• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    A method for automatically interpreting a subsurface characteristic within geophysical data, the method including: storing, in a computer memory, geophysical data obtained from a survey of a subsurface region; and extract, with a computer, a volume of characteristic probability by processing the geophysical data with one or more fully convolutional neural networks, trained to relate the geophysical data to at least one subsurface characteristic, where the extraction includes merging the outputs of one or more fully convolutional neural networks.
    公开号:BR112020001110A2
    申请号:R112020001110-0
    申请日:2018-08-09
    公开日:2020-07-21
    发明作者:Wei D. LIU；Diego A. Hernandez；Niranjan A. Subrahmanya；D. Braden Fitz-Gerald
    申请人:Exxonmobil Upstream Research Company；
    IPC主号:
    专利说明:

    [0001] [0001] This application claims priority benefit from United States Provisional Patent Application No. 62 / 550,069, filed on August 25, 2017, entitled AUTOMATED SYSMIC INTERPRETATION USING COMPLETE CONVOLUTIONARY NEURAL NETWORKS, the disclosure of which is incorporated herein by reference. TECHNOLOGICAL FIELD
    [0002] [0002] The exemplary modalities described here generally refer to the field of geophysical prospecting and, more particularly, to the analysis of seismic data or other image data from geophysical subsurface. Specifically, the disclosure describes exemplary modalities that use convolutional neural networks to automatically detect and interpret subsurface characteristics that can be highlighted using a contiguous region of pixels / voxels in a seismic volume. BACKGROUND
    [0003] [0003] This section intends to introduce several aspects of the technique, which can be associated with exemplary modalities of the present invention. This discussion is believed to help provide a framework to facilitate a better understanding of particular aspects of the present invention. Therefore, it should be understood that this section should be read in this light, and not necessarily as admissions to the prior art.
    [0004] [0004] A conventional hydrocarbon exploration workflow currently focuses on seismic imagery and geological interpretation of the resulting imagery. Although much effort has been made to improve and automate many aspects of seismic imaging [1,2], interpretation has remained largely a labor-intensive process. Specifically, horizon interpretation and fault interpretation are two critical and time consuming aspects of the seismic interpretation workflow that require a significant amount of time and manual effort. Recent developments in vendor technology have helped to reduce the time required for horizon interpretation [3] through the development of automated / semi-automated computational techniques [4], but robust methods for interpreting failures are lacking. There have been several attempts to fully or partially automate fault interpretation [5] using stacked / migrated seismic data, but due to the inherent uncertainty of the problem, no viable approaches have been developed that significantly reduce overall interpretation time.
    [0005] [0005] Based on the recent success in the application of convolutional and deep learning neural networks to image recognition problems [6], there have been recent attempts to apply this technology to the problem of interpreting seismic faults. Specifically [7, 8] describe a technique for applying deep learning to raw traits directly using a special loss function (Wasserstein loss), but the computational complexity of the technique requires significant sampling, resulting in a loss of accuracy in interpreting faults . In [9], a snippet-based approach to seismic faults and feature interpretation (such as channel) is described using in-depth learning. This is the approach that most closely matches the proposed approach, but its use of a section around each pixel / voxel to detect the characteristic in the center of the section makes it very computationally intensive for application in large data sets. In addition, the use of a network that is not entirely convolutional (VGG Net [10]) means that the size of the input section is fixed and, when applied to seismic volumes, produces an increase in computational expenditure due to redundant calculations (such as a separate section needs to be processed to label each voxel). When a network is not fully convolutional, it cannot have arbitrary stretch sizes. Flexible or arbitrary section sizes are useful for detecting seismic characteristics, given the nature of their different sizes and scales, depending on the type of geological environments and structure styles.
    [0006] [0006] Conventional methods have one or more of the following deficiencies.
    [0007] [0007] They require the creation and manipulation of volumes of additional attributes (such as appearance).
    [0008] [0008] They require the selection of a fixed section size, which is generally small (for example, 32 x 32), not only for computational efficiency, but also to enable the training of the statistical model. For example, if large stretch sizes are used, then, if a stretch is moved over a fault by a pixel, the contents of the stretch will be highly correlated due to the large overlap, but the labels will change from positive to negative as the center the section moves over a fault. This makes models that use sections as input and try to predict the feature in the center of the section numerically challenging to be trained as the section size increases. However, in practice, features such as faults can be recognized in areas with a low signal-to-noise ratio only by looking at large regions (such as 512 by 512) to improve the presence of fault tips. In addition, due to the high degree of correlation in the input images as we proceed over fine features, such as flaws, there is a loss in the accuracy with which these features can be located on the output map.
    [0009] [0009] They require the application of the method in a stretch for each pixel / voxel that needs to be labeled (making the execution time computationally intense). The method we use can generate labels for all pixels in a stretch at one time, reducing computational complexity during implementation by some orders of magnitude (for example, up to 5 orders of magnitude for 512 x 512 size runs). SUMMARY
    [0010] [0010] A method to automatically interpret a subsurface characteristic within geophysical data, the method including: storing, in a computer memory, geophysical data obtained from a survey of a subsurface region; and extract, with a computer, a volume of characteristic probability by processing the geophysical data with one or more fully convolutional neural networks, trained to relate the geophysical data to at least one subsurface characteristic, where the extraction includes merging the outputs of one or more fully convolutional neural networks.
    [0011] [0011] In the method, geophysical data can be a migrated or stacked seismic volume.
    [0012] [0012] In the method, geophysical data may include attributes extracted from a migrated or stacked seismic volume.
    [0013] [0013] The method may also include training one or more fully convolutional neural networks with training data, where the training data includes synthetically generated physical subsurface models, consistent with the geological background provided and computer simulated data based on in the governing equations of geophysics and in the synthetically generated physical subsurface model.
    [0014] [0014] In the method, training data can include seismic data migrated or stacked with manual interpretations.
    [0015] [0015] In the method, the training data can be a mixture of synthetic and real data.
    [0016] [0016] In the method, one or more fully convolutional neural networks can be based on a U-network architecture.
    [0017] [0017] In the method, one or more fully convolutional neural networks can be based on increases in a U-network architecture.
    [0018] [0018] In the method, one or more artificial neural networks can use convolution or 3D filtering operations.
    [0019] [0019] In the method, a plurality of neural networks can be used and the plurality of neural networks have different architectures and the training includes training the plurality of neural networks with different sets of data.
    [0020] [0020] In the method, the merger can be done using voxelwise operations.
    [0021] [0021] In the method, voxelwise operations include averaging.
    [0022] [0022] In the method, voxelwise operations include obtaining a maximum value.
    [0023] [0023] In the method, the fusion can be done by feeding several forecast volumes, and the original data, into another artificial neural network.
    [0024] [0024] In the method, at least one subsurface characteristic is one or more of faults, channels or deposition environments.
    [0025] [0025] In the method, at least one subsurface characteristic is a fault.
    [0026] [0026] In the method, the extraction may include the interpretation of seismic characteristics via voxelwise labeling.
    [0027] [0027] In the method, the extraction may include the execution of a 2D or 3D model learned in the entirety of a seismic volume to obtain a failure interpretation of the seismic volume at once.
    [0028] [0028] In the method, the extraction can include the generation of an output label map that is related to an input image size. BRIEF DESCRIPTION OF THE DRAWINGS
    [0029] [0029] Although the present disclosure is susceptible to several modifications and alternative forms, specific exemplary modalities of the same have been shown in the drawings and are described in detail here. It should be understood, however, that the description here of specific exemplary modalities is not intended to limit the disclosure to the particular forms disclosed here, but, on the contrary, this disclosure must cover all modifications and equivalents, as defined by the appended claims. It should also be understood that the drawings are not necessarily to scale, the emphasis being instead placed on clearly illustrating the principles of the exemplary modalities of the present invention. In addition, certain dimensions can be exaggerated to help visually convey these principles.
    [0030] [0030] Figure 1 illustrates an example of a fully convolutional network architecture (it is based on the U-network architecture, but highlights important modifications for greater precision in locating fine features, such as faults).
    [0031] [0031] Figure 2A illustrates an example of seismic data stacked as an input.
    [0032] [0032] Figure 2B illustrates an example of a fault mask interpreted manually as an output.
    [0033] [0033] Figure 3 illustrates an example of synthetic seismic data as input and synthetically induced faults as output.
    [0034] [0034] Figure 4A illustrates an example of a vertical range with manual interpretation of faults.
    [0035] [0035] Figure 4B illustrates an example of a range in a direction orthogonal to manual interpretation.
    [0036] [0036] Figure 4C illustrates an example of a time interval with manual interpretation.
    [0037] [0037] Figure 5A illustrates an example of entry training stretches.
    [0038] [0038] Figure 5B illustrates an example of target failure masks for incoming training stretches.
    [0039] [0039] Figure 5C illustrates failure masks foreseen for incoming training stretches.
    [0040] [0040] Figure 6 illustrates an exemplary method that incorporates the present technological advance.
    [0041] [0041] Figures 7A, 7B and 7C illustrate exemplary results obtained from the present technological advance.
    [0042] [0042] Figures 8A, 8B and 8C illustrate exemplary results obtained from the present technological advance.
    [0043] [0043] Figures 9A, 9B and 9C illustrate exemplary results obtained from the present technological advance. DETAILED DESCRIPTION
    [0044] [0044] Exemplary modalities are described here. However, insofar as the following description is specific to a specific modality, this is for exemplary purposes only and simply provides a description of the exemplary modalities. Therefore, the invention is not limited to the specific modalities described below, but includes all alternatives, modifications and equivalents that fit the true spirit and scope of the attached claims.
    [0045] [0045] The present technological advance can be incorporated as a method based on fully convolutional neural networks with image-to-volume or volume-to-volume training to automatically detect and interpret the subsurface characteristics that can be highlighted using a contiguous region of pixels / voxels in a seismic volume (for example, a subsurface characteristic, such as faults, channels, deposition environments, etc.). The present technological advance can work with stacked or migrated seismic data with or without additional attributes, such as appearance. The method output can be a characteristic probability volume that can be further processed to extract objects that integrate with a subsurface interpretation workflow. A characteristic probability volume is a 4D tensor that transmits a vector in each voxel, indicating the probability that that voxel belongs to a certain class (for example, failure, channel, hydrocarbon trap, salt, etc.). The following discussion will use fault interpretation as an example of the application of this technological advance. However, this is not intended to be limiting, as the present technological advance can be used to detect channels, bodies of salt, etc. when user-provided labels are available.
    [0046] [0046] The present technological advance can overcome all the problems mentioned above of conventional techniques. The present technological advance can use the latest ideas from the field of deep learning. RNAs (artificial neural networks), particularly deep neural networks (RNP), are built on the premise that they can be used to replicate any arbitrary continuous functional relationship, including non-linear relationships. RNPs can include weighted node "layers" that are activated by previous layer inputs. These networks are trained with examples in which the correct / true output (label) is known for a given input; the weighted parameters in the network nodes evolve due to the minimization of the error between the forecast and the true value. This makes the network an increasingly better predictor of training examples and, ultimately, any examples of data of a similar nature to training data. Convolutional neural networks are a class of deep neural networks that are especially suitable for processing spatial data [6].
    [0047] [0047] The present technological advance can leverage a technical approach used in the field of computer vision, specifically semantic segmentation. In semantic segmentation, each pixel / voxel is labeled with a class of objects (ie, fail / fail). Convolutional neural networks, where layers are restricted to perform only certain operations (limited to convolution, correlation, nonlinearity by elements, downward sampling, upward sampling, upward convolution) can be called "fully convolutional networks" and these networks can receive inputs arbitrary size and produce results of corresponding size with efficient inference and learning. A fully convolutional network is made up of a select set of operations that can be applied to inputs of any size (this includes operations such as convolution, census, upward sampling, channel concatenation, addition of channels, etc.). More traditional networks usually have some convolutional layers, followed by a layer that vectorizes all layers and then uses a multilayer perceptron network to perform the final classification task. The multi-layer perceptron network can only handle a fixed number of inputs (hence a fixed number of pixels) and is therefore not suitable for application to input images of varying sizes. These fully convolutional networks are ideal for spatially dense forecasting tasks [11]. The interpretation of seismic characteristics by means of voxelwise labeling falls into this category. Although the original document mainly describes 2D data processing, similar concepts can be used to develop fully convolutional 3D networks to process 3D seismic data. The present technological advance can be used with the training of 2D and 3D models. Specifically, an exemplary modality of the present technological advance can use the “U-shaped” architecture outlined in
    [12] [12] and its customized extensions (as shown in figure 1) as the main candidate to generate fully convolutional networks in 2D and 3D. This network architecture, initially applied to biomedical imaging tasks, has the desirable characteristic of merging information of various scales to increase the accuracy in the segmentation of images.
    [0048] [0048] While the U-network architecture is described, the present technological advance can use other fully convolutional neural network architectures and can even work with a set of neural networks with different architectures and can be trained on different data sets. One way to use multiple networks is to train a network to identify a characteristic that sees it from different views (X, Y, Z views) and then combine the predictions for each voxel (using a fusion rule, how to use the maximum three forecasts). Another method may involve training different networks to detect characteristics at different scales, that is, one network analyzes portions of size 128x128 pixels while another analyzes portions of size 512x512 pixels. These methods are presented as examples of the use of several networks and should not be considered as the only way in consideration.
    [0049] [0049] Figure 1 illustrates an example of a fully convolutional augmented U-network architecture. Each square or rectangular shaded box corresponds to a map of multichannel characteristics. The number of channels or filters is indicated at the top of the box. Filters are not fixed, and they learn from the data. White boxes represent maps of copied features. The arrows indicate the different operations.
    [0050] [0050] The network architecture in figure 1 is based on the U network architecture described in [12] with potential increase / modification, as described below. The augmented U-network architecture includes a contract path (left side), an expansive path (right side) and additional convolutional layers at the highest resolution. The contracting path includes the repeated application of 3x3 convolutions, followed by a rectified linear unit (ReLU) and descending sampling (using maximum census or striated convolutions). The number of convolutional filters in each layer and the descending sampling scale are defined by the user. Each step on the expansive path includes an upward sampling of the feature map, followed by a 2x2 convolution, a concatenation with the corresponding cut feature maps of the hiring path, and several 3x3 convolutions, each followed by a ReLU. Cropping is used due to loss of edge pixels if fill is not used. Finally, multiple convolutional layers (or residual layers [13]) can be added at a resolution equivalent to the input image. In the final layer, a 1x1 convolution is used to map each multicomponent feature vector to the desired number of classes. This vectorized network output at each pixel is stored at the pixel location to generate a 4D tensor that is the probability volume of the characteristic. Departures can take place in different orientations, depending on the training scheme. In this specific approach, the network was trained in the x, y, z directions. Therefore, the results were merged to provide the final 3D probability of failure volume.
    [0051] [0051] The main computational cost of this U-network architecture occurs only once, in advance, during network training. After the convolutional network is trained, it is possible to produce predictions for entire intervals (in 2D) or volumes (in 3D) in a fraction of the training time. The accuracy of this network is significantly better than traditional approaches not based on deep learning. It is also significantly better than previous deep learning approaches for seismic interpretation, as it needs some orders of magnitude of less time for predictions. This means that automated seismic interpretation can now become feasible, both in terms of achieving a level of accuracy in predictions that can have a significant impact in reducing interpretation time and in executing the task in an acceptable period of time.
    [0052] [0052] Below is a discussion of exemplary steps that can be used to implement a modality of the present technological advance. Not all steps may be necessary in all the modalities of this technological advance. Figure 6 illustrates an exemplary method that incorporates the present technological advance. The volumes of probability of failure generated from models executed in the three orthogonal views can be merged to generate the final failure volume (see step 604 in figure 6).
    [0053] [0053] Data generation - step 601. The training of a totally convolutional neural network requires the supply of several pairs of seismic inputs and sections of target label or volumes. An excerpt refers to a part extracted from a seismic image (2D or 3D) that represents the region being analyzed by the network. The excerpt must contain sufficient information and context for the network to recognize the characteristics of interest. This can be done by extracting sufficiently sized portions of actual seismic data (see figure 2A) with fault masks interpreted manually (see figure 2B) as labels. Due to the highly variable signal / noise ratio present in the seismic data, manual interpretations will always have an error. To overcome this problem, the current technological advancement can adopt the simple approach of annotating pixels / voxels around a fault interpreted manually as faults as well. For synthetic data, the present technological advance can create appropriate volumes of rock properties and artificially introduce flaws by sliding the volumes of rock properties. The seismic image is then generated from that "failed" volume using wave propagation models or convolution models (see figure 3). Magnifying the image (mirroring the data, rotating the data, etc.) can be used to make the training data cover a wider region of applicability.
    [0054] [0054] The geophysical data described in this example is seismic data, but other types of data (gravity, electromagnetic) can be used. The present technological advance is not limited to seismic data. Geophysical data can be a migrated or stacked seismic volume. Geophysical data can include attributes extracted from migrated or stacked data.
    [0055] [0055] Training - step 602. The training of a totally convolutional neural network involves learning millions of parameters that define the filters applied to the input data at various scales. The network can learn these millions of parameters by optimizing the value of the parameters to minimize a discrepancy measure based on comparing the network's predictions with the training material provided by the user. The discrepancy measure can include several standard loss functions used in machine learning, such as pixel / voxel losses ("square loss", "absolute loss", "binary cross entropy", "categorical cross entropy") and losses that appear in larger regions, such as “adverse loss” [14]. This is a very large scale optimization problem and is best used with specialized hardware (GPU workstations or high-performance computers) to train models in a reasonable amount of time (hours to days). Specifically, an exemplary training procedure may include the use of a specific stochastic gradient descent optimization variant (called "Adam" [15]) with data parallelism using multiple GPUs, where multiple samples of data are evaluated on each GPU and the gradient estimate for all GPUs was averaged to obtain the batch gradient estimate used by the optimizer. Many standard neural network training options (such as dropout regularization, batch standard, etc.) can be used to improve the quality of trained models.
    [0056] [0056] The training data for the artificial neural network may include models of synthetically generated subsurface physical properties, consistent with the geological background provided, and computer simulated data based on the governing equations of geophysics and models of physical properties of subsurface generated.
    [0057] [0057] Training data for the artificial neural network can include migrated or stacked geophysical data (for example, seismic data) with interpretations made manually.
    [0058] [0058] The artificial neural network can be trained using a combination of synthetic and real geophysical data.
    [0059] [0059] Dealing with Directionality in 2D Networks - step 603. For 2D networks, the present technological advance can extract stretches in the three orthogonal directions and train a different network for visualizations in each direction. The results of these networks can be merged to provide the final 3D probability of failure. 3D networks are robust to this variation in data visualization (for example,
    [0060] [0060] Prediction: Use of 2D Networks - 604. The use of fully convolutional networks allows the prediction of input images with a different size than the stretch used in training. The input image can be propagated through the trained network using a sequence of operations defined by the network (figure 1) and the parameters learned during training. Networks always generate an output label map that is the same size as the input image. The present technological advance can therefore run 2D models in entire "intervals" of seismic data. Alternatively, if the memory is limited, the sections can be extracted from the test volume, propagated through the network and the output can be sewed again in a volume corresponding to the size of the test volume. If there are several models trained in different views, the present technological advance can generate a volume of probability of failure for each model and the final decision would involve the merger of these volumes. The method can run a scheme to select significant predictions from the same orientation used during training mode (x, y, z). This scheme can merge the predictions to provide the final volume of 3D probability of failure. Fusion itself can be a simple method (for example, using the multiplication of individual probability volumes, averaging individual probability volumes, calculating the maximum individual probability, etc.) or the current technological advancement can train a network 3D with each of these individual volumes (and potentially the seismic data as well) as a channel to learn the best way to merge the volumes.
    [0061] [0061] Prediction: Use of 3D Networks - 604. The present technological advance can execute the 3D model learned in the entire seismic volume to obtain the interpretation of the failure in one go (that is, all at once). Computationally, the GPU's memory can be a limiting factor in the implementation of this and the 3D volume may need to be divided into manageable parts to perform the forecast. However, this is not a limitation of current technological advancement, but it is a limitation of some GPU computers.
    [0062] [0062] Post-processing. All post-processing steps (for example, Median Filtering, DBScan-based out-of-curve detection, Ridge detection [16]) that use an attribute volume for fault interpretation can still be applied to the volume generated by the steps above to adjust the results. For example, one can review the method mentioned in [4]. Post-processing can also include feeding the output of a neural network into another neural network (recursively in the same network or another network specifically trained for post-processing). The entrance to the next neural network can also include the original seismic image. It is possible to have more than 2 steps in a segmentation of post-processing instructions. Numerical Examples
    [0063] [0063] The numerical examples below show that the present technological advance can build interpretations of failures with good precision. Fault interpretation refers to techniques associated with creating maps from seismic data, representing the geometry of the subsurface fault structure. In this particular case, a volume of probability of applying convolutional neural networks will be used to provide a reasonable prediction of the presence of faults. However, the accuracy of the results obtained by the present technological advance can improve with the use of more sophisticated RNP architectures (for example, ResNets [13]) and larger data sets.
    [0064] [0064] For these examples, the training data included faults interpreted manually from a cut seismic volume.
    [0065] [0065] Measurements from seismic data, such as amplitude, dip, frequency, phase or polarity, often called seismic attributes or attributes. A seismic attribute is an amount extracted or derived from seismic data that can be analyzed to improve information that can be more subtle in a traditional seismic image. Statistics are provided below. Dimensions of the seismic volume: 1663 x 1191 x 541 Voxel count: 1,071,522,453.
    [0066] [0066] Figures 4A-C are exemplary intervals of a training volume. Figure 4A illustrates an example of a vertical range with manual interpretation of faults. Figure 4B illustrates an example of a range in a direction orthogonal to manual interpretation. Figure 4C illustrates an example of a time interval with manual interpretation. Reference numbers 401 indicate faults interpreted manually.
    [0067] [0067] Figures 5A-C illustrate samples of 2D sections extracted from the training volume and network forecasts for the selected sections. Figure 5A illustrates an example of entry training sections. Figure 5B illustrates an example of target failure masks for incoming training stretches. Figure 5C illustrates predicted failure masks for the incoming training stretches.
    [0068] [0068] It is interesting to note that, even in the training data set, the network is able to identify faults that were missing during manual interpretation (see arrows 501), confirming the hypothesis that the network “learned” to recognize faults and can generalize to invisible failures in addition to training data.
    [0069] [0069] Figures 7A-C illustrate another example for the present technological advance. Figure 7A illustrates a range of a volume of amplitude. Figure 7B illustrates faults interpreted manually, where faults 701 are identified. Figure 7C illustrates the results of this technological advance. While the present technological advance has not identified all the flaws in figure 7B, it has identified the previously unidentified flaw 702.
    [0070] [0070] Figures 8A-C illustrate another example for the present technological advance. Figure 8A illustrates a range of a volume of amplitude. Figure 8B illustrates faults interpreted manually, in which faults 801 are identified. Figure 8C illustrates the results of this technological advance. While the present technological advance has not identified all the failures in Figure 7B, it has identified the 802 failure not previously identified.
    [0071] [0071] Figures 9A-C illustrate another example for the present technological advance. Figure 9A illustrates a range of a volume of amplitude. Figure 9B illustrates faults interpreted manually, where faults 901 are identified. Figure 9C illustrates the results of this technological advance. While the present technological advance has not identified all the faults in Figure 7B, it has identified the previously unidentified fault 902.
    [0072] [0072] The interpreted faults can be used to explore or manage hydrocarbons. Fault and horizon interpretations have been used to describe the subsurface structure and capture mechanisms for hydrocarbon exploration. Many or the largest camps in the world are compartmentalized and trapped by failure; therefore, in the sense of exploration, the interpretation of subsurface could be one of the most critical tasks for finding oil and gas. Different geoscientists and seismic interpreters use a variety of approaches and philosophies in their interpretations, however, all traditional methods are time consuming and data dependent. Automation via the application of convolutional neural networks has the potential to accelerate this long process and reduce the time needed to identify and explore the type of opportunity. As used here, hydrocarbon management includes hydrocarbon extraction, hydrocarbon production, hydrocarbon exploration, identification of potential hydrocarbon characteristics, identification of well locations, determination of injection and / or well extraction rates, identification of water connectivity. reservoir, acquisition, disposal and / or abandonment of hydrocarbon characteristics, review previous hydrocarbon management decisions and any other acts or activities related to hydrocarbons.
    [0073] [0073] In all practical applications, the present technological advance must be used in conjunction with a computer, programmed in accordance with the disclosures in this document. Preferably, in order to efficiently execute the present technological advance, the computer is a high performance computer (HPC), known to those skilled in the art. These high-performance computers typically involve clusters of nodes, each node with multiple CPUs and computer memory that allow parallel computing. Models can be viewed and edited using any interactive visualization programs and associated hardware, such as monitors and projectors. The architecture of the system can vary and can be composed of any number of suitable hardware structures capable of performing logical operations and displaying the output in accordance with the present technological advance. Those skilled in the art know the suitable supercomputers available from Cray or IBM.
    [0074] [0074] The aforementioned request is directed to particular modalities of the present technological advance in order to illustrate it. It will be apparent, however, to one skilled in the art, that many modifications and variations in the modalities described in this document are possible. All such modifications and variations must be within the scope of the present invention, as defined in the appended claims. Those skilled in the art will readily recognize that, in the preferred embodiments of the invention, some or all of the steps of the present inventive method are performed using a computer, that is, the invention is implemented by a computer. In such cases, the resulting gradient or the updated physical property model can be downloaded or saved to the computer's storage.
    [0075] [0075] The following references are incorporated by reference in their entirety:
    [1] [1] Akcelik, V., Denli, H., Kanevsky, A., Patel, K.K., White, L. and Lacasse M.-D. “Multiparameter material model and source signature full waveform inversion”, SEG Technical Program Expanded Abstracts, pages 2406-2410, 2012;
    [2] [2] Denli, H., Akcelik, V., Kanevsky A., Trenev D., White L. and Lacesse, M.-D. “Full-wavefield inversion for acoustic wave velocity and attenuation”, SEG Technical Program Expanded Abstracts, pages 980-985, 2013;
    [3] [3] Paleoscan software available from Eliis;
    [4] [4] “Computer-Assisted Fault Interpretation of Seismic Data”, publication of the US patent application (Pub. No. US 2015/0234070 A1), August 20, 2015;
    [5] [5] Gibson, D., Spann, M., & Turner, J. "Automatic Fault Detection for 3D Seismic Data." DICTA. 2003;
    [6] [6] LeCun, Y., Bengio, Y., & Hinton, G., “Deep Learning.”, Nature 521, 436– 444 (May 28, 2015) doi: 10.1038 / nature14539;
    [7] [7] Zhang, C., Frogner, C. & Poggio, T. "Automated Geophysical Feature Detection with Deep Learning." GPU Technology Conference, 2016;
    [8] [8] Araya-Polo, M., Dahlke, T., Frogner, C., Zhang, C., Poggio, T., & Hohl, D., "Automated fault detection without seismic processing.", The Leading Edge, Special Edition: Data analysis and machine learning, March 2017;
    [9] [9] Jiang, Y. & Wulff, B., “Detecting prospective structures in volumetric geo-seismic data using deep convolutional neural networks.”, Poster presented on November 15, 2016 at the annual meeting of the foundation board of the International Center for Bonn-Aachen Information Technology (b-it);
    [10] [10] Simonyan, K. & Zisserman, A., “Very Deep Convolutional Networks for Large-Scale Image Recognition.”, ArXiv technical report, 2014;
    [11] [11] Jonathan Long, Evan Shelhamer, and Trevor Darrell., "Fully Convolutional Networks for Semantic Segmentation." CVPR 2015;
    [12] [12] Olaf Ronneberger, Philipp Fischer, Thomas Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation”, Computing medical images and computer-assisted intervention (MICCAI), Springer, LNCS, Vol.9351: 234--241 , 2015;
    [13] [13] K He, X Zhang, S Ren, J Sun, “Deep Residual Learning for Image Recognition” IEEE 2016 Conference on Computer Vision and Pattern Recognition (CVPR);
    [14] [14] Goodfellow, I. et. al., Generative Adversarial Nets, Advances in Neural Information Processing Systems 27, NIPS 2014;
    [15] [15] Kingma, P. D., & Ba, J., “Adam: A Method for Stochastic Optimization”,
    [16] [16] Koenderink, Jan J., van Doorn, Andrea J. (May 1994). "2 + 1-D differential geometry". Pattern. Recognition Letters. 15: 439–443.
    权利要求:
    Claims (12)
    [1]
    1. Method for automatically interpreting a subsurface characteristic within geophysical data, the method characterized by the fact that it comprises: storing, in a computer memory, geophysical data obtained from a survey of a subsurface region; and extract, with a computer, a volume of characteristic probability by processing the geophysical data with one or more fully convolutional neural networks, trained to relate the geophysical data to at least one subsurface characteristic, where the extraction includes merging the outputs of one or more fully convolutional neural networks.
    [2]
    2. Method, according to claim 1, characterized by the fact that the geophysical data is a migrated or stacked seismic volume, or attributes extracted from a migrated or stacked seismic volume.
    [3]
    3. Method according to any one of claims 1 to 2, characterized by the fact that it further comprises: training the one or more fully convolutional neural networks with training data, in which the training data includes synthetically generated physical subsurface models consistent with the geological background provided and computer simulated data based on geophysical governance equations and the synthetically generated physical subsurface model.
    [4]
    4. Method, according to claim 3, characterized by the fact that the training data includes seismic data migrated or stacked with manual interpretations.
    [5]
    5. Method, according to claim 3, characterized by the fact that the training data is a mixture of synthetic and real data.
    [6]
    6. Method according to any one of claims 1 to 5, characterized in that the one or more fully convolutional neural networks are based on a U-network architecture or enhancements to a U-network architecture.
    [7]
    Method according to any one of claims 1 to 6, characterized in that the one or more artificial neural networks use 3D convolution or filtering operations.
    [8]
    8. Method, according to claim 3, characterized by the fact that a plurality of neural networks is used and the plurality of neural networks have different architectures and the training includes training the plurality of neural networks with different data sets.
    [9]
    9. Method according to any one of claims 1 to 8, characterized by the fact that the merger is done using voxelwise operations, such as averaging or obtaining a maximum value.
    [10]
    10. Method according to any one of claims 1 to 9, characterized by the fact that the fusion is carried out by feeding several forecast volumes and, optionally, the original data, into another artificial neural network.
    [11]
    11. Method according to any one of claims 1 to 10, characterized by the fact that at least one subsurface characteristic is one or more of faults, channels or deposition environments.
    [12]
    12. Method according to any one of claims 1 to 11, characterized by the fact that the extraction includes at least one between performing interpretation of seismic characteristics via voxelwise labeling, executing a 2D or 3D model learned in the entirety of a seismic volume for obtaining a flawed interpretation of the seismic volume at once, or generating an output label map related to an input image size.
    Convolutions of high Hiring Path Expansive Path additional resolution
    Variable 32 32
    Petition 870200008376, of 01/17/2020, p. 7/51 Dedicated inlet section (single or characteristic Mask channel Multichannel pass filters) (optional)
    Additional Convolutional / Resnet Layers (optional) 1/9
    Convolution + Relu The number of layers of Concatenation descending sampling / sampling census and descending sampling ascending can be Sampling ascending chosen to optimize accuracy and performance Resize with census and ascending sampling
    Failure truth Failure truth Failure truth
    Seismic with labeled faults for training Multimodal average calculation and computational scale expansion for the total 3D volume
    Petition 870200008376, of 01/17/2020, p. 12/51 Synthetic Data Generation Real Data + Manual Interpretation 6/9
    Model Z Model X Model Y of internal Z Representation
    Input space, X Output space, Y The trained model can then fully convolutional neural network interpret gaps in invisible data
    ∈ Expected Truth
    ∈ Expected Truth
    ∈ Expected Truth
    类似技术:
    公开号 | 公开日 | 专利标题
    BR112020001110A2|2020-07-21|automated seismic interpretation using fully convolutional neural networks
    AlRegib et al.2018|Subsurface structure analysis using computational interpretation and learning: A visual signal processing perspective
    Wang et al.2018|Successful leveraging of image processing and machine learning in seismic structural interpretation: A review
    Li et al.2016|Manhattan-world urban reconstruction from point clouds
    AU2013378058B2|2017-04-20|Detecting subsurface structures
    US10036820B2|2018-07-31|Expert guided knowledge acquisition system for analyzing seismic data
    US9733391B2|2017-08-15|Method and system for geophysical modeling of subsurface volumes
    Li et al.2019|Seismic fault detection using an encoder–decoder convolutional neural network with a small training set
    CN109709603A|2019-05-03|Seismic horizon identification and method for tracing, system
    BR112021011253A2|2021-08-24|Machine Learning Systems Training for Seismic Interpretation
    de Lima et al.2019|Deep convolutional neural networks as a geological image classification tool
    BR112021011246A2|2021-08-24|Inversion guided by automated seismic interpretation
    Shi et al.2020|Waveform embedding: Automatic horizon picking with unsupervised deep learning
    Carvalho et al.2020|Automated microfossil identification and segmentation using a deep learning approach
    Larsen et al.2018|Is machine learning taking productivity in petroleum geoscience on a Moore’s Law trajectory?
    Chen et al.2018|Large-scale structure from motion with semantic constraints of aerial images
    Zhang et al.2021|Automatic seismic facies interpretation using supervised deep learning
    WO2017152123A1|2017-09-08|Context based bounded hydrocarbon formation identification
    Oke et al.2020|Application of Machine Learning-Assisted Fault Interpretation on Large Carbonate Field with Subtle Throws
    Alfarhan et al.2020|Concurrent detection of salt domes and faults using resnet with u-net
    Wang et al.2021|Seismic Stratum Segmentation Using an Encoder–Decoder Convolutional Neural Network
    Alfarhan et al.2020|Multiple Events Detection In Seismic Structures Using A Novel U-Net Variant
    Jafrasteh et al.2020|Generative adversarial networks as a novel approach for tectonic fault and fracture extraction in high resolution satellite and airborne optical images
    Wang2018|Computational seismic interpretation using geometric representation and tensor-based texture analysis
    Bogen et al.2021|Using Interactive Visualization and Machine Learning for Seismic Interpretation
    同族专利:
    公开号 | 公开日
    CA3070479A1|2019-02-28|
    WO2019040288A1|2019-02-28|
    EP3673298A1|2020-07-01|
    US20190064378A1|2019-02-28|
    US11119235B2|2021-09-14|
    SG11202000350RA|2020-03-30|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US5251286A|1992-03-16|1993-10-05|Texaco, Inc.|Method for estimating formation permeability from wireline logs using neural networks|
    US5444619A|1993-09-27|1995-08-22|Schlumberger Technology Corporation|System and method of predicting reservoir properties|
    FR2738920B1|1995-09-19|1997-11-14|Elf Aquitaine|METHOD FOR AUTOMATIC SEISMIC FACIAL RECOGNITION|
    GB9819910D0|1998-09-11|1998-11-04|Norske Stats Oljeselskap|Method of seismic signal processing|
    US6374185B1|2000-02-18|2002-04-16|Rdsp I, L.P.|Method for generating an estimate of lithological characteristics of a region of the earth's subsurface|
    US6438493B1|2000-09-29|2002-08-20|Exxonmobil Upstream Research Co.|Method for seismic facies interpretation using textural analysis and neural networks|
    US7203342B2|2001-03-07|2007-04-10|Schlumberger Technology Corporation|Image feature extraction|
    US6662112B2|2001-08-31|2003-12-09|Exxonmobil Upstream Research Company|Method for classifying AVO data using an interpreter-trained neural network|
    EP1815272B1|2004-11-24|2021-07-28|Paradigm Geophysical Ltd.|System and method for fault identification|
    EP2496967B1|2009-11-05|2021-11-17|ExxonMobil Upstream Research Company|Method for creating a hierarchically layered earth model|
    US9618639B2|2012-03-01|2017-04-11|Drilling Info, Inc.|Method and system for image-guided fault extraction from a fault-enhanced seismic image|
    CA2918268C|2013-08-16|2021-06-01|Michael David Ewing|Generating representations of recognizable geological structures from a common point collection|
    US9804282B2|2014-02-17|2017-10-31|General Electric Company|Computer-assisted fault interpretation of seismic data|
    WO2017188858A1|2016-04-28|2017-11-02|Schlumberger Canada Limited|Reservoir performance system|
    US10430725B2|2016-06-15|2019-10-01|Akw Analytics Inc.|Petroleum analytics learning machine system with machine learning analytics applications for upstream and midstream oil and gas industry|
    WO2018026995A1|2016-08-03|2018-02-08|Schlumberger Technology Corporation|Multi-scale deep network for fault detection|
    EP3497302A4|2016-08-08|2020-03-25|Services Petroliers Schlumberger|Machine learning training set generation|
    AU2017343749A1|2016-10-14|2019-04-11|Chevron U.S.A. Inc.|System and method for seismic facies identification using machine learning|
    EP3529639A1|2016-10-18|2019-08-28|Services Petroliers Schlumberger|Determining subsurface layers using machine learning|
    WO2018082790A1|2016-11-07|2018-05-11|Schlumberger Technology Corporation|Seismic data processing artificial intelligence|
    EP3580586A1|2017-02-09|2019-12-18|Services Pétroliers Schlumberger|Geophysical deep learning|US11164067B2|2018-08-29|2021-11-02|Arizona Board Of Regents On Behalf Of Arizona State University|Systems, methods, and apparatuses for implementing a multi-resolution neural network for use with imaging intensive applications including medical imaging|
    EP3894901A2|2018-12-11|2021-10-20|Exxonmobil Upstream Research Company |Automated seismic interpretation-guided inversion|
    EP3894905A2|2018-12-11|2021-10-20|ExxonMobil Upstream Research Company|Automated seismic interpretation systems and methods for continual learning and inference of geological features|
    US20200183035A1|2018-12-11|2020-06-11|Exxonmobil Upstream Research Company|Data Augmentation for Seismic Interpretation Systems and Methods|
    WO2020231379A1|2019-05-10|2020-11-19|Schlumberger Technology Corporation|System and method for identifying subsurface structures|
    US11131640B2|2019-05-13|2021-09-28|Exxonmobil Upstream Research Company|Methods using advanced NMR analysis for core samples|
    CN110321933B|2019-06-11|2021-09-14|武汉闻道复兴智能科技有限责任公司|Fault identification method and device based on deep learning|
    CN110412662A|2019-06-27|2019-11-05|中国石油化工股份有限公司|Method for prediction of reservoirs of thin interbeded based on seismic multi-attribute deep learning|
    CN110473288A|2019-08-23|2019-11-19|清华四川能源互联网研究院|Dam model method for reconstructing, device and electronic equipment|
    WO2021226612A1|2020-05-06|2021-11-11|Exxonmobil Upstream Research Company|Geological reasoning with graph networks for hydrocarbon identification|
    WO2021226613A1|2020-05-06|2021-11-11|Exxonmobil Upstream Research Company|Structured representations of subsurface features for hydrocarbon system and geological reasoning|
    WO2021226614A1|2020-05-06|2021-11-11|Exxonmobil Upstream Research Company|Framework for integration of geo-information extraction, geo-reasoning and geologist-responsive inquiries|
    US20220035069A1|2020-07-31|2022-02-03|Chevron U.S.A. Inc.|Systems and methods for identifying subsurface features as a function of position in a subsurface volume of interest|
    CN112014882B|2020-09-01|2022-02-15|电子科技大学|Pre-stack seismic reflection pattern analysis method based on tensor discrimination dictionary|
    CN112444841B|2020-12-09|2021-10-08|同济大学|Thin-layer-containing lithology earthquake prediction method based on scale-division multi-input convolution network|
    CN112925012B|2021-01-26|2021-11-09|中国矿业大学|Seismic full-waveform inversion method and device|
    CN112489047B|2021-02-05|2021-06-01|四川大学|Deep learning-based pelvic bone and arterial vessel multi-level segmentation method thereof|
    法律状态:
    2021-11-03| B350| Update of information on the portal [chapter 15.35 patent gazette]|
    优先权:
    申请号 | 申请日 | 专利标题
    US201762550069P| true| 2017-08-25|2017-08-25|
    US62/550,069|2017-08-25|
    PCT/US2018/045998|WO2019040288A1|2017-08-25|2018-08-09|Automated seismic interpretation using fully convolutional neural networks|
    [返回顶部]