CN111596978A - Web page display method, module and system for lithofacies classification by artificial intelligence - Google Patents

Abstract

The invention discloses a method for interpreting logging curves and seismogram data with artificial intelligence, and the data input and result output are displayed in the form of web pages, which facilitates remote deployment and data sharing. The invention includes collecting a part of sample data sets of known petrofacies classification as training data, using machine learning and deep learning methods to automatically identify petrofacies, and then dividing petrofacies in strata in unknown areas. The present invention includes artificial intelligence interpretation methods, modules and systems. The server on which the present invention is deployed will display the following functions: the mutual communication interface on the web page, including the data entry part, the data verification part, the data preprocessing, the establishment of the model, the training of the model, the iteration of the model, the use of the model and the result display modules. The invention covers displaying the calculation result of artificial intelligence on a web page by using a Python framework.

Description

Web page display method, module and system for lithofacies classification using artificial intelligence

The invention discloses a method for interpreting logging curves and seismogram data with artificial intelligence, and the data input and result output are displayed in the form of web pages, which facilitates cloud computing operations, cloud server deployment, and cloud data aggregation. Class and calculation result security, storage, sharing, etc. The logging curve data here refers to the physical properties of underground formation rocks, formation fluids and their mixtures; the seismogram here refers to two-dimensional, three-dimensional, four-dimensional and multi-dimensional data sets obtained by artificially transmitting seismic waves. The invention includes using a part of the collected sample data set with known facies classification as training data, and using the method of deep learning to identify the petrofacies. The lithofacies in the stratum are divided, and the thickness of the identified layers ranges from 15 cm to hundreds of meters, and the accuracy depends on the input data and the resolution of the processed data. The present invention also includes artificial intelligence interpretation methods, modules and systems. The server deploying the present invention will perform the following steps: the interactive interface on the web page, including the data entry part, the data verification part, the data preprocessing, the establishment of the model, the training of the model, the iteration of the model, the use of the model and the analysis of the results display modules. The present invention also covers displaying the calculation results of artificial intelligence on web pages using Python frameworks, including Django, Pylons, Tornado, Bottle and Flask. This lays the foundation for cloud deployment.

describe

Web page display method, module and system for petrographic classification using artificial intelligence.

technical field

The invention relates to petroleum exploration, petroleum development, logging data processing, seismogram processing, artificial intelligence, machine learning and its computing module, deep learning and its computing module, and a framework of Python plus website pages.

Background technique

In recent years, modern science and technology have been effectively utilized in the field of petroleum exploration and development, which has promoted the rapid development of the petroleum industry and brought huge benefits to the national economy. However, with the continuous improvement of oil exploration level, it becomes more and more difficult to find new oil and gas fields, which requires us to continuously improve our understanding level and use scientific methods to understand and master the unknown conditions of oil and gas existence. , reservoir development and other data to unearth more new information to predict oil and gas reservoirs.

Especially in the past two years, the petroleum industry at home and abroad is undergoing major adjustments, and many emerging disciplines have been introduced into the field of petroleum exploration and development. In the field of artificial intelligence, my country is also making arrangements. Therefore, the application of the present invention can be said to speed up the pace of informatization of my country's petroleum industry. It is expected that in the near future, the oil exploration and development technology will have breakthrough progress in artificial intelligence, and the present invention is one of its applications. The invention includes a set of software packages, which can be deployed on the cloud server for use by global customers, and is a complete application system combining petroleum data and computer modules.

In the process of exploration and development of petroleum, the identification and definition of lithofacies is of great significance. It can not only analyze the microfacies in lithofacies and their temporal and spatial evolution, establish a deposition model, but also analyze the accumulation of source, reservoir, cap, etc. elements and combinations, establish a reservoir-forming model, further explore the relationship between oil and gas accumulation and sedimentary microfacies, interpret the distribution of known sand bodies, guide sand body prediction in uncontrolled areas, and help determine geological reserves, predict oil and gas areas and Provide the basis for well location deployment.

Lithofacies analysis is an important step in interpreting seismic data for reservoir descriptions. Lithofacies interpretation plays an important role in the initial exploration prospect evaluation, reservoir characterization, and the final stage of field development. A lithofacies is a stratigraphic unit or region with characteristic reflection patterns that can be distinguished from other regions. Regions of different lithofacies are often delineated using descriptive terms that reflect large-scale seismic patterns. Such as reflection amplitude, continuity and the internal configuration of the reflector bounded by the stratigraphic horizon. The application and scale of lithofacies analysis varies widely, from watershed-wide applications to detailed reservoir characterization. At the basin scale, lithofacies analysis has been applied to the study of hydrocarbon systems to broadly identify sources, reservoirs, and seal-prone areas for exploration. These regions are usually identified based on their reflection geometry as well as amplitude strength and continuity. Regional high-amplitude, semi-continuous reflectors are often used to identify potential hydrocarbon-bearing reservoirs, such as deepwater channels, while low-amplitude continuous to semi-continuous regions can be used to identify seal-prone cells.

Petrofacies analysis can also be applied to a single reservoir to help constrain detailed physical characterization. In these local-scale applications, the definitions of continuity and amplitude are often not strictly defined, and the environment is calibrated based on rock properties or sedimentary interpretation. Relationships between seismic characteristics and physical properties can be demonstrated, and lithofacies volumes can then be used to predict rock property distributions and conditional geological models.

The standard technique for petrographic analysis and mapping is a manual process in which a seismic interpreter makes visual decisions about the characteristics of seismic reflection data within an interval of interest and plots this data on a map. The petrofacies are then used for various purposes, but mainly to explain the distribution of petrofacies and rock properties. Intuition and experience make important contributions to the success of petrographic studies, however, this approach can also lead to petrographic analysis being a subjective, time-consuming, and often laborious, inefficient labor. Several related techniques have been used in the petroleum industry to improve automation and enhance the interpretation of petrofacies from seismic data.

R. J. Matlock and G. T. Asimakopoulos, "Can Seismic Stratigraphy Be Solved Using Automatic Pattern Analysis and Recognition" (The Leading Edge, Geophys Explor, Vol. 5, no. 9, pp.51-55, 1986), which established a concept Framework for training algorithms to automate the seismic interpretation process. However, these authors did not demonstrate any working prototypes or describe any details of possible properties or classification algorithms. "Seismic Texture Classification: A Computer-Aided Stratigraphic Analysis Method" by R. Vintner, K. Mosegaard et al. (SEG International Exposition and 65th Annual Meeting, Paper SL14, October 8-13, 1995) and R. Vintner, K. Mosegaard, Abatzis, C. Anderson, VO Vebaek, and PH Nielson (3D Seismic Texture Classification, Society of Petroleum Engineers 35482, 1996), discussing texture analysis of seismic data and classification using texture attributes A version of principal component analysis and probability distributions . These publications do not utilize probabilistic neural networks or dynamically use probabilistic values to optimize classification when using texture analysis methods for seismic data. These methods also did not use an interactive training program, and texture analysis was not guided. The process of guiding computations through formation layers defined by the dip angle of the seismic reflector is called dip steering. D. Gao, "First- and Second-Order Seismic Textures: Implications for Quantitative Seismic Interpretation and Oil and Gas Exploration" (1999), describes the use of standard texture analysis to generate seismic texture properties that quantify reflection intensity, continuity, and geometry. However, this summary does not describe the classification method of texture attributes. Specifically, Gao (1999), does not use probabilistic neural networks, nor does it use an interactive interpreter for neural network training.

SUMMARY OF THE INVENTION

The present invention is aimed at the defects of the prior art, provides an intelligent petrographic identification method based on logging information and geological information, and can effectively solve the problems existing in the prior art. The present invention is composed of three major technical systems:

System 1. Petroleum professional system;

System 2. Website webpage creation system;

System 3. Data exchange system between the two;

The problem to be solved by the present invention is the problem of lithofacies identification in petroleum industry, shale gas industry and geological mining industry. The invention uses the artificial intelligence method to automatically identify petrofacies.

The invention is a method of identifying well logging data sets and seismic data sets by means of artificial intelligence. First, a number of initial texture attributes representing the amount of seismic data are calculated. Next, a probabilistic neural network is constructed from the computed initial texture properties. The final texture properties are then calculated across the entire seismic data volume. Finally, the computed texture attributes are classified using the constructed probabilistic neural network. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention is a method for lithofacies identification and prediction. The method is also applicable to other log attributes and seismic attributes, especially to identify microfacies in seismic amplitude data. Thus, the analysis of seismic texture mimics the vision-based analysis process of a seismic interpreter in a way that traditional attribute analysis does not. The interpreter translates a set of traces into an image to present the classification. This different analysis method offers the potential to capture reflection geometry across the entire study area. One such technique for quantizing image texture in texels employs an image transformation that results in a gray-level co-occurrence matrix. The grayscale co-occurrence matrix describes the spatial relationship between pixels in a small area within a larger image, ie texels. In practice, the gray-level co-occurrence matrix is computed in overlapping texels so that any transitions between texture classes within the entire image can be fully observed. Overlapping texels scans the image until the entire image is processed. The gray level co-occurrence matrix is a matrix of size N×N, where N is the number of gray levels used to quantize the image. Texture analysis, which constructs a grayscale co-occurrence matrix from image texels, is actually a two-dimensional (or three-dimensional) extension of one-dimensional Markov chain analysis. The structure of the seismically derived grayscale co-occurrence allows for a heuristic understanding of the matrix. In a homogeneous region, defining uniformity or continuity in a given direction, the difference between pixel values will be low, so elements close to the diagonal of the gray-level co-occurrence matrix will have higher values. Less uniform regions will yield higher differences between adjacent pixel values, so the resulting gray-level co-occurrence matrix will have values further away from the diagonal. The average pixel value is also represented in the grayscale co-occurrence matrix. Low-amplitude regions have gray-level co-occurrence matrices with values clustered around the center. On the other hand, regions with higher amplitudes have more distributed gray-level co-occurrence matrix values, either along a diagonal continuous texture or across the entire gray-level co-occurrence matrix in more discontinuous textures.

In order to approximate the process followed by the seismic interpreter, 2D texture properties are optimized for computation, which are then filtered in time slices to mimic full 3D operation. Alternatively, 3D texture properties can also be computed and used to characterize petrofacies. Gray-level co-occurrence matrices cannot be explained directly and efficiently, and are described more efficiently by scalar statistical measures of texture properties. Texture attributes can be divided into first-order and second-order descriptors. First-order statistics quantify the global distribution of pixel values within an image and can be computed directly from texels using standard statistical techniques, even without intermediate gray-level co-occurrence matrix transformations. The mean absolute magnitude and standard deviation of magnitude values within a texel are examples of first-order texture properties and can be used to delineate magnitude anomalies and reflection strengths. Derived properties such as instantaneous magnitude, phase, and frequency can also be used to generate first-order statistics. First-order statistics are the starting method for detailed texture quantization, although some geophysical regions can be roughly defined from different pixel intervals. In general, the value of a single texel cannot be adequately described in terms of its first-order statistics alone. For example, high-amplitude chaotic regions of a seismic image are not necessarily separated from continuous regions of high or even moderate amplitude using only average amplitude values. The second-order statistics of an image quantify the spatial relationship of pixels within the image and compute it to a gray-scale co-occurrence matrix through intermediate transformations. Second-order statistics, statistics of the gray-scale co-occurrence matrix, capture trajectory shape features, reflection geometry and reflection continuity, and amplitude intensities. The second-order statistics of texels are multi-trajectory image properties that allow the capture of reflection geometry and continuity by analyzing tilt-controlled gray-level co-occurrence matrices.

These gray-level co-occurrence matrices correspond to textures of organized and differentially contrasted features with only a few gray levels at the same distance and azimuth. Lower texture uniformity values will correspond to larger values of the grey level co-occurrence matrix further away from the diagonal of the matrix, i.e. many different grey levels at the same distance and azimuth. These features make texture uniformity particularly useful for quantifying continuity. The first texture attribute, texture correlation, represents the similarity of the elements of the metric space grayscale co-occurrence matrix in the row or column direction, and the magnitude of the correlation value reflects the correlation of local grayscales in the image. When the matrix element values are uniform, the correlation value is large, and when the matrix element values are very different, the correlation value is small. The second texture property, texture inertia, represents the contrast of the grayscale co-occurrence matrix and is the opposite measure of texture homogeneity. For high-contrast images, texture uniformity is low and texture inertia is high. The third texture property, texture entropy, measures the lack of spatial organization within the computational window. When all elements of the gray level co-occurrence matrix are equal, the texture entropy is high, corresponding to rough textures, and low textures are more uniform or smoother. A fourth texture attribute, texture energy, also represents the organization within the spatial computation window. In this case, it is also possible to compute all or most of the gray levels within the window, which are characteristic of rough textures. Conversely, the highest values of texture energy indicate the presence of high-valued grayscale co-occurrence matrices. In this case, only a few gray levels prevail. The area within this calculation window is more uniform, or exhibits some regular characteristics.

Neural networks are interconnected assemblies of simple processing elements. Through the process of adapting or learning a set of training patterns, the processing power of a neural network is stored in the obtained connection strengths or weights. One advantage of neural networks is the ability to train or modify the strength of connections within the network to produce desired results. In classification applications, neural networks can be thought of as a special case of supervised classification schemes, since the training of neural networks is a supervised exercise. Once sufficiently trained on multiple calibration images, the neural network can be applied to the remaining images in the data volume. In the calculation, the connectivity, weight, and input vector of the modified attributes of the nodes in the general neural network are passed to the next layer of the network through the modified values. Through training, the weights of the network are modified so that in a specific set of training examples, the modification of the input attribute vector produces the desired result. The training of the network and the modification of the connection weights result in the generation of decision surfaces for the network.

One of the advantages of neural network algorithms over more standard classification schemes is the ability to generate nonlinear boundaries. A typical classification or prediction problem usually has only three layers, the first is the input layer; the second is the "hidden" layer; and the third is the output layer. A probabilistic neural network is a parallel implementation of a standard Bayesian classifier. Probabilistic neural networks are three-layer networks that can efficiently perform pattern classification. Mathematically, these probabilistic neural networks are very similar to kriging, where proximity to known points guides the classification and prediction of unknown points. In its standard form, a probabilistic neural network, is not trained in the same way as the more traditional neural networks described above. Instead, the training vector simply becomes the weight vector in the first layer of the network. This simpler approach enables probabilistic neural networks, with the advantage of not requiring extensive training. For example, in seismic texture analysis, the texture attributes of training images provide weight vectors in the first layer of the network. This has significant speed advantages during the training phase compared to traditional types of neural network architectures, such as fully connected backpropagation architectures. Furthermore, probabilistic neural networks tend to generalize well, whereas more traditional networks, even with large amounts of training data, are not guaranteed to converge and generalize to data not used during the training phase. When an input pattern is presented to a probabilistic neural network, the first layer of the network (the input layer) computes the distance from the input vector to the training input vector and produces a vector whose elements indicate how close the input is to the training input. The second layer is to sum these contributions of each class of input to produce a probability vector as its net output, which is another advantage of using a probabilistic neural network. In addition to classifying the maximum probability from the third or output layer, this is the ability to directly extract the classification probability from the second or hidden layer.

Computing the window size for the entire volume using a single gray-level co-occurrence matrix causes this negative effect. Improve results by changing the window size for the entire volume. As the data frequency decreases with depth, the window size becomes larger. This mode is used in conjunction with a dynamically adjusted window size based on a user-defined confidence level. In another alternative embodiment for dealing with degraded seismic data quality, the data may first be filtered with a convolution or median filter to smooth the data prior to input.

If the confidence level falls below a user-defined level, the calculation window size may be automatically adjusted until the confidence level rises above. In another alternative embodiment, the generation of the gray-level co-occurrence matrix may be subject to tilt control, and the generation of the gray-level co-occurrence matrix may be re-scaled accordingly. Calculate and reclassify phases. Stratigraphic framing is an important aspect that seismic interpreters have been taking into account, albeit unintentionally of a specific geological setting. Lithofacies interpreters do not only consider the continuity of the time plane, but they also judge the continuity of the strata after the strata defined by the dip of the seismic reflector.

Texture analysis and construction of the grayscale co-occurrence matrix of texels depend on the viewing direction or azimuth, where pixels within a texel are correlated. Texture analysis applied to seismic data is extremely sensitive to the stratigraphic frame of the texel and must also follow the reflector's stratigraphic dip to properly simulate the process performed by human translation. Maximize the continuity of the image after the formation dip in the grayscale co-occurrence matrix calculation, as represented in the gray-scale co-occurrence matrix. The process of guiding calculations by the formation dip is called dip steering. Texture analysis needs to be highly resolved in the stratigraphic geometry to correctly grasp the computation of the grayscale co-occurrence matrix. To obtain the required resolution, the multi-traces of texels, images, properties, and depressions within the image are estimated by gradient-based techniques.

The invention organically combines the processing module of petroleum data with artificial intelligence technology and website webpage display technology to form a unified system. The present invention uses artificial intelligence for lithofacies classification as a method to help customers process their logging data and seismic data. It realizes the remote quick entry of customer data, the model operation on the shared server, and the independent display and distribution of the calculation results, which improves the calculation speed and user experience.

Accurately identifying underground oil and gas reservoirs, especially thin oil and gas reservoirs with a thickness of less than 3 meters, and improving geological reserves are the work and main challenges of researchers in geology, exploration, development, and well logging. The technique currently used is to study petrophysical discontinuities along the wellbore by studying sharply varying petrophysical data. In the invention software, an artificial neural network (ANN) model is included, which can predict the feature vector of the oil reservoir by inputting 10 logging parameters as input. Through learning and training, it can be considered as detecting the identification of the oil reservoir. standard. The software of the invention is a support vector machine using a machine learning algorithm, and assigns petrofacies to logging data for training. Based on the core description of experts and the logging data of about 3 to 9 wells, model training is carried out. This data is used to train support vector machines and to intelligently identify lithofacies based on logging data.

The use of artificial intelligence for reservoir identification will greatly reduce the workload of geological, exploration, and development researchers, greatly improve the accuracy and efficiency of reservoir identification, and achieve end-to-end docking. The convolutional layer and sampling layer of the algorithm convolutional neural network in artificial intelligence are like cooperation between partners, one extracts features, and the other performs local non-overlapping sampling. Repeatedly, the reservoir identification results will become more and more accurate.

The essence of oil and gas reservoir identification is the identification and evaluation of rock pore fluid properties and saturation. The volume and mass of reservoir pore fluid only account for a very small part of the reservoir rock, and it is filled in the pores of the solid rock framework; The logging method can be discriminated, but the seismic response is very weak. Seismic records, if responsive to changes in rock pore fluids, are only likely to reflect the small structure of seismic events. The wave equation describing the propagation of seismic waves is an approximate equation obtained under certain assumptions (such as a completely elastic medium). Seismic records are an objective reflection of the response of the actual geological medium without any approximation. If the magnitude of the seismic response of rock pore fluids is observable, then it must be present in the seismic record. The crux of the problem is how to identify the pore fluid response on the seismic record. Deep learning in artificial intelligence can perform automatic feature extraction, and the software of the present invention applies deep learning to the field of seismic exploration.

The purpose of the software of the present invention is to replace the manual judgment of the position of the oil and gas layer, because the human eye will be tired for recognition, and the phenomena of missed judgment, wrong judgment and inaccurate interpretation often occur. At the same time, it is necessary to synthesize more than a dozen curves at a time to judge the position of the oil layer by using the logging curve, and it is sometimes impossible to identify it manually.

Because the exploration reservoirs are deeply buried underground, in the analysis of underground lithofacies, only through logging and rock data can we analyze and observe the characteristics of the reservoirs and the signs of sedimentary facies. Drilling coring is generally discontinuous, and the whole-well coring rate of an exploratory well is usually only a few percent to ten percent, which brings great difficulties to the study of sedimentary facies. Although logging facies analysis using logging data can provide continuous sedimentary facies interpretation for the whole well, it does not fully utilize important information such as formation stacking mode and sedimentary body shape. Even if the interpretation is completely correct, it is very local. In order to further grasp the plane distribution characteristics of sedimentary facies, it is necessary to have a large amount of sufficiently dense drilling data, which is difficult to meet in the exploration stage. Therefore, there is a need for a new method and method that can better grasp the variation characteristics of the sedimentary facies plane with only a small amount of drilling data. The identification and prediction of lithofacies are produced to meet the above urgent needs.

Lithofacies refers to a three-dimensional seismic reflection unit within a certain distribution range. The seismic characteristic parameters (such as reflection structure, geometric shape, amplitude, frequency, continuity, etc.) in this unit are different from those of adjacent units. The lithological assemblage, bedding and sedimentary characteristics of the sediments. Lithofacies prediction is to identify and map the lithofacies units according to certain procedures according to the seismic characteristic parameters and other downhole and surface data within the sedimentary stratigraphic unit, which lays a necessary foundation for comprehensive interpretation of the depositional environment and depositional system. Seismic data, which is necessary for lithofacies prediction, are essential basic data in petroleum exploration. They can be obtained in the early stage of exploration and generally cover the entire work area, and contain extremely rich stratigraphic, structural and sedimentary facies information. The identification and prediction of lithofacies in the present invention is to carry out regional stratigraphic interpretation, determine sedimentary system, lithofacies characteristics and interpret depositional development history, and finally convert lithofacies to sedimentary facies, which is used as the basis for studying petroleum geological sources, reservoirs and caps. Based on the combination and distribution law, the favorable oil-generating area and reservoir facies belt can be predicted.

The key to lithofacies prediction lies in the accurate division of lithofacies. The traditional method of lithofacies division is to observe images, that is, to observe and describe the reflection characteristics on the seismic section with the naked eye, and classify lithofacies with similar characteristics into one category. This purely manual method is time-consuming, labor-intensive, and highly subjective, which is not conducive to the identification of abnormal reflection features that are not prominent on the seismic profile. With the continuous improvement of seismic data acquisition technology, the seismic information contained in the seismic section is more abundant, many of which cannot be detected by naked eyes, and must be extracted and analyzed with the help of seismic data processing technology and computer technology. At first, people used seismic structure attributes to divide lithofacies, but the method of extracting seismic structure attributes was not mature at that time, and the division results were limited by the signal-to-noise ratio of seismic data. Later, people used the gray-scale co-occurrence matrix to extract the seismic structure attributes, described the amplitude distribution of the seismic data from a statistical and mathematical sense, and improved the accuracy of using the structural attributes to divide the lithofacies. However, after all, the seismic structure attributes only represent several physical parameters of the seismic signal, and the description of the overall abnormality of the seismic signal is still lacking. In recent years, artificial intelligence neural network technology has been introduced into the division of lithofacies. The adaptability, fault tolerance and large-scale parallel processing ability of neural network further improve the accuracy of petrofacies division. At present, the methods of using neural network to classify lithofacies can be classified into two categories: one is unsupervised pattern recognition; the other is supervised pattern recognition. The unsupervised pattern recognition divides the lithofacies in the work area based on the input data and the classification number set in advance by the interpreter; the supervised pattern recognition adds drilling data as control information in the classification process, so that the lithofacies division results have a clear direction properties, such as a certain lithofacies representing a certain lithology or a hydrocarbon-bearing area.

Two common petrofacies prediction techniques are described below, which are based on unsupervised neural networks and supervised neural networks, respectively.

The first is lithofacies identification and prediction technology based on seismic waveform classification. Because different sedimentary environments will form different sedimentary bodies, different sedimentary bodies are different in terms of lithology, physical properties and oil content, which is reflected in the seismic information as changes in the amplitude, frequency and phase of seismic waves, that is, seismic waveforms The change. Therefore, the self-organizing feature mapping neural network can be used to automatically identify and classify the seismic trace waveform and its reflected geological features, so as to complete the identification and prediction of lithofacies. The technology mainly includes three steps: the first step is to use the self-organizing neural network in the target interval to learn and train the actual seismic traces for the entire work area, and obtain a series of seismic traces that can reflect the changes of the seismic traces in the interval. Model tracks, these model tracks are arranged in a gradient of shape, each model track represents a type of lithofacies, and is assigned a color and a number in sequence; the second step is to assign each actual earthquake in the target interval of the whole work area Compare the actual seismic trace with the model trace, classify the actual seismic trace into the category of the model trace with the highest degree of correlation and assign the corresponding color and number number; the third step is to draw the target interval according to the different color and number number. lithofacies. So far, the lithofacies prediction of the target interval has been completed.

The second is lithofacies prediction technology based on seismic attributes and multilayer perceptron neural network. Seismic attributes contain rich geological information such as stratigraphic structure, lithology, and physical properties. Using them to drive a multilayer perceptron neural network can accurately predict the lithofacies in the study area when drilling is limited. Early people used this technique to successfully predict lithofacies. The technology includes the following steps: the first step is to extract seismic attributes that can clearly describe the lithofacies characteristics of the target interval, such as seismic amplitude attributes, energy attributes and coherent body attributes; the second step is to use seismic attributes as multi-layer The input of the perceptron neural network, the drilling data is used as the control point information, and the error back propagation algorithm is used to train the network. Prediction of lithofacies in this interval.

The lithofacies prediction technology based on seismic waveform classification utilizes self-organizing neural network to classify lithofacies. Its disadvantages are as follows: 1. The number of classifications of lithofacies needs to be preset manually, which often leads to inaccurate setting of classification numbers, and generally requires multiple calculations to estimate this parameter; 2. It requires multiple iterations to make The classification results converge to accurate results. In practical applications, in order to ensure the best network convergence, 20 to 40 iterative operations are usually required; 3. Self-organizing neural networks are unsupervised pattern recognition methods, and the geological significance of the classification results is unclear. , and further interpretation is required in combination with drilling data to obtain a lithofacies map with clear indication.

Lithofacies prediction technology based on seismic attributes and multilayer perceptron neural network. Its shortcomings are as follows: 1. The multi-layer perceptron neural network uses the error back propagation algorithm to train the network. This training method has a slow convergence speed and usually takes a lot of computing time; 2. When using the multi-layer perceptron neural network When the network performs lithofacies classification, it is usually done in one step, that is, the existing supervision information is divided into several types, and the lithofacies of the reservoir are predicted at one time, which often results in inaccurate classification. Because supervision information generally comes from geological and logging data, and their scales are different from seismic data, in general, the classification of supervision information may not necessarily correspond to the classification of lithofacies.

The purpose of the present invention is to provide a reservoir identification system based on artificial intelligence method. It is essentially a deep learning method using artificial intelligence to automatically extract features from logging curves and seismograms, and its manifestation is hierarchical feature extraction. Low-level features belong to local features, high-level features are nonlinear combinations of low-level features, and belong to abstract structural features. High-level features are more discriminative and category-indicative. The software of the invention innovatively introduces the feature extraction method of machine learning and deep learning, and extracts the micro-features and weak seismic response features of the reservoir logging curve, which can determine the reservoir features more simply and efficiently, and improves the utilization rate of the logging curve and the seismic exploration. Data recognition accuracy.

The core function of the software of the present invention is reservoir image recognition, which refers to the use of computers to process, analyze and understand images to recognize targets and technologies of various patterns. Image recognition is an important area of artificial intelligence. At present, the main image recognition methods include image recognition method based on neural network, image recognition method based on wavelet moment and so on.

In order to develop computer programs that simulate human reservoir image recognition activities, different reservoir image recognition models are proposed here.

1. Template matching model

This model believes that to identify a certain reservoir image, there must be a memory pattern of this reservoir image in past experience, also called a template. If the current stimulus matches the template in the brain, the image of the reservoir is identified. For example, in a three-dimensional (3D) seismic reservoir image, if the oil layer map has the feature of "two curves intersect and bend down", the big data model has such a template. If the seismogram has such a size, orientation, and shape In full agreement with this template, the oil layer is identified. This model is simple and clear, and it is easy to get practical application. However, this model emphasizes that the reservoir image must be completely consistent with the template in the big data to be recognized. In fact, the big data can not only identify the reservoir image that is completely consistent with the template in the human brain, but also can identify the image that is not completely consistent with the template. reservoir image.

2. Prototype matching model

In order to solve the problem of template matching model, a prototype matching model is proposed. This model argues that what is stored in long-term memory is not an infinite number of templates to be identified, but rather some "similarity" of reservoir images. The "similarity" abstracted from the reservoir image can be used as a prototype to test the reservoir image to be identified. If a similar prototype can be found, the reservoir image is identified. This model is more suitable than the template matching model both in terms of neural and memory search process, and can also illustrate the recognition of some irregular reservoir images that are similar in some aspects to the prototype. Generally used in the petroleum industry, three-dimensional (3D) seismograms are used, and then the software is used to process the color difference and grayscale difference of the pictures to identify useful oil and gas layer information.

The software of the present invention is based on cloud storage and cloud computing on the server side, data input on the web page, and result output on the web page. For reaching the above-mentioned goal, the scheme that the present invention adopts is as follows:

Artificial intelligence module, including machine learning module, deep learning module, plus Python web framework. The software system of the present invention includes a client and a server, and the client is connected with the server through a network, and the method comprises the following steps:

1) Open the webpage and enter the data entry interface;

2) Enter the data set, set the calculation parameters and module parameters;

3) Enter the model training interface, run the training module and debug the parameters, so as to improve the model learning efficiency;

4) After the model training is completed, enter the data processing interface to identify and divide lithofacies;

5) Display and distribution of the results, and add the training model to the training set to be used next time.

The present invention is realized through the following technical elements:

Log data refers to the physical properties of a portion of the destination formation, the present invention includes using a portion of samples with known lithofacies classification as training data, dividing the training data into two subsets, a calibration set and a cross-validation set, using automated supervised learning The lithofacies identification method, based on the calibration set to determine the identification and prediction of lithofacies in the subsurface destination layer, by comparing the predicted and observed lithofacies for the cross-validation set to calculate the confusion matrix for the supervised learning facies identification method , calculate a phase transition matrix characterizing the change between successive facies, and use the preliminary identification, the petrographic transition matrix, and the confusion matrix to iteratively calculate the identification of the petrographic facies.

Based on the deep learning method of logging curve data and seismic data, the calculation method is used to carry out the module operation, including the following steps:

(1) Use logging, well logging and synthetic seismic records to demarcate the target layer;

(2) After obtaining log data, specify and classify on multiple log samples. For example by using core descriptions. It will be appreciated that the classification may have been pre-specified, or may be specified through expert analysis as part of the implementation of the method. These designations are considered known phases. Portions of the log data with known facies are selected and removed, and set aside until further processing. That is, data with known phases is divided into a training subset and a test subset, where the test subset may be referred to as "ignored" or "intersection and validation" data. The ignored data can be randomly selected and the percentage of ignored data can be set by the user as a parameter or can be a fixed percentage. When there are many data samples, the percentage of ignored data can approach 50%.

The method is performed by implementing any conventional computer-implemented supervised pattern recognition or machine learning method for identifying phases and training using a training set. It will be appreciated that there are a wide variety of such methods, including backpropagation, neural networks, decision trees, and any number of additional supervised learning algorithms that can be applied to log data.

Once the machine learning method has been trained, it is used to predict the phase on all samples, including the ignored data, and by comparing the output of the trained machine learning algorithm with respect to the classification previously specified for those parts of the data Compare to generate confusion matrix.

A phase transition matrix is generated that characterizes the changes between previously specified phases in the log data. A preliminary predicted phase transition matrix is generated, which characterizes the changes between phases in the preliminary predicted classification.

The transition matrix describes each pair of continuous phases and the relationship of that pair to each other. For example, when successive pairs show the change from shale to sandstone, the transition matrix will capture the relationship as well as the change from sandstone back to shale, as shown in Figure 2.

Once the observed and initially predicted transition matrices are calculated, the target probability matrix can be formed. In this regard, target probabilities can be calculated based on predictions, or a transition probability matrix can be set strictly based on observed transitions.

The seismic data is extracted along the specified time window width of the target horizon as the training data of the deep learning pre-training model, wherein a single training data sample is formed by connecting the specified time window data of each surrounding track, and the moving distance of the time window is generally less than or equal to the length of the time window;

Pre-training deep learning model parameters using Restricted Boltzmann Machine (RBM) or Continuous Restricted Boltzmann Machine (CRBM);

Select the optimal model depth, the number of neuron nodes in each layer of the model, the neuron activation function and the sparse limit through experiments;

Extract well side channel seismic data along the target horizon with specified time window width as training data for deep learning fine-tuning model, the categories of fine-tuning model include oil, gas and water;

Fine-tune deep learning model parameters using batch stochastic gradient descent;

Calculate the base of each layer of the deep learning model, extract the seismic response value of the target layer, and use the correlation between the sample and the base to determine the deep learning target characteristics. Reservoir differences.

The well-passing or well-connected seismic data is extracted according to the training data extraction method, and the target features are obtained by inputting the well-passing or well-connected seismic data into the trained deep learning model.

According to the well data, the difference of seismic deep learning characteristics caused by different lithologies and fluids is determined, and then the different seismic deep learning characteristics caused by different lithologies and fluids are extrapolated to the area without wells, and then lithology and hydrocarbon detection are carried out.

The calculation of high-level nonlinear features of deep learning can be applied to two-dimensional and three-dimensional data, and the calculation methods are flexible and diverse.

Theoretical basis of the software: Deep learning in artificial intelligence forms the theoretical basis of this software. The oil and gas reservoir distribution prediction based on deep learning makes use of the difference in the sensitivity of oil and gas reservoirs between longitudinal waves and converted shear waves to effectively identify oil and gas reservoir characteristics and improve the accuracy of seismic reservoir distribution boundary characterization. The specific method is to combine a convolutional neural network with the support vector machine method, and use the Wright criterion to eliminate the possible outliers in the generated multi-wave seismic attributes and reduce the number of network variants. Then, a hidden layer is constructed through a clustering algorithm and an unsupervised learning algorithm that can highlight the characteristics of oil and gas reservoirs, which are used to increase network sharing and extract oil and gas characteristics. Finally, the known well point samples after adding the network penalty value are used as the input samples for SVM prediction, and the sampled convolution layer attributes are used as the learning set to predict the oil and gas reservoirs from known to unknown. Further, according to The intelligent prompting method for searching of the present invention is characterized in that the calculation of the statistical frequency of keywords in the step S511 includes the step of calculating the statistical frequency weighted by time.

Software realization function:

1) Oil and gas reservoir feature extraction from three-dimensional (3D) seismic reservoir images;

2) Modeling of oil and gas reservoir characteristics;

3) Triggering, realization and debugging of model parameters;

4) Reservoir image recognition training: fast learning and iteration;

5) Boundary processing of reservoir image recognition model;

6) Experimental application of automatic interpretation of machines;

7) Commercial application of reservoir interpretation of big data platform;

8) The work plan includes finding 2~3 oil companies for cooperation, so that the results can be quickly transformed into productivity

Application scenarios

With the rapid development of petroleum exploration and development and 3D seismic technology, the oil and gas reserves of domestic proven hidden oil and gas reservoirs have exceeded half of domestic proven oil and gas reserves. the exploration and development stage of hidden oil and gas reservoirs. The identification of hidden oil and gas reservoirs is difficult, there are many control factors, and the law of accumulation is complex, so the knowledge of geophysical exploration technology and oil reservoirs is relatively high. Reservoir geophysical technology organically combines seismic exploration technology and oil reservoir research technology, which meets the requirements of vertical resolution and plane resolution for understanding of oil reservoirs. Production is more targeted. Most of the hidden oil and gas reservoirs discovered in the early years of exploration have become old oilfields in the middle and late stages of development. They are rich in production materials, high in recovery, and water cut rises rapidly. It is urgent to make up for the declining production through intensification adjustment and edge expansion. Therefore, the re-understanding of hidden oil and gas reservoirs that have been exploited for many years and the research on the distribution of remaining recoverable reserves are new areas of seismic identification technology. The distribution characteristics of remaining reserves in old oilfields guide the deployment of production wells to further increase production and tap potential.

In oil exploration, reservoir prediction is one of the key tasks in finding oil and gas resources and evaluating oil and gas reserves. Due to the complexity of the underground geological structure and the ambiguity of the distribution of logging parameters, the traditional lithology identification methods have limited identification accuracy, and the interpretation results are often unsatisfactory. The main reason is that when logging, the wellbore curve is drawn by geophysical methods such as resistivity, ultrasonic waves, radioactivity, and nuclear magnetic resonance, and then the comparison curve is used to judge whether there is oil and gas in a certain layer in the well, which will inevitably lead to judgment deviations, omissions, and even mistakes etc. Later, with the help of computer-aided calculation and logging data to evaluate formation parameters such as formation lithology, electrical properties, porosity, saturation and permeability, the accuracy has been improved. The lithology identification results play an important role in finding oil and gas reservoir resources. . Due to the many factors that affect the measured value of the logging curve, and the lithology of the stratum varies greatly in different regions, the lithology is various, and the geological structure is complex, etc., the existing identification methods have the disadvantages of low identification accuracy, and it takes time to build different models in different regions. Therefore, how to accurately identify the lithology becomes a key issue in logging processing.

On the other hand, a problem plaguing the field of oil exploration and development is the precise interpretation of 3D seismic images. First, the interpretation of reservoir images ultimately depends on manual work, and manual judgment depends on its expert experience. Expert experience is the reflection of the human brain. How to map the experience of identifying oil layers in the human brain with computer models This is one of the goals of this project. On the other hand, artificial judgment of the location of oil and gas layers will often lead to missed judgments, wrong judgments and inaccurate interpretations due to the fatigue of human eye recognition. The third aspect is to use the logging curve to judge the position of the oil and gas layer, and it is necessary to synthesize as many as a dozen curves at a time. There is also such a problem with manual identification.

With the emergence of advanced technologies such as big data, cloud computing, and artificial intelligence, especially the successful application of face recognition, a new way has been opened up for machine identification of oil and gas layers in seismic reservoir images. The basic principle is: use the existing oil and gas reservoir images, store their reservoir image features as the same as a single face into the data model, and then let the machine automatically find similar oil and gas in the customer's 3D seismic reservoir images. This processing and interpretation process is not only fast but also accurate; at the beginning, manual inspection can be added to let the machine continue to learn and iterate repeatedly, in order to reach the ideal state. Through this processing method, the disadvantages of unstable and inaccurate manual interpretation can be overcome, and the application of artificial intelligence technology in the oil field can be realized. At present, the oil industry urgently needs such software to be practically implemented.

The specific main application scenarios are:

First, the geological conditions are complex, and the structural interpretation is difficult. Most of the exploration areas have complex underground structures, fault zones and small fault blocks are developed, and the underground velocity changes rapidly laterally; the reservoir lithology is complex, the potential reservoir types are diverse, and the types of gas reservoirs It is complex and diverse, and it is difficult to describe in detail.

Second, the utilization rate of the existing seismic data is low. Due to the limitation of software and hardware conditions, the high-precision seismic data collected by customers at a high cost can only be used for structural interpretation and simple post-stack reservoir prediction, especially for low-permeability oil and gas reservoirs. However, the collected high-precision and all-round 3D seismic data cannot independently and effectively carry out pre-stack reservoir and fracture prediction, which seriously restricts the exploration process of low-permeability blocks. level of data waste.

The third is that seismic exploration projects are limited by professional seismic software and professional technology. It is difficult for exploration and deployment personnel and decision makers to grasp more real, reliable and scientific evaluation data at the first time, which increases exploration risks to some extent. At the same time, customers’ scientific research The personnel also lose the opportunity to further study and master the core exploration technology, which restricts the establishment of the core technology system of the client's oil and gas exploration assets.

The present invention can build an exploration and development geophysical integrated platform, which is of great significance to the construction of the customer's upgrading technology system. Post-stack seismic inversion and reservoir prediction, determine the distribution range of favorable reservoirs, and propose exploration and development targets. It can excavate the abundant geological information of the existing massive seismic data to the greatest extent, effectively reduce the risk of oil and gas exploration and development, and contribute to the exploration of geological reserves for customers.

Necessity and needs analysis

The invention can carry out structural interpretation and comprehensive evaluation of oil and gas exploration and development projects. It can be built into a comprehensive integrated platform of comprehensive geological analysis-structural interpretation-reservoir prediction-physical property prediction-fluid detection-industrial mapping-target optimization, with depth domain interpretation capabilities.

The advantages of the present invention are:

(1) An integrated data management platform can be built to provide the functions of creating a work area and integrating data management and analysis. The resource tree mode is used to manage various data, which can query data conveniently and quickly, and supports both Chinese and English. Data loading has wide adaptability and high intelligence, and supports the loading of well data from Excel tables.

(2) Seismic data optimization processing toolkit. The platform provides a variety of seismic data optimization processing tools. Provides high signal-to-noise ratio or high-resolution seismic profiles for seismic interpretation.

(3) Wavelet estimation and well-seismic calibration. It has a variety of advanced wavelet estimation methods to help users extract the most ideal wavelet to make synthetic records (vertical wells, inclined wells).

(4) Multi-well comprehensive geological interpretation. Using geological and well logging data, it helps geological engineers to perform comprehensive geological interpretation of single wells - continuous well analysis, comprehensive interpretation and mapping of connected well formations, reservoirs and sedimentary facies profiles; at the same time, it provides automatic lifting of dominant facies and sedimentary cycle division Oil test data display analysis tools and formation flattening tools.

(5) Rock physics analysis and pre-stack seismic forward modeling. Provides single-well attenuation coefficient curve, pre-stack elastic and viscoelastic seismic forward modeling results, and seismic forward modeling gather attribute analysis results; at the same time, it can perform 2D petrophysical model and fluid replacement petrophysical simulation, 2D pre-stack incident angle gather forward modeling and analysis of its properties.

(6) Construction of virtual logging data. Based on the prestack angle gathers and raw velocity field data, a genetic algorithm is used to create virtual well curves to simulate P-wave velocity, shear wave velocity and density. It provides well constraint conditions for realizing seismic inversion under the condition of no well or few wells, and achieves the purpose of predicting reservoir.

(7) Structural explanation. It can directly carry out 2D and 3D forward and reverse fault structure interpretation and direct plane mapping of forward and reverse faults. It supports single-point multi-time value structure. One cdp can take multiple points, that is, a layer name can contain two or more time values. The structural plan of the upper and lower walls of the reverse fault is drawn at one time. At the same time, it has the ability of multi-user collaborative interpretation and deep domain interpretation.

(8) Speed modeling and deep conversion. It has velocity modeling under the constraints of structural information; the generated velocity field not only retains the vertical velocity variation uphole, but also has the characteristics of the lateral velocity variation of the velocity spectrum; it has the functions of correction and multi-directional quality control analysis. It can have block filling and block interpolation speed modeling methods, including inverse distance weighting, kriging interpolation and co-kriging three interpolation modeling methods. Through this fine velocity modeling method, the purpose of fine time-depth conversion of horizons, faults, seismic data, grids, and scatter points is achieved.

(9) Seismic inversion. Use well data, seismic data, and structural interpretation data to establish an initial seismic inversion model. It provides linear interpolation, fractal interpolation and collaborative kriging and other facies-controlled modeling algorithms, which can deal with complex structural systems of normal faults and reverse faults, and also consider the constraints of various depositional modes. The global optimization fast inversion algorithm is adopted, and high-resolution acoustic impedance/velocity data are obtained through iterative optimization for reservoir interpretation and reservoir parameter inversion.

(10) Explanation of lithologic bodies. This module provides a complete set of tools for fine lithologic interpretation and mapping of target geological bodies.

(11) Prediction of reservoir parameters. A variety of mathematical methods are used to carry out comprehensive prediction of multi-parameter reservoirs, which enriches geological target evaluation methods.

(12) Calculation and analysis of seismic attributes. It can provide high-resolution time-spectrum and instantaneous spectrum analysis algorithms based on wavelet transform and three-parameter wavelet transform, and calculate various seismic attributes such as energy attributes, attenuation attributes, frequency attributes, and instantaneous attributes, which can be applied to reservoir prediction and fluid detection.

(13) Prestack seismic inversion. Two pre-stack inversion methods are provided, namely pre-stack elastic wave impedance inversion and pre-stack extended elastic wave impedance inversion.

On the basis of prestack wave impedance inversion, elastic parameters such as longitudinal wave impedance volume, shear wave impedance volume, Lame coefficient, shear modulus, Poisson's ratio, elastic gradient, Young's modulus, and brittleness index can be obtained simultaneously. The prestack extended elastic wave impedance inversion part includes the correlation analysis between the extended elastic impedance curve in the well and the target curve and the determination of the optimal rotation angle, including the calculation of the optimal rotation angle projection seismic data, including the prestack extended elastic wave impedance modeling and Inversion results.

(14) Plane mapping and data analysis. It can provide rich seismic interpretation, plane analysis of reservoir data and powerful comprehensive map compilation functions. Through this function, data can be filtered, meshed, smoothed, etc.; at the same time, various industry-standard map compilations such as various structural maps, thickness maps, attribute distribution maps, and sedimentary facies plans can be completed; and Manage all kinds of scattered point data, grid data, contour data, survey area data, boundary data, and fault polygon data in layer mode; it can calculate distance and trap area, and also provide built-in snapshot tools.

Detailed ways:

This example demonstrates prediction of lithofacies from well log data. The dataset used here is log data from 9 wells that have been marked with lithofacies types based on core observations. We will use this test data to train a support vector machine to classify lithofacies types. A Support Vector Machine (or SVM) is a supervised learning model that can be trained on data to perform classification and regression tasks. The SVM algorithm uses the training data to best fit the different classes with a hyperplane. We will implement it using SVM in scikit-learn.

The first step is to organize the dataset: load the training data for 9 wells, create a cross plot to see the changes in the data, and use the cross validation set for model parameter selection.

The second step, build and adjust the classifier: Apply the trained model to classify facies in unlabeled wells, here the classifier is applied to two wells, in the future the classifier can be applied to any number of wells in the area.

Step 3, Implementation: The dataset is from 9 wells (with 8722 instances) and consists of a set of seven predictors and a lithofacies (class), each instance vector and validation (test) data (from two wells 830 examples) with the same seven predictors in the eigenvectors. The phasing is based on testing cores from nine wells tested vertically at half-foot intervals. Predictors include five from wireline measurements and two from geological knowledge. The seven predictors are: five line log curves including 1, gamma ray (GR) 2, resistivity log (ILD_log10) 3, photoelectric effect (PE) 4, neutron density porosity difference 5, mean neutron density Porosity (DeltaPHI and PHIND) Note that some wells do not have PE. These facies are not discrete but gradually merge with each other. Some adjacent adjacent phases. Mislabeling can be expected within these adjacent phases. The following table lists petrofacies, their abbreviated labels, and their approximate neighborhoods. Phase labels adjacent phases.

Description of drawings

In order to more clearly describe the embodiments of the present invention or the technical solutions in the prior art, it will become more apparent to those skilled in the art. The following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. The accompanying drawings in the following description are only some embodiments of the present invention.

Description Figure 1: System construction framework.

Figure 2 in the description: the module relationship of the machine learning part.

Description Figure 3: Lithofacies classification map: The dataset in this example is from 9 wells (with 4149 samples) and consists of a set of seven predictors and a lithofacies (class), each instance vector sums the validation ( test) data (830 examples from two wells) with the same seven predictors in the eigenvectors. Phases are based on inspection of cores from 9 wells tested vertically at 15.24 cm intervals. Predictors include five from wireline measurements and two from geological knowledge. These are basically continuous variables sampled at a 15.24 cm sampling rate:

The seven predictors are:

The five line-log curves include 1, Gamma Ray (GR) 2, Resistivity Log (ILD_log10) 3, Photoelectric Effect (PE) 4, Neutron Density Porosity Difference 5, Average Neutron Density Porosity (DeltaPHI and PHIND ) Note that some wells do not have PE:

Two geo-constrained variables: non-ocean-ocean index (NM_M) and relative position (RELPOS)

The nine discrete facies (rock classes) are: 1. Non-marine sandstone (FS) 2. Non-marine coarse siltstone (XFS) 3. Non-marine fine siltstone (XFS) 4. Marine siltstone and shale (CJ) 5 , Mudstone (Limestone MS) 6. Basalt Sandstone (Limestone WS) 7. Dolomite (NY) 8. Mud and granular limestone (Limestone PS) 9. Leafy algae reef (Limestone BS).

Figure 4 in the description: Well logging sample curve 1.

Figure 5 of the description: Well logging sample curve 2.

Figure 6 of the description: Figure 1 of the seismic sample.

Figure 7 of the description: Figure 2 of the seismic sample.

Figure 8 of the description: lithofacies classification diagram.

Although some implementation results of the present invention are shown, the content described is only a description method adopted to facilitate the understanding of the present invention, and is not intended to limit the present invention. Any person skilled in the technical field to which the present invention belongs, without departing from the spirit and scope of the present invention, can make any modifications and changes in the form and details of the implementation, but the scope of patent protection of the present invention, The scope as defined by the appended claims shall still prevail.

Claims (6)

1. The present invention discloses a web page display method for lithofacies classification with artificial intelligence, a method for interpreting logging curves and seismogram data with artificial intelligence methods, and the data entry and result output can use remote login web pages. In the form of cloud computing operation, cloud server deployment, cloud data clustering and cloud storage of computing results, the rights claimed in the present invention include the following contents. 2. The integrated method of combining the input and output of petroleum data with website pages, the composition and calculation method of each module of artificial intelligence, the display, storage and integration of data results and other methods of the present invention. 3. The method according to claim 2, utilizes the machine learning of artificial intelligence, the module calculation result of deep learning, and displays to the web (web page, website) end through various python frameworks, including constructing ideas, framework diagrams, modules and methods . 4. The method according to claim 2, wherein the combination technology and method of the Django framework and modules, as well as the output forms and usage skills of the website and web page display mode, color, layout, etc. 5. The method of claim 2, wherein the dataset may consist of seven features (five wireline measurements and two indicator variables) and phase labels at depth intervals, seven predictors (or input variables) variable) is: The five logs include: Gamma rays (GR) Resistivity Log (ILD_log10) Photoelectric Effect (PE) Neutron Density Porosity Difference (DeltaPHI) Average Neutron Density Porosity (PHIND) Two geological constraint variables: non-marine metrics (NM_M); relative position(RELPOS); In practical applications, some wells can be free of PE. 6. The method according to claim 2, the design process of the calculation model is: data entry->split data into training and test sets->preprocess data->build an activity code->build a vector stream ( Tensorflow)->build execution computation graph->build visual computation graph->make test data prediction->evaluate computation.

Images