple groups of profile data.

[0088] In the embodiment of the present invention, sections are randomly extracted along different inline, crossline and tangent directions of labeled pre-stack seismic data and / or labeled post-stack seismic data, so as to ensure the diversity of training data.

[0089] S103: Randomly select at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples.

[0090] In the embodiment of the present invention, the randomly selecting at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples specifically includes:

[0091] Randomly selecting different preset masking ratios, masking the profile data, and constructing multiple groups of mask profile samples;

[0092] The profile data is randomly divided into a plurality of regions, and for the data in each region, different preset masking area blocks are randomly selected for masking, so as to construct a plurality of groups of mask profile samples;

[0093] Different values ​​are randomly selected from the profile data, and the values ​​in the profile data are randomly masked with the randomly selected different values ​​to construct multiple groups of mask profile samples.

[0094] In the embodiment of the present invention, the inventors borrowed from the image processing method, used the seismic data as image data for masking, and set a variety of preset methods, including:

[0095] 1. Modify the proportion of masked data in the profile data. You can set different masking ratios according to actual needs, such as 25%, 50%, and 75%. By changing the masking ratio, you can simulate data missing situations of different severity, thereby evaluating the tolerance for data missing.

[0096] 2. Divide the profile data into multiple areas, and mask the data in each area. According to actual needs, you can randomly use different masking size blocks to mask the data in the area, such as 8x8, 16x16, and 32x32, to simulate data missing in different ranges. This can better understand the impact of data missing in different ranges on subsequent processing and analysis.

[0097] 3. Select data with different values ​​in the profile data for masking. Randomly select different values ​​in the profile data, and randomly mask the values ​​in the profile data with the randomly selected different values. Select according to actual needs, for example, you can fill it with 0 values ​​or Gaussian noise values. You can simulate different types of data missing situations, so as to evaluate the processing effect on different signal types.

[0098] In the process of sample construction, the above preset methods can be combined and adjusted according to actual needs to generate a series of samples under different circumstances and construct multiple sets of mask profile samples. These samples can be used to test and evaluate the performance of algorithms and models based on profile data, and can also provide references for subsequent data processing and analysis. By randomly masking the seismic profile in the manner of image processing and repairing the image through computer vision algorithms, a large earthquake model can be established.

[0099] S104: Based on self-supervised learning, the multiple groups of mask profile samples are respectively input into multiple pre-built computer vision algorithm models for training to obtain multiple base models.

[0100] In the embodiment of the present invention, the self-supervised learning is based on inputting the multiple groups of mask profile samples into multiple pre-built computer vision algorithm models for training to obtain multiple base models, specifically including:

[0101] Inputting the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training respectively, to obtain a plurality of groups of reconstructed profiles; the computer vision algorithms include Unet, FCN and Deeplab computer vision algorithms;

[0102] Comparing the plurality of groups of mask profile samples with the corresponding reconstructed profiles respectively to obtain reconstruction errors;

[0103] Based on the reconstruction error, adjusting the weight of the corresponding computer vision algorithm model;

[0104] The above process is repeated until the reconstruction error meets the preset conditions, and multiple base models corresponding to multiple computer vision algorithm models are obtained.

[0105] In an embodiment of the present invention, the constructed multiple groups of mask profile samples are respectively input into the Unet, FCN and Deeplab neural networks for training through self-supervised learning. By constructing the self-supervised learning task, the massive data accumulated in the seismic exploration direction is fully utilized to realize the training of unlabeled samples.

[0106] During the training process, the neural network will learn to extract useful features, and different neural networks will extract corresponding features. The Unet neural network is used to construct the semantic segmentation task of seismic data, and the semantic features of different scales in the seismic data are extracted; the FCN neural network is used to construct the target detection task of seismic data, and the target features of different scales in the seismic data are extracted; the deeplab neural network is used to construct the image restoration task of seismic data, and the image restoration features of different scales in the seismic data are extracted. Among them, semantic features refer to the semantics of the image expressed by different combinations of features in seismic data, such as strata, faults, etc.; target features refer to the target locations expressed by different combinations of features in seismic data, such as the location of strata; image restoration features refer to the feature combinations required for image restoration.

[0107] During the training process, multiple computer vision algorithm models will output multiple groups of reconstructed sections, compare the reconstructed sections with the corresponding original sections, obtain the reconstruction errors, adjust the weights of the multiple computer vision algorithm models according to the loss function, so that the reconstructed sections output next time are closer to the original sections, repeat the above process until the reconstruction errors meet the preset conditions, and obtain multiple base models corresponding to the multiple computer vision algorithm models. The preset conditions are set according to the actual situation.

[0108] In addition, during the training process, parameter adjustment and training optimization are performed for each computer vision algorithm model. The learning rate is adjusted by setting the initial learning rate, learning rate decay strategy and using an adaptive learning rate method. The complexity of the model is constrained by adding regularization terms. The adjustment of regularization parameters can be determined by cross-validation and other methods to balance the complexity and generalization ability of the model. The optimal training batch size is selected through experiments and the dynamic batch size method is used to adjust the batch size. The initialization of the model is adjusted by selecting an appropriate weight initialization method based on the network structure.

[0109] The specific implementation of the above parameter adjustment and training optimization can refer to the detailed description in the prior art and is not specifically limited here.

[0110] In a specific embodiment, the training process of each computer vision algorithm model, that is, the construction process of the self-supervised learning task, is as follows: Figure 5As shown in the figure, it is divided into two parts: training data construction and self-supervised model training. In the process of training data construction, multiple groups of profile data, i.e., multiple groups of original profiles, are obtained by extracting profiles from the preprocessed seismic data set, i.e., labeled seismic data. Then, the multiple groups of profile data are randomly masked to obtain multiple groups of masked profile samples. In the process of self-supervised model training, the multiple groups of masked profile data are respectively input into multiple computer vision algorithm models for training. The multiple computer vision algorithm models will output the model prediction results, i.e., output the reconstructed profiles. The reconstruction error loss of the reconstructed profiles and the corresponding original profiles is compared. The multiple computer vision algorithm models are updated according to the reconstruction error. The above training process is repeated until the reconstruction error meets the preset conditions, and multiple trained computer vision algorithm models, i.e., multiple base models, are obtained.

[0111] S105: The multiple base models are integrated based on the attention mechanism to construct a large earthquake model.

[0112] In the embodiment of the present invention, the method of fusing the multiple base models based on the attention mechanism to construct a large earthquake model further includes:

[0113] A fusion layer is constructed based on the attention mechanism to fuse the features of different dimensions of the multiple base models to construct a large earthquake model.

[0114] In the embodiment of the present invention, after obtaining multiple base models, namely, the Unet base model, the FCN base model and the deeplab base model, the semantic features of different scales in the seismic data extracted by the Unet base model, the target features of different scales in the seismic data extracted by the FCN base model and the image restoration features of different scales in the seismic data extracted by the deeplab base model are also obtained, and the extracted features of different dimensions are fused, and the fusion layer is constructed using the attention mechanism to construct a large seismic model. The features of different dimensions are weighted and summed by the attention mechanism, and the features of different dimensions are fused to obtain a more comprehensive and accurate feature representation.

[0115] In a specific embodiment, the seismic large model obtained by the seismic large model construction method based on computer vision algorithm is applied to the identification process of reservoir fractures and caves in the Tarim Oilfield. It is necessary to add a decoding network to form a prediction model of fractures and caves together with the constructed seismic large model. The prediction model can be trained in the following way:

[0116] 1. Extract the profile data from the pre-obtained seismic data of the annotated fracture-cavity body to obtain multiple groups of profile data.

[0117] Collect the original seismic data obtained from field exploration, interpret and annotate the fractures and caves in the original seismic data, and prepare for subsequent sample construction.

[0118] Preprocess the seismic data, including: segy seismic data analysis, segy seismic data effective range interception, format conversion and data normalization. Since seismic data is generally saved in segy format, it is necessary to parse the seismic data before using it. According to the header information in the segy file, define the byte number of Inline and cdp, and extract the seismic trace data within the control layer range for subsequent sample construction.

[0119] After obtaining the preprocessed seismic data of the annotated fracture-cavity body, it is necessary to extract the profile data. Taking the fracture-cavity body position as the center, the seismic data profile is extracted along multiple directions such as inline, crossline and arbitrary line to obtain multiple groups of profile data. Among them, each group of profile data has 128 sampling points in the horizontal and vertical directions respectively, and each sampling point has a value.

[0120] 2. Determine the label data corresponding to the multiple groups of profile data and the corresponding standardized profile data according to a preset method.

[0121] For each set of profile data, obtain the label data corresponding to the profile group. The label data has the same size as the profile data. The label data is divided into two parts: fracture body position prediction label and fracture body numerical regression label. The corresponding label is constructed in a preset manner, thereby obtaining the label data corresponding to each set of profile data. Among them, the preset method for constructing the fracture body position prediction label can be based on the fracture body marked in the profile data, the value at the fracture body is recorded as 1, and the values ​​of other positions are recorded as 0; the preset method for constructing the fracture body numerical regression label can be normalized according to the maximum or minimum value in the numerical range statistically calculated in the profile data, so that the value changes continuously from 0 to 1, and the values ​​in the invalid range are set to -1 and do not participate in model training.

[0122] The label data obtained in the above way can be divided into two categories: one is the reflection boundary of the fracture-cavity reservoir, that is, the fracture-cavity position label; the other is the specific wave impedance value of the fracture-cavity body, that is, the fracture-cavity body numerical regression label.

[0123] When processing multiple sets of profile data, the extracted slice data should also be standardized according to the mean and standard deviation of the seismic trace amplitude.

[0124] The label data processed by the above process is a 128×128 data matrix, in which the data consists of 0 and 1; the seismic profile data corresponding to the label data is a 128×128 data matrix, in which the data consists of standardized seismic amplitude values.

[0125] 3. Divide multiple sets of labeled data and corresponding multiple sets of standardized profile data according to preset proportions to obtain training sets and validation sets.

[0126] The obtained multiple sets of labeled data and the corresponding multiple sets of standardized profile data are divided into training sets and validation sets according to the preset ratio. The preset ratio is set according to the actual situation, for example, it can be set to 80%:20% for training set: validation set, and a test set can be further divided to conduct a final evaluation of the model performance.

[0127] 4. Input the training set into a pre-built prediction model for training, and obtain multiple feature maps of different resolutions through the earthquake large model; the pre-built prediction model includes the earthquake large model and a decoding network.

[0128] The prediction model uses a multi-task deep convolutional neural network based on the UNet structure. The neural network consists of two parts: an encoding network (i.e., a large earthquake model) and a decoding network. The encoding part is a multi-resolution encoding network based on a large earthquake model, which extracts multi-resolution high-dimensional semantic features of earthquake data through a pre-trained network. When the training set is input into the pre-built prediction model for training, the sample data in the training set will first pass through the large earthquake model part of the prediction model for feature extraction to obtain feature maps of multiple different resolutions.

[0129] 5. Input multiple feature maps of different resolutions into the decoding network to obtain the position and numerical prediction results of the fracture body and the trained prediction model.

[0130] During the prediction model training process, multiple feature maps of different resolutions output from the large earthquake model are input into the decoding network. The decoding network predicts the location and value of the fracture-cavity body through the nonlinear combination of the features of each part. The structure of the decoding network is as follows: Figure 6 As shown in Figure 1, firstly, an atrous spatial pyramid pooling (ASPP) structure is used to increase the receptive field of the feature map. In the remaining structure, one part is composed of multiple cascaded upsampling modules. The structure of the upsampling module is as follows: Figure 7 As shown in the figure, it includes upsampling layer, convolution layer, normalization layer and activation layer, which contains bilinear interpolation and residual structure, the convolution kernel size is 3×3, and the activation function is GELU function; one part is the prediction end, including a 3×3 convolution module, two 1×1 convolution layers and sigmoid activation layer. The convolution kernel size of 1*1 and sigmoid activation layer are used in the prediction of the fracture position, and only one convolution kernel size of 1×1 is used for the fracture numerical regression. Through the decoding network, the semantic features are decoded into the spatial position of the fracture in the data and the corresponding wave impedance value.

[0131] In the embodiment of the present invention, Figure 8As shown in the figure, the training process of the prediction model is as follows: the training set constructed by seismic data slices is input into the prediction model, feature extraction is performed through the large seismic model, and then the fracture-cavity prediction is performed through the decoding network to obtain the fracture-cavity prediction result.

[0132] 6. Update the trained prediction model according to the location of the fracture body and the prediction accuracy of the numerical prediction result.

[0133] The model training process is implemented by fitting the model output and annotation on the training set, and the validation set data is used to evaluate the generalization performance of the model on unknown data and optimize the model hyperparameters.

[0134] After the training set is input into the prediction model for training, the position and numerical prediction results of the fracture body and the trained prediction model are obtained, the position and numerical prediction results of the fracture body are compared with the actual results, the prediction accuracy of the trained prediction model is determined, and the trained prediction model is updated.

[0135] The prediction model is continuously iterated and trained using the training set and the validation set. After each training, the trained model is evaluated for model accuracy on the validation set. Appropriate model hyperparameters are selected for iterative training based on the prediction accuracy. The above training process is repeated until a trained prediction model that meets the preset conditions is obtained. When the model training converges, the test set can also be used to perform a final evaluation of the model performance. By performing model prediction on the test set and comparing it with the labeled data, the prediction accuracy and generalization performance of the model can be quantitatively evaluated.

[0136] 7. Obtain a trained prediction model that meets the preset conditions.

[0137] The prediction model based on the large seismic model is trained in the above way, and the trained prediction model is used for the prediction of fracture-cavity bodies. It is applied to the identification of reservoir fractures and caves in the Tarim Oilfield. The original seismic data of the actual test are as follows: Fig. 9 As shown in the figure, the original seismic data is input into the trained prediction model after amplitude normalization to predict the position of the fracture-cavity body and the numerical prediction results, that is, the fracture-cavity body reflection boundary prediction results and the regression wave impedance prediction results are as shown in Fig.10 The fracture-cavity prediction results are overlapped with the original seismic data, as shown in Fig.11 As shown in the figure, the background part is the original seismic data, and the superposition range is the predicted fracture-cavity reflection boundary position. The overall result obtained by the trained prediction model is 92.4% consistent with the expert interpretation, which meets the business application requirements, and the timeliness of intelligent identification is significantly improved. Without reducing or even improving the quality of interpretation, the number of samples required is reduced by more than 8 times, and the production efficiency is increased by more than 30 times.

[0138] The method for constructing a large earthquake model based on a computer vision algorithm provided by an embodiment of the present invention improves the diversity of training data by sampling earthquake data and extracting random sections along different inlines, crosslines and arbitrary tangent directions; masks earthquake data by drawing on the image processing method, and masks the multiple groups of profile data by randomly selecting at least one preset method to construct multiple groups of masked profile samples; uses different computer vision algorithms to extract features of different dimensions in earthquake data, and fuses the features based on the attention mechanism, so that the constructed large earthquake model can obtain a more comprehensive and accurate feature representation. This method constructs a large earthquake model with high calculation precision, good accuracy, strong generalization ability, stronger robustness and strong interpretability based on computer vision algorithms, improves the accuracy and timeliness of intelligent identification of earthquake profiles, and breaks through a series of problems such as the fragmentation of current intelligent applications, model customization and AI threshold.

[0139] Embodiment 2

[0140] Based on the same inventive concept, the embodiment of the present invention also provides a large earthquake model construction device based on computer vision algorithm, referring to Fig.12 As shown, the device comprises:

[0141] An acquisition module 101 is used to acquire labeled earthquake data;

[0142] An extraction module 102 is used to extract profile data from the labeled seismic data to obtain multiple groups of profile data;

[0143] A masking module 103 is used to randomly select at least one preset method to mask the multiple groups of profile data to construct multiple groups of masked profile samples;

[0144] A self-supervised learning module 104 is used to input the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training based on self-supervised learning to obtain a plurality of base models;

[0145] The model building module 105 is used to fuse the multiple base models based on the attention mechanism to build a large earthquake model.

[0146] Embodiment 3

[0147] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method for constructing a large earthquake model based on a computer vision algorithm as described in the above embodiment 1 is implemented.

[0148] Embodiment 4

[0149] Based on the same inventive concept, an embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for constructing a large earthquake model based on a computer vision algorithm as described in the first embodiment above is implemented.

[0150] Embodiment 5

[0151] Based on the same inventive concept, an embodiment of the present invention further provides a computer program product comprising instructions. When the computer program product is run on a computer device, the computer device executes the method for constructing a large earthquake model based on a computer vision algorithm as described in the first embodiment above.

[0152] Embodiment 6

[0153] Based on the same inventive concept, an embodiment of the present invention also provides a chip, the chip includes a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run a computer program or instruction to implement the method for constructing a large earthquake model based on a computer vision algorithm as described in the above embodiment one.

[0154] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0155] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0156] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0157] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0158] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for constructing a large earthquake model based on computer vision algorithm, characterized in that: include: Get labeled earthquake data; Extracting profile data from the labeled seismic data to obtain multiple groups of profile data; Randomly selecting at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples; Based on self-supervised learning, the plurality of groups of mask profile samples are respectively input into a plurality of pre-built computer vision algorithm models for training to obtain a plurality of base models; The multiple base models are fused based on the attention mechanism to construct a large earthquake model.

2. The method according to claim 1, characterized in that Before labeling the seismic data to obtain the labeled seismic data, the method further includes: Perform data statistics on multiple sets of seismic data obtained, clean abnormal data based on the normal distribution method, and obtain cleaned seismic data; By data normalization and standardization, the cleaned seismic data is mapped into a preset range to obtain preprocessed seismic data; The preprocessed seismic data are labeled to obtain corresponding labeled seismic data.

3. The method according to claim 2, characterized in that The seismic data includes pre-stack seismic data and post-stack seismic data; The preprocessed seismic data is labeled to obtain corresponding labeled seismic data, including: For the prestack seismic data, high-quality reservoir position analysis is performed according to the relationship between amplitude and offset distance to obtain labeled prestack seismic data; For the post-stack seismic data, the structural characteristics of the target layer are obtained according to the lateral and vertical changes of the amplitude, and reservoir prediction is performed to obtain labeled post-stack seismic data.

4. The method according to claim 3, characterized in that The extracting of the profile data from the labeled seismic data to obtain multiple groups of profile data includes: The profile data are extracted along different inline, crossline and tangent directions of the labeled pre-stack seismic data and / or the labeled post-stack seismic data to obtain multiple groups of profile data.

5. The method according to claim 1, characterized in that The randomly selecting at least one preset method to mask the multiple groups of profile data to construct multiple groups of mask profile samples includes: Randomly selecting different preset masking ratios, masking the profile data, and constructing multiple groups of mask profile samples; The profile data is randomly divided into a plurality of regions, and for the data in each region, different preset masking area blocks are randomly selected for masking, so as to construct a plurality of groups of mask profile samples; Different values ​​are randomly selected from the profile data, and the values ​​in the profile data are randomly masked with the randomly selected different values ​​to construct multiple groups of mask profile samples.

6. The method according to claim 1, characterized in that Based on the self-supervised learning, the plurality of groups of mask profile samples are respectively input into a plurality of pre-built computer vision algorithm models for training to obtain a plurality of base models, including: Inputting the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training respectively, to obtain a plurality of groups of reconstructed profiles; the computer vision algorithms include Unet, FCN and Deeplab computer vision algorithms; Comparing the plurality of groups of mask profile samples with the corresponding reconstructed profiles respectively to obtain reconstruction errors; Based on the reconstruction error, adjusting the weight of the corresponding computer vision algorithm model; The above process is repeated until the reconstruction error meets the preset conditions, and multiple base models corresponding to multiple computer vision algorithm models are obtained.

7. The method according to claim 1, characterized in that The method of fusing the multiple base models based on the attention mechanism to construct a large earthquake model also includes: A fusion layer is constructed based on the attention mechanism to fuse the features of different dimensions of the multiple base models to construct a large earthquake model.

8. A device for constructing a large earthquake model based on computer vision algorithm, characterized in that: include: An acquisition module, used to acquire labeled earthquake data; An extraction module, used for extracting profile data from the labeled seismic data to obtain multiple groups of profile data; A masking module, used for randomly selecting at least one preset method to mask the multiple groups of profile data, and constructing multiple groups of masked profile samples; A self-supervised learning module, for inputting the plurality of groups of mask profile samples into a plurality of pre-built computer vision algorithm models for training based on self-supervised learning to obtain a plurality of base models; The model building module is used to fuse the multiple base models based on the attention mechanism to build a large earthquake model.

9. A computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal, the terminal executes the method for constructing a large earthquake model based on a computer vision algorithm as described in any one of claims 1 to 7.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for constructing a large earthquake model based on a computer vision algorithm as described in any one of claims 1 to 7 is implemented.

11. A computer program product comprising instructions, which, when executed on a computer device, enables the computer device to execute the method for constructing a large earthquake model based on a computer vision algorithm as claimed in any one of claims 1 to 7.

12. A chip, comprising a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run a computer program or instruction to implement the method for constructing a large earthquake model based on a computer vision algorithm as described in any one of claims 1 to 7.

Images