Fracture identification method and device based on deep learning of hybrid neural network model

Abstract

The invention provides a fracture identification method and device based on deep learning of a hybrid neural network model. According to the invention, the mixed neural network model is adopted to learn the noise characteristics of the seismic data, the random noise in the seismic data is eliminated, and the signal-to-noise ratio of the seismic data is improved; on this basis, using a hybrid neural network model to carry out fracture feature extraction on the seismic data to obtain a fault probability data volume; artificial fault interpretation is utilized to train a hybrid neural network model through transfer learning, the generalization ability and applicability of the model are improved, the trained hybrid neural network model is applied to actual seismic data for fracture feature extraction, fault recognition effects extracted before and after model training are compared, and the fault recognition accuracy is improved. Artificial interpretation fault samples are added until the fault extraction effect of the trained hybrid neural network model is finer and more accords with geological laws; and finally, a fault probability refined data body is obtained, and a set of good technical scheme and thought are provided for low-level sequence fracture identification.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas geophysics, and more specifically, to a fracture identification method and device based on deep learning of a hybrid neural network model. Background Art

[0002] Fractures and cracks are widely developed in oil and gas reservoirs, forming oil and gas accumulation spaces and migration channels, and are important factors affecting the reservoir storage capacity and development potential. Accurate detection of fractures and cracks can effectively guide well location deployment and fracturing construction.

[0003] Fractures and cracks are gap and fissure structures generated in the formation due to stress, and they are widely developed in oil and gas reservoirs. On the one hand, the existence of fractures and cracks provides spaces for oil and gas accumulation and convergence. Sandstones and carbonates with brittle lithologies are prone to form fractures and cracks under the action of fracture lithology differences. On the other hand, fractures and cracks constitute the main channels for oil and gas migration inside the reservoir and between it and other surrounding lithologies. When oil and gas encounter a fracture interface during migration, capillary action will occur and they will adhere and accumulate. Therefore, fractures and cracks not only provide accumulation spaces but also form migration channels, and are key geological factors controlling oil and gas enrichment and distribution.

[0004] Since fractures and cracks have an important impact on oil and gas accumulation and migration, accurately identifying and depicting the spatial distribution of fractures and cracks in the reservoir has a direct guiding effect on analyzing the reservoir storage performance and production potential, and guiding reasonable well location arrangement and scientific fracturing construction. For example, through fracture detection, the permeability distribution of the reservoir can be judged, which can be used as a basis for well location selection and directional drilling; by judging the characteristic parameters of the dominant fractures, it can be used for the optimized design of fracturing parameters, making the fracturing more in line with the formation conditions, making full use of natural fractures, and improving fracturing efficiency. Therefore, high-precision and high-resolution fracture and crack identification is of great significance for reservoir evaluation and subsequent development.

[0005] Currently, there are many methods for fracture and crack detection based on seismic data, and mainly post-stack fracture and crack detection methods. The basic principle of the post-stack method is to use mathematical methods such as spatial scanning of adjacent seismic traces and similarity analysis to detect the changes in seismic reflection waveforms, amplitudes, phases, etc. caused by the development of fractures, so as to realize the description of the spatial distribution of fractures and cracks. Specific mathematical methods include the small window method, multi-path similarity, condition number method, differential curvature, etc. These methods are collectively referred to as post-stack geometric attributes, including coherence, curvature, edge detection, etc. Their characteristics are high anti-noise performance, high efficiency, easy implementation, and can better realize the identification of large-scale fractures. However, for seismic data with low signal-to-noise ratio and low-order fractures, the identification effect is average.

[0006] In addition to traditional post-stack data analysis, with the development of seismic acquisition technology, the pre-stack original full-waveform seismic data also contains rich information such as the offset distance between the source and receiver and the azimuth of seismic wave propagation. Based on the wave characteristics of these pre-stack data, the anisotropic effect of seismic waves passing through fractures can be analyzed more deeply to achieve the prediction of small-scale fracture structures at low-order levels. For example, high-angle fractures have a direction-dependent impact on the seismic wave propagation speed and arrival time, which is called wave anisotropy. By collecting seismic data with multi-azimuth, wide-band, and wide-offset, and combining signal processing methods such as elliptic function fitting, the anisotropic parameters of fractures can be extracted to predict the structural characteristics such as the fracture strike, dip angle, and density at specific locations. This anisotropic prediction can identify finer low-order fractures, but it has a large amount of calculation and a long cycle, making it difficult to meet the high-efficiency production requirements.

[0007] In summary, in current fracture and crack identification technologies, there are certain limitations in identifying low-order fractures for various methods. Both pre-stack and post-stack data can detect fracture information, but with different accuracies and resolutions; relatively speaking, post-stack attribute calculation is efficient, while pre-stack wave anisotropy analysis is more refined but computationally complex.

[0008] Therefore, there is a great need for a technical solution that can combine the advantages of different types of data, is computationally efficient, and can identify low-order fine fractures to more comprehensively and accurately predict the spatial distribution of reservoir fractures. Summary of the Invention

[0009] In view of this, the present invention discloses a fracture identification solution based on deep learning of a hybrid neural network model, which can significantly improve the identification accuracy of low-order fractures and save a large amount of calculation time.

[0010] According to one aspect of the present invention, a fracture identification method based on deep learning of a hybrid neural network model is proposed, and the method includes:

[0011] Step 1, prepare a pre-stack depth migration seismic data volume and fault interpretation data;

[0012] Step 2, use the first neural network model in the hybrid neural network model to process the pre-stack depth migration seismic data volume to eliminate noise and output a denoised seismic data volume;

[0013] Step 3, use the second neural network model in the hybrid neural network model to process the denoised seismic data volume and output a fault probability data volume;

[0014] Step 4, use the fault interpretation data to train the second neural network model through transfer learning;

[0015] Step 5: Process the denoised seismic data volume using the second neural network model trained in Step 4, and compare the calculated fault probability data volume with the fault probability data volume obtained in Step 3. If the fault identification effect of the calculated fault probability data volume is better than that of the fault probability data volume obtained in Step 3, end Step 5; otherwise, further increase the artificial fault interpretation data samples, and then return to Step 4;

[0016] Step 6: Refine the fault probability data volume calculated in Step 5 to obtain the final refined fault probability data volume.

[0017] In some embodiments, the first neural network model is a deep neural network model based on a residual network model.

[0018] In some embodiments, the second neural network model is a deep neural network model based on a U-Net model.

[0019] In some embodiments, Step 6 specifically includes:

[0020] Within the time window of the target layer, set a fault probability threshold value, and use a voting mechanism to perform weighted averaging on the probability values in the fault probability data volume calculated in Step 5 that are higher than the fault probability threshold value to obtain the final refined fault probability data volume.

[0021] In some embodiments, the method further includes:

[0022] Train the first neural network model in the hybrid neural network model using the seismic data noise training set to learn the noise characteristics in order to eliminate the noise in the seismic data. The seismic data noise training set includes synthetic data and actual data.

[0023] According to another aspect of the present invention, a fracture identification device based on deep learning of a hybrid neural network model is also proposed. The device includes:

[0024] A training set preparation unit for preparing a pre-stack depth migration seismic data volume and fault interpretation data;

[0025] A seismic data denoising unit for processing the pre-stack depth migration seismic data volume using the first neural network model in the hybrid neural network model to eliminate noise and output the denoised seismic data volume;

[0026] A fault probability identification unit for processing the denoised seismic data volume using the second neural network model in the hybrid neural network model and outputting a fault probability data volume;

[0027] A transfer training unit for training the second neural network model using fault interpretation data through transfer learning;

[0028] A fracture recognition iteration unit, which is used to process the denoised seismic data volume by using the second neural network model trained in step 4, and compare the calculated fault probability data volume with the fault probability data volume obtained in step 3. If the fault recognition effect of the calculated fault probability data volume is better than that of the fault probability data volume obtained in step 3, end step 5; otherwise, further increase the artificial fault interpretation data samples, and then return to step 4;

[0029] A refinement processing unit, which is used to refine the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

[0030] In some embodiments, the first neural network model is a deep neural network model based on a residual network model.

[0031] In some embodiments, the second neural network model is a deep neural network model based on a U-Net model.

[0032] In some embodiments, the refinement processing unit specifically is used for:

[0033] Within the time window of the target layer, set a fault probability threshold value, and use a voting mechanism to perform weighted averaging on the probability values higher than the fault probability threshold value in the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

[0034] According to another aspect of the present invention, an electronic device is further provided. The electronic device includes:

[0035] A memory, which stores executable instructions;

[0036] A processor, which runs the executable instructions in the memory to implement the fracture recognition method based on deep learning of the hybrid neural network model described above.

[0037] According to another aspect of the present invention, a computer-readable storage medium is further provided. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the fracture recognition method based on deep learning of the hybrid neural network model described above.

[0038] The present invention proposes a technical solution for identifying low-order fractures based on deep learning of a hybrid neural network model. Each embodiment and implementation manner of the present invention has the following advantages:

[0039] 1. Improve the recognition accuracy of low-order fractures

[0040] The method based on the hybrid neural network model proposed by the present invention can automatically and efficiently extract complex low-order fault features from seismic data through deep learning technology. The recognition effect is significantly better than that of the traditional post-stack geometric attribute technology, solving the problem of low-order fault recognition;

[0041] 2. Saving a large amount of computing time

[0042] Compared with the traditional prestack wave motion anisotropy analysis method, the present invention is based on deep learning modeling, avoiding a large amount of manual feature engineering and complex operations, with high computing efficiency, greatly saving the workload and meeting the high-efficiency production requirements;

[0043] 3. Improving applicability

[0044] Through transfer learning and model iterative training technology, and continuously increasing artificial fault interpretation data during iteration, the present invention can continuously improve the model, adapt to the geological conditions of different regions, and improve the applicability of the method;

[0045] 4. Improving the recognition accuracy effect of faults on the inline, crossline and time slice

[0046] Adopting a voting mechanism, the prediction results of faults with a relatively high probability of meeting the set fault probability threshold value within the time window range of the target layer are weighted and averaged to obtain a fault probability refined data volume;

[0047] 5. Automatically mining data features

[0048] Through the deep learning network model for representation learning of data samples, automatically mining data features, no longer requiring manual calculation of attributes, and realizing the mining of fracture-fracture information based on the depth of the convolutional neural network model;

[0049] 6. Improving the ability of the model to solve complex problems

[0050] A hybrid neural network model is formed by ResNet (residual network model) and U-Net. U-Net takes into account both the underlying and high-level information of data features, while ResNet alleviates the problem of model degradation, deepens the number of layers of the network model, and significantly improves the ability of the network model to solve complex problems;

[0051] 7. Providing support for development

[0052] High-precision and high-efficiency low-order fault recognition can clarify the microfracture structure, accurately judge the formation permeability, and provide strong technical support for subsequent development activities such as fracturing parameter optimization and well location selection;

[0053] 8. Promotion and application prospects

[0054] The technical route of the present invention is novel and the effect is remarkable. It is expected to be popularized and applied in the oil and gas field, extended to the extraction of geological body characteristics of more types, and has good application prospects and popularization value.

[0055] In summary, the present invention has obvious innovations and breakthroughs in the technical route of low-order fault identification, with excellent effects, bringing many direct and potential economic benefits, and is a high-level original technical achievement.

[0056] The method and device of the present invention have other characteristics and advantages, which will be obvious in the accompanying drawings and subsequent specific embodiments incorporated herein, or will be described in detail in the accompanying drawings and subsequent specific embodiments incorporated herein. These accompanying drawings and specific embodiments are jointly used to explain the specific principles of the present invention. Brief Description of the Drawings

[0057] By describing the exemplary embodiments of the present invention in more detail in conjunction with the accompanying drawings, the above and other objects, features and advantages of the present invention will become more obvious. Among them, in the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.

[0058] Figure 1 The flowchart of a fault identification method based on deep learning of a hybrid neural network model according to an embodiment of the present invention is shown.

[0059] Figure 2 (a) and (b) show the time slice attributes of the target layer before and after fault enhancement processing of seismic data in a certain work area in the depth domain.

[0060] Figure 3 (a) and (b) show the comparison diagram of the high-precision coherence attribute of the target layer in a certain work area and the fault probability volume attribute extracted according to the embodiment of the present invention.

[0061] Figure 4 (a) and (b) show the fault probability volume refinement data body attribute of the target layer obtained after the fault probability volume refinement processing of the fault probability volume attribute extracted by the hybrid neural network model of the present invention for the target layer in a certain work area and further refinement processing.

[0062] Figure 5 (a) and (b) show the fault probability volume refinement attribute extracted by the hybrid neural network model of the present invention for the target layer in a certain work area and the fault probability volume attribute of the fault probability volume extracted by the hybrid neural network model after transfer learning training according to the present invention. Detailed Description of the Invention

[0063] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0064] The following introduces the conceptual background of the present invention.

[0065] The inventors have deeply considered the defects existing in the current fracture-crack detection methods. The data-driven machine learning method can provide a better solution. Deep learning is a type of machine learning method with a relatively deep layer based on the artificial neural network model. With the continuous enhancement of hardware computing power and the rise of big data, more and more deep learning algorithms have been applied to the field of seismic data interpretation. Deep learning can perform representation learning on data samples, automatically mine data features, and no longer require manual calculation of attributes. Based on the convolutional neural network model, deep learning is used to mine fracture-crack information.

[0066] As described above, the current fracture-crack detection technologies, whether it is the traditional pre-stack and post-stack data analysis methods or the wave anisotropy technology, have certain limitations. The recognition effect of small-scale low-order fractures in low-signal-to-noise ratio data in complex environments is poor, and it cannot meet the need for high-precision feature extraction. Considering this contradiction, the data-driven machine learning or deep learning technology can provide a new and effective way.

[0067] Deep learning is a new research hotspot in the field of machine learning. It establishes a multi-layer neural network model for feature learning by simulating the network model structure of the human brain to analyze and process information. Compared with traditional machine learning, deep learning can complete end-to-end data modeling, without the need for manual feature extraction and rules, and gradually learns the high-order abstract feature representation in the data through hierarchical representation, with strong autonomous analysis and simulation capabilities.

[0068] In recent years, relying on the accumulation of big data and the improvement of computing power, deep learning algorithms have made great progress and achieved success in fields such as image recognition and natural language processing. Seismic data, as a typical multi-dimensional spatio-temporal sequence signal, can also be modeled end-to-end through a deep neural network model to carry out automatic learning of formation features. This provides a new idea for solving the dilemmas faced by the current methods.

[0069] The essence of seismic data is the record of fluctuations in the movement of clastic rock particles in the strata. This dense microscopic information is hidden in complex random noise, and it is difficult to fully utilize this potential information through manually extracted geological attributes. Deep learning can automatically abstract patterns in seismic data through representation learning and perform feature representation. This data-driven mining of deeper features can identify more subtle stratum changes and is particularly suitable for the detection of low-order geological bodies such as fine fractures.

[0070] The convolutional neural network model is a typical deep network model structure that can capture spatiotemporal features. Through convolutional layer cascading and pooling downsampling, the more layers the network model has, the more abstract and complex the data features that can be extracted. The convolutional neural network model has been successful in processing time series data such as images and videos, and can also be transferred to seismic data interpretation tasks. Through end-to-end training, automatic feature learning and classification can be achieved. This provides a new technical path for low-cost and efficient fracture prediction.

[0071] In summary, the inventors think that deep learning has opened up a new perspective of data-driven intelligent feature mining. The seismic field can also use its technical methods to break through the limitations of current geological body identification. It is especially suitable for the detection of small-scale geological bodies such as low-order fine cracks, and is an important breakthrough for improving the accuracy of seismic data interpretation in the future. Combined with big data accumulation and model integration technology, deep learning will surely have a broader application prospect.

[0072] Example 1

[0073] Figure 1 A flow chart of fracture identification based on deep learning of a hybrid neural network model according to an embodiment of the present invention is shown. As shown in the figure, the method includes steps 1 to 6.

[0074] Step 1: Prepare prestack depth migration seismic data volume and fault interpretation data.

[0075] Prestack depth migration seismic data volume and fault interpretation data are two main types of data used in the present invention, which are briefly introduced respectively.

[0076] Prestack depth migration seismic data is a raw full-waveform seismic data, which is formed by seismic data processing. It not only records the amplitude and frequency information of the seismic signal, but also contains the offset distance information of the source-receiver and the azimuth information of the signal propagation path. These additional geometric information can be used to analyze the anisotropic effect of seismic wave propagation and improve the resolution of stratigraphic details.

[0077] Fault interpretation data refers to the interpretive result data of the fault structure and properties in a certain area proposed by experts based on geological principles. It is derived from the interpretation of data such as drilling data and reflects geological information such as the location, type, and movement direction of faults in this area. These interpretation data containing artificial prior knowledge can be used to transcend the limitations of seismic data itself and enhance the accuracy of the results.

[0078] The significance of using these two types of data in the present invention is that the pre-stack depth migration seismic data volume, as the original full-waveform data, records rich formation information, while the fault interpretation data provides prior constraint knowledge on the fault distribution in this area. The combination of the two can achieve the organic integration from data-driven to knowledge-driven, making the final model result conform to both the data itself and geological principles, ensuring the accuracy of the method.

[0079] Step 2: Process the pre-stack depth migration seismic data volume using the first neural network model in the hybrid neural network model to eliminate noise and output the denoised seismic data volume.

[0080] In some embodiments, the method further includes training the first neural network model in the hybrid neural network model using the seismic data noise training set to learn the noise characteristics in order to eliminate the noise in the seismic data. The seismic data noise training set includes synthetic data and actual data. High signal-to-noise ratio seismic data is beneficial for achieving the purpose of enhancing fracture feature processing subsequently.

[0081] The use of synthetic data and actual data to form a training set in the present invention has the following advantages:

[0082] 1. Expand the scale of the training set

[0083] Often, the number of actual data samples alone is insufficient. By generating synthetic data, the training set can be greatly expanded to provide richer training samples;

[0084] 2. Improve the generalization of the model

[0085] Synthetic data can simulate more complex and boundary situations, enabling the model to be exposed to more diverse data distributions and enhancing the model's generalization and adaptation ability;

[0086] 3. Enhance the robustness of the model

[0087] Synthetic data with added noise and simulation errors can increase the perturbation range of the data, making the model more robust to more complex input situations and having a higher error tolerance rate;

[0088] 4. Provide an infinite number of samples

[0089] The synthetic data algorithm can automatically generate a large number of training samples, providing an infinite sample source and alleviating the problem of insufficient data;

[0090] 5. Reduce data bias

[0091] Synthetic data can reduce the selection bias of data by controlling the generation process, generate a more balanced and comprehensive sample distribution, and reduce overfitting;

[0092] 6. Reduce label cost

[0093] Synthetic data is directly generated from the physical model without the need for cumbersome manual annotation work, greatly reducing the labor cost of constructing the training set.

[0094] In summary, this embodiment uses a training set composed of synthetic data and actual data to train the deep learning network model, which can greatly improve the generalization, robustness and applicability of the model.

[0095] In some embodiments, the first neural network model is a deep neural network model based on the Residual Network (ResNet) model.

[0096] The Residual Network (ResNet) is a deep neural network model structure, mainly characterized by introducing residual (shortcut) connections in the neural network model. The Residual Network model can alleviate the problem of model degradation, deepen the number of layers of the network model, and significantly improve the ability of the network model to solve complex problems.

[0097] The residual connection realizes a direct signal propagation path across several layers, so that the feature information from the bottom layer can directly reach the higher layer of the network model, effectively alleviating the problems of gradient disappearance and gradient explosion. This enables the Residual Network model to stack more than 100 or even 1000 layers, achieving an unprecedented deepening effect of the network model.

[0098] Compared with the ordinary convolutional network model, the modeling error of the residual structure is lower and the accuracy is higher. The present invention uses a deep learning network model based on the Residual Network model to denoise seismic data, which is beneficial to obtaining the best denoising effect. In addition, the parameters and computational amount of the Residual Network model do not increase significantly, and the model is also easy to optimize.

[0099] The deep learning network model based on the Residual Network model is used in the hybrid neural network model of this embodiment to realize denoising of the pre-stack depth migration seismic data volume, which has the following advantages:

[0100] 1. Strong feature learning and expression ability

[0101] The Residual Network model structure is deep, and through stacking multiple layers of network models, hierarchical abstraction learning of multi-order features is realized, which can fit complex data mapping relationships and has strong learning and representation capabilities for the patterns of seismic data;

[0102] 2. End-to-End Denoising Joint Optimization

[0103] The residual network model realizes end-to-end modeling, with the input being noisy data and the output being denoised data. The network model parameters are jointly optimized through the loss function to maximize the denoising recovery quality without the need for manual design of denoising algorithms;

[0104] 3. Computationally Efficient

[0105] Despite the large depth of the network model, the residual structure avoids training difficulties and the number of parameters does not increase sharply. The calculation speed is fast, meeting the usage requirements of the production environment;

[0106] 4. Noise Adaptation

[0107] The network model is trained with a large number of noise samples and has strong adaptability to different types and complexities of noise, without the need for artificial setting of the denoising model.

[0108] Step 3: Use the second neural network model in the hybrid neural network model to process the denoised seismic data volume and output a fault probability data volume.

[0109] The hybrid neural network model extracts fracture features from the denoised data volume obtained in step 2, quickly and efficiently identifies the main fracture features, and obtains a fault probability data volume.

[0110] The second neural network model may include identification models for conventional fault types such as normal fault identification models, reverse fault identification models, horizontal moving fault identification models, strike-slip fault identification models, etc.

[0111] In some embodiments, the second neural network model is a deep neural network model based on the U-Net model.

[0112] U-Net is a convolutional neural network model structure for image semantic segmentation. U-Net can take into account both the underlying and high-level information of data features, improving the accuracy of fault identification. In this embodiment, the U-Net model is used to obtain the fault probability data volume, and the specific advantages are analyzed as follows:

[0113] 1. The U-shaped structure can retain spatial information

[0114] Through the encoder-decoder U-shaped structure of U-Net, spatial information can be retained while the image is downsampled, and then details can be restored during upsampling, which is very important for maintaining the spatial correspondence relationship of seismic data;

[0115] 2. Skip connections combine multi-scale information

[0116] Skip connections directly pass the encoded features to the decoding part, enabling the direct fusion of low-level and high-level information, combining features at different abstraction levels, and facilitating the learning of multi-scale fracture features;

[0117] 3. Reducing the number of parameters

[0118] U-Net significantly reduces the number of parameters through skip connections and parameter sharing mechanisms, and can be effectively trained with a small amount of labeled data;

[0119] 4. Good segmentation effect

[0120] U-Net is designed for image semantic segmentation, has a good classification effect at the pixel level, and can accurately extract the fracture probability information in seismic data;

[0121] 5. End-to-end training

[0122] U-Net realizes the end-to-end mapping from seismic data to fracture segmentation results. Through training, it can directly learn the fracture features in the data without artificial design and extraction of fracture attributes.

[0123] In summary, the inventor uses U-Net to obtain the fault probability data volume, which can maintain spatio-temporal features, is suitable for processing seismic data, and the prediction result presents a pixel-level delicate fracture probability distribution, which well meets the purpose of the present invention.

[0124] Step 4: Use the fault interpretation data to train the second neural network model through transfer learning.

[0125] Training the hybrid neural network model through transfer learning using the artificial fault interpretation results can improve the generalization ability and applicability of the model. The specific analysis of the beneficial effects of using the fault interpretation data to train the second neural network model through transfer learning is as follows:

[0126] 1. Improving the model generalization ability

[0127] The fault interpretation data provides more accurate and comprehensive geological knowledge. Transfer learning enables the model to absorb this knowledge for in-domain adaptation, thus having better adaptability to new regions;

[0128] 2. Enhancing the model robustness

[0129] The interpretation data is equivalent to adding perturbations to the original data, which expands the input distribution range that the model can adapt to and enhances the model's robustness;

[0130] 3. Introducing artificial prior constraints

[0131] Fault interpretation reflects human understanding of the geological laws in this area. Incorporating these constraints enables the model to combine data-driven and knowledge-driven approaches, enhancing the interpretability of the results;

[0132] 4. Improve the adaptability to small datasets

[0133] Through transfer learning, the performance of the target domain can be efficiently and rapidly improved using a small amount of labeled interpretive data, reducing data dependence;

[0134] 5. Reduce the annotation requirements for the target domain

[0135] Utilizing the annotation knowledge from the source domain eliminates the need to label a large amount of data from scratch in the new domain, alleviating the data processing pressure;

[0136] 6. Expand the applicable fields of the model

[0137] Transfer learning enables the model to be extended to more similar downstream tasks based on the source domain, reducing repetitive work and expanding the scope of model usage;

[0138] 7. Improve the model accuracy

[0139] Introducing interpretive data makes the model training approach the actual geological distribution, resulting in more accurate and reliable prediction results, which is beneficial for subsequent production dynamic management.

[0140] In step 5, the denoised seismic data volume is processed using the second neural network model trained in step 4, and the calculated fault probability data volume is compared with the fault probability data volume obtained in step 3. If the fault identification effect of the calculated fault probability data volume is better than that of the fault probability data volume obtained in step 3, then step 5 ends; otherwise, the artificial fault interpretation data samples are further increased, and then it returns to step 4.

[0141] In step 5, the fault recognition effect can be significantly improved through iterative calculation of the fault probability volume. Through iterative calculation, more artificial interpretation knowledge is introduced each time based on the previous round of prediction, gradually optimizing the model, thereby improving the fault recognition effect and being more accurate and efficient; closely combining data-driven and knowledge-driven, the original seismic data ensures the data-driven nature of the results, and the artificial interpretation knowledge provides problem constraints, effectively combining data-driven and knowledge-driven and giving full play to the advantages of both; and introducing an interactive verification mechanism, a closed-loop iterative verification is formed between the prediction results and artificial interpretation. The model prediction takes into account human verification, and human verification inherits the model basis and gradually tends to be consistent, making the results more accurate and credible; enabling the present invention to adapt to different geological environments. By continuously iteratively learning the geological knowledge of different regions, the model can continuously absorb this knowledge for adaptation, enhancing the ability to adapt to different regions; significantly improving the generalization ability of the model. Iterative learning continuously adds new knowledge, expanding the scope of model adaptation, contacting different geological environments, enhancing the generalization adaptability of the model, and having a wider application range, thus providing a basis for later recognition. The model and fault interpretation knowledge obtained through iterative calculation in step 5 lay a foundation for fault prediction in similar regions in the later stage, making subsequent work more efficient and accurate.

[0142] Step 6: Refine the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

[0143] In some embodiments, step 6 specifically includes:

[0144] Within the time window of the target layer, set a fault probability threshold value, and use a voting mechanism to perform weighted averaging on the probability values higher than the fault probability threshold value in the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

[0145] The weighting coefficient can be adjusted according to the prediction accuracy of the classifier.

[0146] Through the refinement process, the recognition accuracy effect of faults on the line, trace, and time slice can be improved, and a final refined fault probability data volume can be obtained.

[0147] The voting mechanism (Voting) is an ensemble learning method. Its main idea is:

[0148] 1. Construct multiple different models (also called classifiers), and let each model make a prediction or classification once;

[0149] 2. Statistically sum up or weight and sort the prediction results of these models, and vote to select the final prediction result.

[0150] This idea is similar to voting in an election. For a simple example:

[0151] Suppose it is predicted that a sample belongs to one of two categories, A (fault) or B (not a fault), and three classification models are constructed, where:

[0152] Model 1 predicts that the sample belongs to A

[0153] Model 2 predicts that the sample belongs to B

[0154] Model 3 predicts that the sample belongs to A.

[0155] Then finally, the prediction results of the three models can be integrated by the voting method.

[0156] This idea of integrating through multiple "voters" can improve the accuracy and stability of the final fault identification result. Some commonly used voting strategies include:

[0157] 1. Simple majority voting method

[0158] 2. Weighted voting method

[0159] 3. Soft voting method

[0160] In this embodiment, the weighted voting method is adopted, which can integrate the judgments of multiple models and obtain a better final prediction effect.

[0161] Specifically, a probability threshold value for fault prediction can be set first. In this embodiment, it is called the fault probability threshold value. Only when the predicted probability of a fault exceeds this fault probability threshold value is it considered a fault prediction; for those fault prediction results that meet the probability threshold (i.e., the prediction results that exceed the set fault probability threshold value), a weighted average operation is performed.

[0162] The weight coefficients in the weighted average are related to the accuracy / credibility of each prediction. The higher the prediction accuracy, the larger the weight coefficient.

[0163] After weighted average, a fault probability value that comprehensively considers each prediction result and its accuracy is finally obtained.

[0164] For example, a simple example is given for illustrative purposes.

[0165] Suppose the fault probability threshold is set at 60%. Then, if the predicted fault probability is lower than 60%, it will be excluded from consideration. Assume there are 3 predictions exceeding the fault probability threshold, at the positions of 70%, 80%, and 90% fault probabilities respectively. They all first meet the condition of being greater than 60%, and then weighted averaging is performed. Since the prediction with 90% has the highest accuracy, it is given the highest weight of 0.4; the 80% accuracy is the second highest, with a weight of 0.3; the 70% accuracy is the lowest, with a weight of 0.3. The weighted average calculation: (0.4 * 90% + 0.3 * 80% + 0.3 * 70%) / (0.4 + 0.3 + 0.3) = 81%. Then 81% is the final prediction result.

[0166] In this way, both the predicted values of each judgment and the factor of judgment accuracy are considered.

[0167] In this embodiment, by setting the fault probability threshold, the suspected false alarms with lower probabilities are filtered out, making the remaining high-probability predictions more accurate and reliable. And by performing weighted fusion on the predictions of different models, the accidental errors of individual models can be eliminated, improving the stability; the weighting coefficients are set according to the model accuracy, making the more accurate judgments account for a higher proportion and the results more accurate; by integrating multiple prediction results, the limitations of a single result are avoided, and a more comprehensive judgment can be provided; through probabilization and quantification, the result is transformed from a qualitative description to a quantitative description, significantly improving the accuracy and resolution; the adaptability of the present invention to complex environments is improved, and the detail description effect is good.

[0168] Therefore, in summary, through data fusion and probabilization, the refined processing obtains a fault probability refined data volume with higher precision, which is beneficial for analysis and judgment, and significantly improves the identification accuracy on the fault online, trace, and time slice.

[0169] This embodiment and each implementation manner propose a method and system for deep learning to identify low-order faults based on a hybrid neural network model. Among them, the hybrid neural network model is used to learn the noise characteristics of seismic data, eliminate random noise in the seismic data, and improve the signal-to-noise ratio of the seismic data; on this basis, the hybrid neural network model is applied to extract fault characteristics from the seismic data to obtain a fault probability data volume; the artificial fault interpretation results are used to train the hybrid neural network model through transfer learning to improve the generalization ability and applicability of the model. The trained hybrid neural network model is applied to actual seismic data for fault feature extraction, and the fault identification effects extracted before and after model training are compared. By increasing the artificial interpreted fault samples until the fault extraction effect of the trained hybrid neural network model is finer and more in line with geological laws; finally, a voting mechanism is used to perform weighted averaging on the fault prediction results with a higher probability of meeting the set fault probability threshold value within the time window of the target layer to obtain a refined fault probability data volume, ultimately providing a good technical solution and idea for low-order fault identification, and having good application prospects and promotion value.

[0170] Example 2

[0171] According to an embodiment of the present invention, a fault identification device based on deep learning of a hybrid neural network model is provided. The device includes:

[0172] A training set preparation unit for preparing pre-stack depth migration seismic data volume and fault interpretation data;

[0173] A seismic data denoising unit for processing the pre-stack depth migration seismic data volume using the first neural network model in the hybrid neural network model to eliminate noise and output the denoised seismic data volume;

[0174] A fault probability identification unit for processing the denoised seismic data volume using the second neural network model in the hybrid neural network model and outputting a fault probability data volume;

[0175] A transfer training unit for training the second neural network model using the fault interpretation data through transfer learning;

[0176] A fault identification iteration unit for processing the denoised seismic data volume using the second neural network model trained in step 4, and comparing the calculated fault probability data volume with the fault probability data volume obtained in step 3. If the fault identification effect of the calculated fault probability data volume is better than the fault probability data volume obtained in step 3, then end step 5. Otherwise, further increase the artificial fault interpretation data samples, and then return to step 4;

[0177] A refinement processing unit for refining the tomographic probability data volume calculated in step 5 to obtain a final refined tomographic probability data volume.

[0178] In some embodiments, the first neural network model is a deep neural network model based on a residual network model.

[0179] In some embodiments, the second neural network model is a deep neural network model based on a U-Net model.

[0180] In some embodiments, the apparatus further includes a first network model training unit for training the first neural network model in the hybrid neural network model to learn noise features using a seismic data noise training set to eliminate noise in the seismic data, where the seismic data noise training set includes synthetic data and actual data.

[0181] This embodiment and each implementation manner propose a method and system for identifying low-order faults based on deep learning of a hybrid neural network model. Among them, the hybrid neural network model is used to learn the noise characteristics of seismic data, eliminate random noise in the seismic data, and improve the signal-to-noise ratio of the seismic data; on this basis, the hybrid neural network model is applied to extract fracture characteristics from the seismic data to obtain a tomographic probability data volume; the hybrid neural network model is trained using the artificial tomographic interpretation results through transfer learning to improve the generalization ability and applicability of the model. The trained hybrid neural network model is applied to actual seismic data for fracture feature extraction, and the tomographic identification effects extracted before and after model training are compared. By increasing the artificial interpretation fault samples until the tomographic extraction effect of the trained hybrid neural network model is finer and more in line with geological laws; finally, a voting mechanism is used to perform weighted averaging on the fault prediction results with a higher probability of meeting the set tomographic probability threshold value within the time window of the target layer to obtain a refined tomographic probability data volume, ultimately providing a good technical solution and idea for low-order fault identification, and having good application prospects and promotion value.

[0182] For other detailed descriptions and advantages of this embodiment, reference may be made to the corresponding descriptions in the foregoing embodiments, which will not be elaborated herein.

[0183] Example 3

[0184] According to another aspect of the present invention, an electronic device is further provided. The electronic device includes:

[0185] A memory storing executable instructions:

[0186] A processor that runs the executable instructions in the memory to implement the fracture identification method based on deep learning of a hybrid neural network model according to the present invention.

[0187] Specifically, the memory may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0188] The processor may be a central processing unit (CPU) or other forms of processing units with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In an embodiment of the present invention, the processor is used to run the computer-readable instructions stored in the memory.

[0189] For a detailed description of this embodiment, reference may be made to the corresponding descriptions in the foregoing embodiments, and details are not repeated herein.

[0190] Example 4

[0191] According to another aspect of the present invention, there is also provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the fracture recognition method based on deep learning of a hybrid neural network model according to the present invention.

[0192] The computer-readable storage medium according to an embodiment of the present invention stores non-temporary computer-readable instructions. When the non-temporary computer-readable instructions are run by a processor, all or part of the steps of the methods of the foregoing embodiments of the present invention are executed.

[0193] The above-mentioned computer-readable storage media include but are not limited to: optical storage media (such as CD-ROM and DVD), magneto-optical storage media (such as MO), magnetic storage media (such as magnetic tape or removable hard disk), media with built-in rewritable non-volatile memory (such as memory card), and media with built-in ROM (such as ROM cartridge).

[0194] Those skilled in the art should understand that in order to solve the technical problem of how to obtain a good user experience effect, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included in the protection scope of the present invention.

[0195] For a detailed description of this embodiment, reference may be made to the corresponding descriptions in the foregoing embodiments, and details are not repeated herein.

[0196] Example 5

[0197] The effectiveness of the fracture identification solution based on the deep learning of the hybrid neural network model according to the present invention is verified through tests in a certain work area as follows.

[0198] Figure 2 (a) shows the time slice attributes of the target layer before fault enhancement processing of the seismic data in the depth domain of a certain work area. Figure 2 (b) shows the time slice attributes of the target layer after fault enhancement processing according to the embodiment of the present invention. As can be seen from the figure, after random noise removal by the hybrid neural network model of the present invention, the slice signal-to-noise ratio is significantly improved and the fracture characteristics are enhanced.

[0199] Figure 3 (a) shows the high-precision coherence attributes of the target layer in a certain work area. Figure 3 (b) shows the fracture probability volume attributes extracted by the hybrid neural network model according to the embodiment of the present invention. As can be seen from the figure, the fracture prediction effect of the hybrid neural network model is better.

[0200] Figure 4 (a) shows the fracture probability volume attributes extracted by the hybrid neural network model of the target layer in a certain work area according to the embodiment of the present invention. Figure 4 (b) shows the fracture attributes of the target layer after refinement processing of the fault probability volume according to the embodiment of the present invention. As can be seen from the figure, after the refinement processing of the fault probability volume, the fracture characteristics are more obvious and the identification accuracy of low-order fractures is further improved.

[0201] Figure 5 (a) shows the refined attributes of the fault probability volume extracted by the hybrid neural network model of the target layer in a certain work area according to the embodiment of the present invention. Figure 5 (b) shows the fracture probability volume attributes extracted by the hybrid neural network model after transfer learning training according to the embodiment of the present invention. As can be seen from the figure, by adding an artificial interpretation fracture model, the hybrid neural network model is more adapted to the fracture characteristics of the study area through transfer learning, and the identification effect of low-order fractures is significantly improved.

[0202] For other detailed descriptions of this exemplary embodiment, reference may be made to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0203] The present invention proposes a technical solution for identifying low-order fractures based on the deep learning of a hybrid neural network model. Each embodiment and implementation manner of the present invention has the following advantages:

[0204] 1. Improve the identification accuracy of low-order fractures

[0205] The method based on the hybrid neural network model proposed by the present invention can automatically and efficiently extract complex low-order fracture characteristics from seismic data through deep learning technology. The identification effect is significantly better than the traditional post-stack geometric attribute technology, solving the problem of low-order fracture identification.

[0206] 2. Saving a large amount of computing time

[0207] Compared with the traditional prestack wave equation anisotropy analysis method, the present invention is based on deep learning modeling, avoiding a large amount of manual feature engineering and complex operations, with high computing efficiency, greatly saving the workload and meeting the high-efficiency production requirements;

[0208] 3. Improving applicability

[0209] Through transfer learning and model iterative training techniques, and continuously increasing artificial fault interpretation data during iteration, the present invention can continuously improve the model, adapt to the geological conditions in different regions, and improve the applicability of the method;

[0210] 4. Improving the recognition accuracy of faults on inline, trace and time slice

[0211] Adopting a voting mechanism, weighted averaging is performed on the fault prediction results with a relatively high probability of meeting the set fault probability threshold value within the time window range of the target layer to obtain a refined data volume of fault probability;

[0212] 5. Automatically mining data features

[0213] Through the deep learning network model for representation learning of data samples, data features are automatically mined, and there is no need for manual calculation of attributes, realizing the mining of fracture - crack information based on the depth of the convolutional neural network model;

[0214] 6. Improving the ability of the model to solve complex problems

[0215] A hybrid neural network model is formed by ResNet (residual network model) and U - Net. U - Net takes into account both the underlying and high - level information of data features, while ResNet alleviates the problem of model degradation, deepens the number of layers of the network model, and significantly improves the ability of the network model to solve complex problems;

[0216] 7. Providing support for development

[0217] High - precision and high - efficiency identification of low - order faults can clarify the fine fracture structure, accurately judge the formation permeability, and provide strong technical support for subsequent development activities such as fracturing parameter optimization and well location selection;

[0218] 8. Popularization and application prospects

[0219] The technical route of the present invention is novel and the effect is remarkable. It is expected to be popularized and applied in the oil and gas field, extended to the extraction of more types of geological body features, and has good application prospects and popularization value.

[0220] In summary, the present invention has obvious innovations and breakthroughs in the technical route of low-level order fracture recognition, with excellent effects, bringing many direct and potential economic benefits, and is a high-level original technical achievement.

[0221] The embodiments of the present invention have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the technical improvement of the technology in the market, or to enable other ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A fracture recognition method based on deep learning of a hybrid neural network model, characterized in that, the method includes: Step 1, prepare the pre-stack depth migration seismic data volume and fault interpretation data; Step 2, use the first neural network model in the hybrid neural network model to process the pre-stack depth migration seismic data volume to eliminate noise and output the denoised seismic data volume; Step 3, use the second neural network model in the hybrid neural network model to process the denoised seismic data volume and output the fault probability data volume; Step 4, use the fault interpretation data to train the second neural network model through transfer learning; Step 5, use the second neural network model trained in Step 4 to process the denoised seismic data volume, and compare the calculated fault probability data volume with the fault probability data volume obtained in Step 3. If the fault recognition effect of the calculated fault probability data volume is better than the fault probability data volume obtained in Step 3, end the calculation; otherwise, further increase the artificial fault interpretation data samples, and then return to Step 4; Step 6, refine the fault probability data volume calculated in Step 5 to obtain the final refined fault probability data volume.

2. The method according to claim 1, characterized in that, the first neural network model is a deep neural network model based on a residual network model.

3. The method according to claim 1, characterized in that, the second neural network model is a deep neural network model based on a U-Net model.

4. The method according to claim 1, characterized in that, Step 6 specifically includes: In the time window of the target layer, set the fault probability threshold value, and use the voting mechanism to perform weighted averaging on the probability values higher than the fault probability threshold value in the fault probability data volume calculated in Step 5 to obtain the final refined fault probability data volume.

5. The method according to claim 1, characterized in that, the method further includes: Use the seismic data noise training set to train the first neural network model in the hybrid neural network model to learn the noise characteristics to eliminate the noise in the seismic data. The seismic data noise training set includes synthetic data and actual data.

6. A fracture recognition device based on deep learning of a hybrid neural network model, characterized in that, the device includes: A training set preparation unit for preparing the pre-stack depth migration seismic data volume and fault interpretation data; A seismic data denoising unit for using the first neural network model in the hybrid neural network model to process the pre-stack depth migration seismic data volume to eliminate noise and output the denoised seismic data volume; A fault probability recognition unit for using the second neural network model in the hybrid neural network model to process the denoised seismic data volume and output the fault probability data volume; A transfer training unit for using the fault interpretation data to train the second neural network model through transfer learning; A fracture identification iteration unit, which is used to process the denoised seismic data volume by using the second neural network model trained in step 4, and compare the calculated fault probability data volume with the fault probability data volume obtained in step 3. If the fault identification effect of the calculated fault probability data volume is better than that of the fault probability data volume obtained in step 3, then end step 5; otherwise, further increase the artificial fault interpretation data samples, and then return to step 4; A refinement processing unit, which is used to refine the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

7. The apparatus according to claim 6, wherein, the first neural network model is a deep neural network model based on a residual network model.

8. The apparatus according to claim 6, wherein, the second neural network model is a deep neural network model based on a U-Net model.

9. The apparatus according to claim 6, wherein, the refinement processing unit is specifically used for: within the time window of the target layer, set a fault probability threshold value, and use a voting mechanism to perform weighted averaging on the probability values higher than the fault probability threshold value in the fault probability data volume calculated in step 5 to obtain a final refined fault probability data volume.

10. An electronic device, wherein, the electronic device includes: a memory storing executable instructions; a processor that runs the executable instructions in the memory to implement the method according to any one of claims 1-5.

11. A computer-readable storage medium storing a computer program, which when executed by a processor implements the method according to any one of claims 1-5.

Images