Well logging data super-resolution method based on multi-view information integration

Abstract

The present invention belongs to the field of well logging data analysis and processing, and provides a well logging data super-resolution method based on multi-view information integration. The method first constructs a multi-scale low-resolution well logging data set, and then comprehensively uses three super-resolution processing technologies: fractal theory, neural network, and wavelet transform to process the data set from multiple angles. This processing process not only fully considers the temporal characteristics of the well logging data, but also effectively addresses the problems of data heterogeneity and local changes. The present invention effectively integrates the super-resolution processing results from different perspectives and constructs a unified super-resolution model. This method effectively overcomes the limitations of traditional technologies in dealing with complex geological environments, and provides a new technical means for the fine characterization of reservoirs.

Description

Technical Field

[0001] The invention belongs to the field of well logging data analysis and processing, and in particular relates to a well logging data super-resolution method based on multi-view information integration. Background Art

[0002] With the continuous development of oil and gas resources, oil and gas fields have gradually entered the exploitation stage of complex geological structures and unconventional reservoirs. In order to improve the exploration and development efficiency of oil and gas resources, it is particularly important to fully tap the fine reservoir information contained in logging big data. However, due to the complex geological environment downhole and the limitations of instrument resolution, the vertical resolution of traditional logging data is low, and it is difficult to capture subtle changes in the formation. It is not enough to cope with the fine characterization task of complex unconventional reservoirs. Therefore, improving the level of geological characterization analysis has always been a challenging task in the exploration and development of unconventional oil and gas reservoirs.

[0003] To address this problem, researchers have proposed a variety of vertical resolution enhancement techniques to improve the accuracy of reservoir characterization. For example, although deconvolution and resolution matching methods can improve resolution, they also increase the error between the numerical value and the true value in thin layers, limiting the feasibility of their large-scale application. In contrast, transform-domain-based super-resolution methods, such as the Walsh function method, invert the true value of the formation through parameter estimation, which not only improves the resolution, but also ensures that the error between the calculated value and the true value of the formation does not exceed the error of the logging data itself. However, since this method relies on a single preset curve pattern, its invariance and parameter selection limitations may lead to increased inversion errors. In addition, the windowed Fourier transform constructs a connection between high-resolution and low-resolution data by converting logging data from the spatial domain to the frequency domain, but this method easily amplifies outliers or errors in the original data to high-resolution results. In recent years, with the rapid development of machine learning technology, especially neural networks, have shown great potential in improving the vertical resolution of logging data, but existing methods are still insufficient in restoring local sharp structures and high-frequency information.

[0004] In order to address the limitations of the above existing methods and fully consider the temporal and spatial variation characteristics of logging data, it is necessary to effectively utilize the complementary information of logging data at different scales to cope with the heterogeneity of data and local geological changes. Therefore, a new logging super-resolution method is needed to fully mine the spatial and temporal information in logging data and achieve higher-precision reservoir characterization. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a logging data super-resolution method based on multi-view information integration to solve the problems in the prior art. The technical solution adopted by the present invention is:

[0006] The logging data super-resolution method based on multi-view information integration includes the following steps:

[0007] Step 1: Use the outlier detection method based on standard deviation and Gaussian filtering method to filter out outliers and noise information in the logging data set;

[0008] Step 2: Extract and analyze the self-similar features of the well logging data set using the re-standardized range analysis method;

[0009] Step 3: Upsample and downsample the low-resolution logging data set to construct a low-resolution logging data set;

[0010] Step 4: Combining the self-similar characteristics of the logging data set in step 2 and the low-resolution logging data set in step 3, different methods are used for super-resolution processing to construct a multi-view logging curve super-resolution model;

[0011] Step 5: Using an integrated machine learning method, the multi-view logging curve super-resolution model obtained in step 4 is combined with the original logging data set to construct an information integration model;

[0012] Step 6: Use the machine learning model to perform super-resolution processing on the logging data of unknown wells, and quantitatively evaluate the super-resolution results through restoration evaluation indicators.

[0013] The present invention has the following beneficial effects: The present invention proposes a new multi-view information integrated logging super-resolution method, which can effectively capture subtle changes under complex geological conditions and achieve more refined formation information extraction through multi-scale and multi-angle logging data analysis and the complementary advantages of multiple super-resolution methods, and is particularly suitable for unconventional reservoirs such as thin layers and fractures. The present invention combines multiple technologies such as fractal theory, neural networks, and wavelet transforms, which fully considers the temporal characteristics of logging data and can cope with the heterogeneity and local changes of data, thereby providing a more accurate and comprehensive reservoir characterization. This method not only significantly improves the resolution of logging data, but also overcomes the limitations of traditional methods in complex geological environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the process flow of the well logging data super-resolution method based on multi-view information integration according to the present invention;

[0015] Figure 2 This is a schematic diagram of the abnormal values ​​identified by the standard difference constant detection method for the GR curve of a certain experimental well;

[0016] Figure 3 A schematic diagram of low-resolution logging data of different scales constructed for the GR curve of a certain experimental well;

[0017] Figure 4A schematic diagram showing the comparison of super-resolution results obtained by three models: fractal interpolation, SRCNN network, and wavelet transform for a certain experimental well GR curve at a specific scale.

[0018] Figure 5 This is a schematic diagram of the final super-resolution result of the GR curve of a certain experimental well after model fusion;

[0019] Figure 6 This is a schematic diagram showing the comparison between the results of super-resolution processing of six conventional logging curves, namely GR, SP, DEN, AC, LLD and CAL, of a certain experimental well and the original data;

[0020] Figure 7 It is a schematic diagram comparing the super-resolution processing results of the GR curve of a certain well by the well logging data super-resolution method based on multi-view information integration described in the method of the present invention and the bicubic linear interpolation and random forest super-resolution method. DETAILED DESCRIPTION

[0021] The following will be combined with the embodiments of the present invention Figure 1-Figure 7 , the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0022] The present invention focuses on multi-view information integration technology, aiming to improve the recognition accuracy of reservoir characteristics under complex geological backgrounds, while solving the problem that existing super-resolution technology fails to fully utilize the complementary information between logging data of different scales. This method first constructs a multi-scale low-resolution logging data set, and then comprehensively uses three super-resolution processing technologies: fractal theory, neural network, and wavelet transform to process the data set from multiple angles. This processing process not only fully considers the temporal characteristics of the logging data, but also effectively addresses the heterogeneity and local variation of the data. The present invention effectively integrates the super-resolution processing results from different perspectives and constructs a unified super-resolution model. This method effectively overcomes the limitations of traditional technology in dealing with complex geological environments, and provides a new technical means for the fine characterization of reservoirs.

[0023] like Figure 1 , a logging data super-resolution method based on multi-view information integration, which includes the following steps:

[0024] Step 1: Use the outlier detection method based on standard deviation and Gaussian filtering method to filter out outliers and noise information in the logging data set;

[0025] Step 2: Use the rescaled range analysis method in fractal theory to analyze the self-similar characteristics of the well logging data set;

[0026] Step 3: Upsample and downsample the low-resolution logging data set to construct low-resolution logging data sets of different scales;

[0027] Step 4: Combining the self-similar characteristics of the well logging data set in step 2 and the low-resolution well logging data sets of different scales in step 3, different methods are used for super-resolution processing to construct a multi-view well logging curve super-resolution model;

[0028] Step 5: Using an integrated machine learning method, the multi-view logging curve super-resolution results obtained in step 4 are combined with the original logging data set to build an information integration model;

[0029] Step 6: Use the machine learning model to perform super-resolution processing on the logging data of unknown wells, and quantitatively evaluate the super-resolution results through restoration evaluation indicators.

[0030] Specifically, the method for filtering outliers and noise information from the data of a randomly selected well in step 1 is:

[0031] First, by calculating the mean and standard deviation of the data set, an outlier detection method based on standard deviation is used to identify and remove outliers in the well logging data that are outside a specific range:

[0032] ;

[0033] like:

[0034] ;

[0035] Then judge is an outlier, Figure 2 An example diagram of outliers identified from the GR curve of a certain experimental well is given.

[0036] in, Represents the total number of data points in the logging data set, Represents the value of a well logging data point, represents the mean of the well logging data set, represents the standard deviation of the well logging data set, Indicates a multiple, usually a positive integer.

[0037] After removing outliers, the Gaussian filter method is used to process the noise in the logging data set. The convolution operation weakens the high-frequency noise in the data while retaining the main features of the logging signal:

[0038] ;

[0039] in, represents the output value of Gaussian filtering, represents the standard deviation of the well logging data set, and Respectively represent the corresponding position coordinates in the well logging data set, specifically: Represents the horizontal coordinate. Indicates the vertical coordinate.

[0040] Specifically, the method for extracting and analyzing the self-similar features of the well logging data set in step 2 is:

[0041] The Hurst index H is calculated by rescaling range analysis method to determine the randomness, correlation and long-term trend of the logging data sequence:

[0042] ;

[0043] ;

[0044] in, Indicates the current measurement point data. Represents the total number of data points in the logging data set, Indicates the number of data points from the first to the The accumulated deviation of data points is Indicates The value of the well logging data point, represents the average value of the well logging data set, represents the range of the well logging data set, represents the standard deviation of the well logging data set, represents the re-standard range ratio, H represents the Hurst exponent, c Indicates independence from parameters The self-similar feature means that the shape and structure of the logging data show similar characteristics at different scales. In other words, no matter whether the logging data is enlarged or reduced, the local structure and the overall structure of the curve still show similarity.

[0045] The Hurst index of logging data can be calculated by formula (9). According to fractal theory, the Hurst index is used to characterize the randomness of time series, and its value range is 0 to 1. Different Hurst indexes reflect different properties and states of time series. When , it indicates that the logging sequence is completely random or unrelated, so the current data has no effect on future data; when Indicates that the logging sequence is a fractal sequence. Specifically, when When , it means that the logging sequence has negative correlation and exhibits a short-term memory effect, that is, the future change trend is opposite to the past. When , it indicates that the logging sequence has a long-term positive correlation, which means that the sequence will show the same change trend as the previous period for a long time in the future.

[0046] Specifically, the method for constructing low-resolution logging data sets of different scales in step 3 is:

[0047] The cubic spline interpolation method is used to upsample the well logging data set, which can well preserve the overall trend of the original well logging data set, avoid excessive oscillation, and improve the accuracy of the interpolation results:

[0048] ;

[0049] in, Indicates that in the interval The function after upper interpolation, represents the coefficient to be determined, Indicates adjacent logging data points. Different scales refer to different degrees of magnification or reduction of low-resolution logging data to generate data with different resolution levels.

[0050] In order to preserve the spatial correlation of the data and maintain the overall trend and geological characteristics while reducing the amount of data, the sampling kriging method downsamples the well logging data of complex geological structures:

[0051] ;

[0052] in, Indicates at location The predicted value at Indicates a known location The logging data value at Represents the weighting coefficient.

[0053] Figure 3 An example diagram of low-resolution logging data of different scales constructed for a certain experimental well GR curve is given.

[0054] Specifically, the method for constructing the multi-view logging curve super-resolution model in step 4 is:

[0055] According to the self-similar characteristics of the logging data set in step 2, the fractal characteristics of the data are determined. Based on the analysis results, three different super-resolution processing methods are selected: fractal interpolation, neural network and wavelet transform. These three methods can process the low-resolution logging data of different scales constructed in step 3 from different angles and integrate them into a multi-view logging super-resolution model.

[0056] The fractal interpolation method uses a given set of interpolation points and constructs an iterated function system to ensure that the image of these points converges to an attractor. Assume that there is a data set ,and , consider the iterated function system , where each function in the iterated function system has an affine transformation of the following form:

[0057] ;

[0058] The following boundary conditions are met:

[0059] ;

[0060] in, is the logging depth, is the logging data corresponding to the corresponding depth, real number is a free parameter, called the vertical scale factor. When , the iterated function system converges to a unique attractor.

[0061] From equations (11) and (12), we can get:

[0062] ;

[0063] According to fractal theory, if , and the interpolation point If they are not collinear, the fractal dimension of the iterated function system attractor satisfies:

[0064] ;

[0065] Where D represents the fractal dimension of the logging data.

[0066] Artificial neural networks consist of input layers, hidden layers, and output layers, each containing multiple neurons. Each neuron receives inputs from other neurons, multiplies these inputs by weights, sums them, and passes the results to the next layer. Each neuron generates an output signal through weighted summation and bias:

[0067] ;

[0068] in, is the neuron output, For time, is the connection weight, is the offset of the neuron, is the output transformation function.

[0069] Wavelet transform decomposes the signal into different frequency components by operating a series of discrete scale and translation factors while retaining time information. The scale factor and translation factor are usually recorded as:

[0070] ;

[0071] in, represents the scale factor, represents the basic scale factor, Indicates the scale level, represents the translation factor, represents the index of the translation factor, Represents a constant used to adjust the wavelet shift.

[0072] Assumptions , the wavelet function cluster is:

[0073] ;

[0074] Through this wavelet function, the logging data It can be decomposed into multi-scale frequency components and its wavelet coefficients can be calculated:

[0075] ;

[0076] Its inverse transform is:

[0077] ;

[0078] in, represents the wavelet function, represents the wavelet coefficients, represents well logging data, Is a constant that is independent of the signal or function.

[0079] In step 3, the well logging data set is first divided into multiple scales to generate low-resolution well logging data at different scales. Then, the above three super-resolution processing methods are applied to the data set at each scale. and treatment methods Under the combination of , the super-resolution result of multi-view logging is obtained, that is, the super-resolution model of multi-view logging:

[0080] ;

[0081] in, Indicates the depth Super-resolution logging data at Represents low-resolution logging data sets of different scales, Represents different super-resolution processing methods, Representation scale and super-resolution methods Acting on depth The super-resolution result obtained after the neighborhood data is processed. Indicates depth Neighborhood dataset.

[0082] The third step of the present invention is to construct low-resolution data of different scales, and the fourth step is to construct a multi-view logging super-resolution model. The multi-view logging super-resolution model specifically refers to the use of three different logging curve super-resolution methods in step 4 to perform super-resolution processing on data of three different scales respectively, and all the super-resolution results obtained are the results of the multi-view logging super-resolution model. Formula (25) is the formula of the multi-view logging super-resolution model.

[0083] Specifically, the method for constructing the information integration model in step 5 is:

[0084] In step 4, super-resolution processing is performed on logging data of different scales to generate super-resolution results under multiple models. In order to effectively fuse the super-resolution information of these different scales and models, a multi-view information integration model is constructed by adopting the method of deep multi-level scale estimation fusion:

[0085] ;

[0086] in, Indicates depth The results of the multi-view information integration model at represents the weight, Representation scale and super-resolution methods Acting on depth The super-resolution result obtained after the neighborhood data is processed. Indicates depth Neighborhood dataset.

[0087] Specifically, the method for quantitatively evaluating the super-resolution results in step six is:

[0088] The super-resolution model of multi-view information integration established in step 5 is used to perform super-resolution processing on unknown wells, and multiple quantitative evaluation indicators are used to evaluate the model performance and super-resolution effect. In order to ensure the generalization ability of the model on unknown wells, multiple indicators such as root mean square error (RMSE), mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and Pearson correlation coefficient (PCCs) are used for comprehensive evaluation. MAE measures the average difference between the predicted value and the actual value, reflecting the overall deviation of the model. The smaller the MAE, the higher the model accuracy; RMSE reflects the square root mean value of the error and is more sensitive to large errors. The smaller the RMSE, the better the model performance; PCC is used to evaluate the linear correlation between the predicted value and the actual value. The value range is [-1, 1]. The closer to 1, the stronger the correlation; PSNR measures the quality of signal reconstruction. The larger the value, the higher the prediction quality; and R² represents the model's ability to explain data variation. The closer to 1, the better the model fitting effect. The calculation formula is as follows:

[0089] ;

[0090] ;

[0091] in, is the total number of logging data samples, Indicates the depth axis The logging data value at the sampling point, Indicates the depth axis The result after super-resolution processing at the sampling points is: Represents the sampling data on the entire depth axis The mean of Represents the super-resolution result on the entire depth axis The mean of .

[0092] In order to verify the effectiveness of the method of the present invention, a certain logging data was used as the research object, and the proposed logging curve super-resolution method was experimentally verified. The experimental results show that compared with other single methods, the method of the present invention significantly improves the vertical resolution of logging data, and the restored logging curve can basically capture the subtle changes in the formation, which is of great significance for the subsequent fine characterization of complex reservoirs and oil and gas reservoir development decisions.

[0093] In terms of logging curve selection, six conventional logging curves, including acoustic time difference (AC), natural gamma (GR), density (DEN), natural potential (SP), well diameter (CAL) and deep lateral resistivity (LLD), were selected for super-resolution experiments. Figure 6 A schematic diagram comparing the super-resolution results of 6 conventional logging curves with the original data is given.

[0094] The present invention aims to address the problems that the existing super-resolution research on well logging data has failed to fully consider the time series and spatial variation characteristics of well logging data, as well as the heterogeneity of data and local geological changes. A super-resolution method for well logging data based on multi-view information integration is proposed. This method analyzes well logging data from multiple scales and angles by integrating multiple super-resolution models, making full use of its spatial, temporal and frequency domain information. First, a fractal interpolation model is used to mine the self-similarity characteristics in the well logging data; then, in combination with a neural network model, the nonlinear relationship and high-frequency details of the well logging data are processed by identifying historical data patterns; finally, the wavelet transform is used to decompose the logging signal into sub-bands of different scales to accurately capture the multi-frequency characteristics under complex geological conditions. The three complement each other, can overcome the limitations of a single model under specific conditions, and provide a more refined reservoir characterization.

[0095] At the same time, the use of BiLSTM to fuse different super-resolution models can effectively capture the time series characteristics in the logging data, and ensure the continuity and consistency of the vertical information through bidirectional processing. Its nonlinear processing ability helps to learn the complex relationship between different models, and at the same time, it copes with the non-uniformity and non-stationarity of the logging data through the memory mechanism, which is particularly suitable for fine formation characterization in complex reservoir environments. Compared with the traditional single use of neural networks, fractal interpolation and wavelet transform methods, the method of the present invention effectively integrates the advantages of multiple models, further improves the robustness and stability of super-resolution processing, especially in capturing subtle formation changes and reducing errors. Figure 7 An example diagram is given of the comparison of the super-resolution results of the GR curve of a certain well using the method of the present invention and traditional super-resolution methods such as bicubic linear interpolation and random forest.

[0096] The embodiments described above are only descriptions of the preferred modes of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations, modifications, and substitutions made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A super-resolution method for well logging data based on multi-view information integration, characterized in that: The following steps are involved: Step 1: Use the outlier detection method based on standard deviation and Gaussian filtering method to filter out outliers and noise information in the logging data set; Step 2: Extract and analyze the self-similar features of the well logging data set using the re-standardized range analysis method; Step 3: Upsample and downsample the low-resolution logging data set to construct a low-resolution logging data set; Step 4: Combining the self-similar characteristics of the logging data set in step 2 and the low-resolution logging data set in step 3, different methods are used for super-resolution processing to construct a multi-view logging curve super-resolution model; Step 5: Using an integrated machine learning method, the multi-view logging curve super-resolution model obtained in step 4 is combined with the original logging data set to construct an information integration model; Step 6: Use the machine learning model to perform super-resolution processing on the logging data of unknown wells, and quantitatively evaluate the super-resolution results through restoration evaluation indicators; In step 4, a multi-view logging super-resolution model is constructed using the following method: According to the self-similar characteristics of the logging data set in step 2, the fractal characteristics of the data are determined; based on the analysis results, three super-resolution processing methods are selected: fractal interpolation, neural network and wavelet transform; the low-resolution logging data of different scales constructed in step 3 are processed by these three methods and integrated into a multi-view logging super-resolution model: ； in, Indicates the depth Super-resolution logging data at Represents low-resolution logging data sets of different scales, Represents different super-resolution processing methods, Representation scale and super-resolution methods Acting on depth The super-resolution result obtained after the neighborhood data is processed. Indicates depth Neighborhood dataset.

2. The logging data super-resolution method based on multi-view information integration according to claim 1, characterized in that: In step 1, outliers and noise information in the logging data set are filtered out using the following method: The outlier detection method based on standard deviation is used to remove outliers in the well logging data set: ； like: ； Then judge is an outlier; in, Represents the total number of data points in the logging data set, Represents the value of a well logging data point, represents the mean of the well logging data set, represents the standard deviation of the well logging data set, Indicates multiples; Gaussian filtering is used to remove noise information in the logging data set: ； in, represents the output value of Gaussian filtering, represents the standard deviation of the well logging data set, and They represent the corresponding position coordinates in the well logging data set.

3. The logging data super-resolution method based on multi-view information integration according to claim 1, characterized in that: In step 2, the self-similar features of the well logging data set are extracted and analyzed using the following method: Restandardized range analysis method: ； in, Indicates the current measurement point data. Represents the total number of data points in the logging data set, Indicates The value of the well logging data point, represents the average value of the well logging data set, represents the range of the well logging data set, represents the standard deviation of the well logging data set, represents the re-standard range ratio, represents the Hurst exponent, c Indicates independence from parameters A positive constant of It indicates the extreme restandardization; H The value range is from 0 to 1. When , it indicates that the logging sequence is completely random; when , indicating that the logging sequence has a negative correlation: , indicating that the logging sequence has a long-term positive correlation.

4. The logging data super-resolution method based on multi-view information integration according to claim 1, characterized in that: In step 3, a low-resolution logging data set is constructed using the following method: Cubic spline interpolation and Kriging are calculated by the following formulas: ； in, Indicates that in the interval The function after upper interpolation, , , , represents the coefficient to be determined, , represents adjacent logging data points, Indicates at location The predicted value at Indicates a known location The logging data value at Represents the weighting coefficient.

5. The logging data super-resolution method based on multi-view information integration according to claim 1, characterized in that: In step 5, the information integration model is constructed using the following method: Using the super-resolution results of logging data of different scales in step 4 under different models, a multi-view information integration model is established by adopting the method of depth-to-multi-level scale estimation fusion: ； in, Indicates depth The results of the multi-view information integration model at represents the weight, Representation scale and super-resolution methods Acting on depth The super-resolution result obtained after the neighborhood data is processed. Indicates depth Neighborhood dataset.

6. The logging data super-resolution method based on multi-view information integration according to claim 1, characterized in that: The evaluation indicators for quantitative evaluation of super-resolution results in step 6 are root mean square error RMSE, mean absolute error MAE, peak signal-to-noise ratio PSNR, Pearson correlation coefficient PCCs and determination coefficient , the calculation formula is as follows: ； in, is the total number of logging data samples, represents the logging data value at the i-th sampling point on the depth axis, represents the result of super-resolution processing at the i-th sampling point on the depth axis, Represents the sampling data on the entire depth axis The mean of Represents the super-resolution result on the entire depth axis The mean of .

Images