An Adaptive Scale Coarsening Method for Logging-While-Drilling Curves

Abstract

The present invention discloses an adaptive scale coarsening method for logging-while-drilling curves, belonging to the field of logging-while-drilling data processing in oil and gas exploration and development, including: intelligently analyzing logging curves through deep learning technology and dynamically adjusting the coarsening strategy to achieve the collaborative optimization of data dimensionality reduction, noise filtering, and retention of key geological features. The present invention is applicable to multiple scenarios such as intelligent directional drilling, reservoir parameter calculation, 3D geological modeling, and well-seismic joint inversion, and involves technical fields such as signal processing, deep learning, and geological data analysis. The cross-integration of these fields makes the adaptive coarsening of logging curves possible, laying a technical foundation for improving the efficiency of logging data processing and the accuracy of geological interpretation.

Description

Technical Field

[0001] The present invention belongs to the field of logging-while-drilling data processing in oil and gas exploration and development, and particularly relates to a method for adaptively scaling coarsening of logging-while-drilling curves. Background Art

[0002] Logging technology is an indispensable means in oil and gas exploration and development. By recording the changes in formation physical parameters, high-resolution logging curves are generated, providing core data support for reservoir evaluation, geological modeling, and oil and gas development. However, with the progress of modern logging instruments, the sampling resolution of logging data has been significantly improved, and the common sampling interval can reach the centimeter level or even the millimeter level. For example, an array acoustic logging tool can generate hundreds of data points per unit depth. While this high-resolution data provides rich geological information, it also poses significant challenges: large data volume, high storage requirements, increased computational complexity, especially in real-time drilling monitoring, large-scale reservoir simulation, or joint analysis of multi-well data, traditional processing methods are difficult to meet the dual requirements of efficiency and accuracy.

[0003] The limitations of traditional methods are mainly reflected in the following aspects: (1) Insufficient resolution matching: Fixed coarsening strategies cannot flexibly adjust according to the resolution differences of logging curves, resulting in information loss in high-resolution regions or redundancy in low-resolution regions. (2) Unintelligent noise processing: Logging data often contains instrument noise, formation heterogeneity interference, etc., and traditional filtering is difficult to distinguish noise from geological signals. (3) Geological information loss: Fixed coarsening provides insufficient protection for geological features, affecting the geological interpretation value of the coarsened data. (4) Lack of adaptability: Different application scenarios have different requirements for coarsening resolution, and traditional methods require manual parameter adjustment, resulting in low efficiency. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method for adaptively scaling coarsening of logging-while-drilling curves, including: preprocessing the original logging curve, including outlier processing, missing value imputation, sampling point alignment, normalization processing, adaptive slicing, filling, and masking processing, and dividing the processed data into a training set and a test set; using a multi-layer one-dimensional convolutional network to extract a preliminary feature sequence, and calculating global dependencies through the self-attention mechanism of the Transformer to generate a global feature representation, while dynamically adjusting the convolutional kernel parameters of the one-dimensional convolutional network; training the model using the training set and testing the trained model using the test set; inputting the logging data to be coarsened after preprocessing into the model to generate an importance scoring vector; determining coarsening parameters according to the importance scoring vector, including the downsampling rate and smoothing intensity; coarsening the preprocessed logging curve according to the coarsening parameters, including downsampling and smoothing processing, and restoring the physical dimension to generate the final coarsened curve.

[0005] Preferably, the process of adaptive slicing includes: calculating the local variance characteristics of well logging curves to determine the signal resolution; dynamically adjusting the slice length through linear mapping based on eigenvalues, using a shorter segmentation length in high-resolution regions and a longer segmentation length in low-resolution regions; traversing the data in depth order, performing slicing according to the dynamic slice length, and marking the starting depth of each segment.

[0006] Preferably, the process of padding and masking includes: setting the maximum slice length, padding slices shorter than this length, with the padding value being a special value -1; generating a mask tensor for each slice, marking actual data points as 1 and padding points as 0.

[0007] Preferably, during the process of using a multi-layer one-dimensional convolutional network to extract the initial feature sequence, the mask is applied to calculate only at positions where the mask value is 1.

[0008] Preferably, when calculating the global dependencies through the self-attention mechanism of the Transformer, an attention mask is used to ignore the padding positions; the Transformer generates weights through the Gumbel-Softmax function to dynamically adjust the convolutional kernel parameters of the one-dimensional convolutional network.

[0009] Preferably, the process of determining the coarsening parameters according to the importance scoring vector includes: using the Sigmoid function to map the importance score to the downsampling rate; using a linear method to map the importance score to the smoothing intensity to complete the generation of the coarsening parameters.

[0010] Preferably, the steepness parameter, threshold, and user influence parameter of the Sigmoid function are adjusted according to geological requirements; the user influence parameter of the linear method is adjusted according to geological requirements.

[0011] Preferably, the process of coarsening the preprocessed well logging curves according to the coarsening parameters includes: downsampling the preprocessed well logging curves according to the downsampling rate to obtain the downsampled point set and the corresponding depth; performing smoothing processing on the downsampled point set using mean filtering, with the intensity being the smoothing intensity; and anti-normalizing the smoothed point set to restore the physical dimension to generate the final coarsened curve.

[0012] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, and when the processor executes the computing program, the method is implemented.

[0013] On the other hand, the present invention also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the method is implemented.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects:

[0015] The present invention provides an adaptive scale coarsening method for logging-while-drilling curves. Compared with traditional logging curve coarsening methods, it has significant technical advantages. First, by slicing the high- and low-resolution characteristic sections of the logging curve at different lengths, and performing slice completion and position masking. Then, by extracting the local features of the logging curve through a one-dimensional convolutional neural network and combining the global dependency modeling of the Transformer model, it can dynamically identify the high-importance regions and low-importance regions in the data, and then map them to coarsening parameters that can be adaptively adjusted, avoiding the problem of key feature loss caused by traditional fixed-rule coarsening methods. Second, the Transformer optimizes the convolutional kernel parameters through the self-attention mechanism, further improving the accuracy and adaptability of feature extraction, making the coarsening process more in line with the actual geological characteristics of logging data. In addition, this algorithm supports users to adjust the input influence parameters according to specific needs and further optimize the coarsening effect through manual intervention, with high flexibility and practicality. Brief Description of the Drawings

[0016] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0017] Figure 1 is a schematic flowchart of the method of the embodiment of the present invention;

[0018] Figure 2 is a flowchart of model training of the embodiment of the present invention. Detailed Embodiments

[0019] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0020] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0021] Embodiment 1: As Figure 1 - Figure 2 shown, this embodiment provides an adaptive scale coarsening method for logging-while-drilling curves, including:

[0022] Preprocess the original logging curves, including outlier processing, missing value imputation, sampling point alignment, normalization, adaptive slicing, padding, and masking, and divide the processed data into a training set and a test set; use a multi-layer one-dimensional convolutional network to extract preliminary feature sequences, and calculate the global dependencies through the self-attention mechanism of the Transformer to generate global feature representations, while dynamically adjusting the convolutional kernel parameters of the one-dimensional convolutional network; use the training set to train the model and the test set to test the trained model; input the logging data to be coarsened after preprocessing into the model to generate an importance scoring vector; determine the coarsening parameters according to the importance scoring vector, including the downsampling rate and the smoothing intensity; coarsen the preprocessed logging curves according to the coarsening parameters, including downsampling and smoothing, and restore the physical dimension to generate the final coarsened curve.

[0023] S1: Logging curve data processing.

[0024] S11: Perform preliminary cleaning on the input original logging curves to identify and process outliers.

[0025] S12: For possible missing values or non-uniform sampling problems in the logging curves, use interpolation methods to fill in the missing data and align the sampling points. The interpolation formula is:

[0026] ;

[0027] where z is the depth, x(z) is the value of the interpolation point, and z1, z2 are the depths of adjacent sampling points.

[0028] S13: Normalize the cleaned and aligned logging curves, map the data value range to [0,1], and eliminate the differences in different physical dimensions. The min-max normalization formula is:

[0029] ;

[0030] where x is the original data value, x' is the normalized value, and min(x) and max(x) are the minimum and maximum values of the data respectively.

[0031] S14: Perform adaptive slicing on the logging curves according to the signal characteristics to adapt to different resolution characteristics - use a shorter segmentation length in the high-resolution region and a longer segmentation length in the low-resolution region. Determine the signal resolution by calculating the local variance characteristics of the local signal, and use a window size W of 20. The calculation formula is:

[0032] ;

[0033] where, is the local variance of the data points, W is the size of the local window, is the value of the data points within the window, is the mean value of the data points within the window.

[0034] Dynamically adjust the slice length through linear mapping based on the eigenvalue, and set the maximum slice length to 50. The mapping rule is:

[0035] ;

[0036] where Segment_length is the slice length of the data point's section, is the local variance.

[0037] Finally, traverse the data in depth order, split according to the dynamic slice length, mark the starting depth for each segment, and record as where is the m-th slice.

[0038] S15: Perform padding and masking on slices of different lengths to unify the input shape. Set the same maximum slice length as in S14, pad the slices shorter than this length, and use a special value "-1" unrelated to the curve value for padding to distinguish the actual data from the padded part. Generate a mask tensor for each slice, mark the actual data points as 1 and the padded points as 0. For example: for a slice of length 10, pad it to 50 points, , (the first 10 are 1 and the last 40 are 0).

[0039] S16: Divide the processed data into a training set and a test set in a ratio of 70% and 30% to prepare for subsequent model training and evaluation.

[0040] S2: Coarse model training and usage.

[0041] S21: Use a multi-layer one-dimensional convolutional network to extract the preliminary feature sequence. Initialize the size and number of convolutional kernels, input the padded slices and masks into the one-dimensional convolutional network. The one-dimensional convolutional network applies the mask during the convolution operation and only calculates at the positions where the mask value is 1 to ensure that the padding value does not affect feature extraction, and generates the preliminary feature sequence.

[0042] S22: Calculate the global dependencies based on the preliminary feature sequence through the self-attention mechanism of the Transformer and ignore the padding positions through the attention mask to generate a preliminary global feature representation. Based on the preliminary global features, the Transformer learns the optimal convolution kernel size and number, and generates weights through the Gumbel-Softmax function. Gumbel-Softmax is a differentiable approximation method designed specifically for discrete choice problems. By introducing Gumbel noise and temperature parameter τ, the discrete choice is approximated as a continuous probability distribution, and its formula is:

[0043] ;

[0044] where , , τ is the temperature parameter, is the original probability of the category.

[0045] Feed these weights back to S21 to dynamically adjust the convolution kernel parameters of the one-dimensional convolutional network.

[0046] S23: The one-dimensional convolution uses the optimized convolution kernel to perform convolution operations on the input data again and outputs the final feature sequence.

[0047] S24: The Transformer adds position encoding to the final feature sequence again and generates an importance scoring vector for each section through the self-attention mechanism for subsequent coarsening parameter adjustment.

[0048] S25: Use the training set divided by S1 to train the model and optimize the parameters of the convolutional network and the Transformer. Use the test set divided by S1 to test the trained model. Verify the performance of the model on unseen data through evaluation metrics.

[0049] S26: Input the well logging data to be recognized into the model after being processed by the preprocessing steps in S1, and generate an importance scoring vector for use in the third step S3.

[0050] S3: Receive the importance scoring vector from S26 , and map it to the coarsening parameters of the well logging curves. First, determine the scoring granularity. Through uniform distribution, the importance scoring is extended to the level of each data point within the section. Use the Sigmoid function to map the downsampling rate, and the formula is:

[0051] ;

[0052] where, is the minimum value of the downsampling rate, set to 0.2, is the maximum downsampling rate, set to 0.9, is the importance score of the i-th point, k is the steepness parameter of the Sigmoid function, set to 10, is the threshold of the Sigmoid, set to 0.5, is the user influence parameter, controlling the conservativeness of the retention probability. The user can input an appropriate influence parameter according to geological requirements, and the default value is 1. High corresponds to high to retain more points, low corresponds to low for substantial compression.

[0053] Use a linear method to map the smoothing intensity. The formula is:

[0054] ;

[0055] where, is the minimum smoothing intensity, set to 3, is the maximum smoothing intensity, set to 15, is the user influence parameter, controlling the conservativeness of the smoothing intensity, and the default value is 1. High corresponds to small to retain details, low corresponds to large to increase smoothing.

[0056] S4: Coarsen the preprocessed well log curve to be measured based on the coarsening parameters of S3. First, perform downsampling according to to obtain the downsampled point set and the corresponding depth , and then perform smoothing on using mean filtering with an intensity of . Calculate the mean of the retained points in the neighborhood of each point to obtain the smoothed point set . Then, is denormalized to restore the physical dimension, and the final coarsened curve is generated. The formula is:

[0057] ;

[0058] where x is the original data value, x' is the normalized value, and min(x) and max(x) are the minimum and maximum values of the data respectively.

[0059] The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An adaptive scale coarsening method for logging-while-drilling curves, characterized in that, Including: Preprocess the original logging curves, including outlier handling, missing value imputation, sampling point alignment, normalization, adaptive slicing, padding and masking, and divide the processed data into a training set and a test set; use a multi-layer one-dimensional convolutional network to extract a preliminary feature sequence, and calculate the global dependence through the self-attention mechanism of the Transformer to generate a global feature representation, while dynamically adjusting the convolutional kernel parameters of the one-dimensional convolutional network; use the training set to train the model and use the test set to test the trained model; input the logging data to be coarsened after preprocessing into the model to generate an importance scoring vector for each section; determine the coarsening parameters according to the importance scoring vector of each section, including the downsampling rate and the smoothing intensity; coarsen the preprocessed logging curves according to the coarsening parameters, including downsampling and smoothing, and restore the physical dimension to generate the final coarsened curve; The process of the adaptive slicing includes: calculating the local variance feature of the logging curve to determine the signal resolution; dynamically adjusting the slice length through linear mapping based on the eigenvalue, using a shorter segmentation length for high-resolution sections and a longer segmentation length for low-resolution sections; traversing the data in depth order, slicing according to the dynamic slice length, and marking the starting depth of each segment; The process of determining the coarsening parameters according to the importance scoring vector of each section includes: using the Sigmoid function to map the importance score to the downsampling rate; using a linear method to map the importance score to the smoothing intensity to complete the generation of the coarsening parameters; The expression for using the Sigmoid function to map the importance score to the downsampling rate is: ； Among them, is the minimum value of the downsampling rate, set to 0.2, is the maximum value of the downsampling rate, set to 0.9, is the importance score of the i-th point, k is the steepness parameter of the Sigmoid function, set to 10, is the threshold of the Sigmoid, set to 0.5, is the user influence parameter, controlling the conservativeness of the retention probability; The expression for using the linear method to map the importance score to the smoothing intensity is: ； Among them, is the minimum smoothing intensity, set to 3, is the maximum smoothing intensity, set to 15, is the user influence parameter, controlling the conservatism of the smoothing intensity, and the default value is 1.

2. The method according to claim 1, wherein The process of the padding and masking includes: setting the maximum slice length, padding the slices shorter than this length, and using a special value -1 for the padding value; generating a mask tensor for each slice, marking the actual data points as 1 and the padding points as 0.

3. The method according to claim 1, wherein During the process of using the multi-layer one-dimensional convolutional network to extract the preliminary feature sequence, the mask is applied to calculate only at the positions where the mask value is 1.

4. The method according to claim 1, wherein When calculating the global dependence through the self-attention mechanism of the Transformer, the attention mask is used to ignore the padding positions; the Transformer generates weights through the Gumbel-Softmax function to dynamically adjust the convolutional kernel parameters of the one-dimensional convolutional network.

5. The method according to claim 1, characterized in that, The steepness parameter, threshold and user influence parameter of the Sigmoid function are adjusted according to geological requirements; the user influence parameter of the linear method is adjusted according to geological requirements.

6. The method according to claim 1, wherein The process of coarsening the preprocessed logging curves according to the coarsening parameters includes: downsampling the preprocessed logging curves according to the downsampling rate to obtain the downsampled point set and the corresponding depth; smoothing the downsampled point set, using mean filtering with an intensity equal to the smoothing intensity; anti-normalizing the smoothed point set to restore the physical dimension to generate the final coarsened curve.

7. An electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, the method described in any one of claims 1-6 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the method described in any one of claims 1-6 is implemented.

Images