Dipolar shear wave velocity extraction method and system

Abstract

This application belongs to the field of geophysical well logging technology and relates to a method and system for extracting dipole shear wave velocity. The method comprises the following steps: processing the acquired acoustic array waveform data of a target well section to obtain a dispersion energy spectrum; extracting and clustering the dispersion points in the dispersion energy spectrum to obtain a dispersion point set; extracting regions in the dispersion energy spectrum with energy values ​​greater than a set energy threshold and processing them into a dispersion spectrum weight matrix; inputting the determined formation and wellbore parameters to be inverted into a pre-built dispersion curve forward model to obtain a theoretical dispersion curve; calculating a joint error function for each depth point based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve; and using the shear wave velocity contained in the formation and wellbore parameters with the minimum joint error function as the shear wave velocity at the current depth point; and arranging the shear wave velocities corresponding to each depth point in the target logging section according to depth to obtain a continuous formation shear wave velocity profile. This application can quickly and accurately obtain dipole shear wave velocity.

Description

Technical Field

[0001] The present application belongs to the field of geophysical well logging technology, and specifically relates to a dipole shear wave velocity extraction method and system. Background Art

[0002] Borehole acoustic wave measurement is a core technology used in formation evaluation, wellbore stability analysis, and seismic data interpretation. By exciting acoustic waves in the wellbore and recording their propagating waveforms with an array of receivers, rich information including formation elastic parameters (such as shear wave velocity, Vs) can be obtained.

[0003] For dipole shear wave measurements, the dispersion characteristics of the flexural wave mode are highly sensitive to formation shear wave velocity. Traditional dipole shear wave velocity inversion relies on the following steps: First, a dispersion curve (typically the relationship between phase velocity or slow velocity and frequency) is extracted from the received array acoustic waveform. Second, a theoretical dispersion curve is calculated through numerical forward modeling based on a physical model of the wellbore acoustic field (e.g., the dispersion equation). Finally, an iterative optimization algorithm (e.g., least squares method, simulated annealing, genetic algorithm, etc.) is used to adjust the formation model parameters (including shear wave velocity) to achieve an optimal fit between the theoretical dispersion curve and the actual extracted dispersion curve.

[0004] Traditional inversion methods based on physical models and iterative optimization have the following problems and shortcomings: (1) The numerical calculation of the forward modeling of the borehole acoustic field (such as finite difference, real axis integration, etc.) is very time-consuming, and the iterative optimization process requires multiple forward modeling calculations, which is computationally inefficient, resulting in low efficiency of the entire inversion process and difficulty in meeting the needs of real-time processing. (2) The performance and results of the optimization algorithm are often affected by the initial model, parameter settings (such as step size, number of iterations) and optimization strategy. The parameters are highly dependent and it is easy to fall into local optimal solutions. Especially in highly nonlinear and multimodal problems such as dispersion inversion, the reliability and stability of the inversion results are low. (3) When the measured data noise is strong, the dispersion energy is weak, or the dispersion point extraction is inaccurate (especially in the low frequency band), the fitting accuracy and stability of the traditional method will be significantly reduced.

[0005] In recent years, with the development of artificial intelligence (AI) technology, deep learning methods have been explored for use in acoustic logging data processing (including automated extraction of dispersion curves and direct inversion of formation parameters). Some acoustic logging data processing methods use convolutional neural networks (CNNs), recurrent neural networks (RNNs), or their variants (such as long short-term memory (LSTM) and gated recurrent units (GRUs)) to learn complex mappings from raw waveform data, time-frequency spectrograms, or extracted dispersion points to formation parameters. While these methods have improved processing efficiency and automation to some extent, they still have the following limitations:

[0006] Insufficient matching between network structure and data characteristics: Some methods adopt a general deep learning structure without deep optimization design for the specific physical properties of borehole acoustic logging data and the structural characteristics of dispersion curves, resulting in limited model expression ability and generalization performance.

[0007] Insufficient fusion of physical information: Some purely data-driven methods may produce inversion results that do not conform to geophysical laws due to the lack of sufficient physical constraints, especially when the training data coverage is incomplete or unconventional formations are encountered. Their "black box" characteristics also make the results difficult to interpret and quality control.

[0008] Insufficient utilization of dispersion information: Some methods still rely primarily on discrete dispersion points for learning or inversion, while ignoring the richer full-band energy distribution information contained in the dispersion spectrum. Especially in the low-frequency region, the extraction of dispersion points is inherently uncertain, and relying solely on these points can lead to reduced inversion accuracy in this important frequency band.

[0009] Limitations of frequency dispersion point extraction: The extraction and clustering algorithms (such as DBSCAN) used for frequency dispersion point input in deep learning may themselves have problems such as parameter sensitivity and high computational complexity, which affects the degree of automation and robustness of the entire process. Summary of the Invention

[0010] In response to the above-mentioned problems of low inversion efficiency and poor accuracy in the prior art, the present application provides a dipole shear wave velocity extraction method and system with high inversion efficiency and accurate inversion results. By constructing a dispersion curve forward modeling proxy model based on a hybrid neural network, combining real-time data with a clustering algorithm to extract reliable dispersion points, and using the full dispersion spectrum energy information as a joint constraint, the formation shear wave velocity is inverted through an optimization algorithm. The dipole shear wave velocity can be obtained quickly and accurately, thereby improving the real-time performance and accuracy of shear wave velocity inversion under complex formation conditions.

[0011] In a first aspect, the present application provides a method for extracting dipole shear wave velocity, the steps of which are as follows:

[0012] Acquire acoustic array waveform data from well logging records of the target well section on site;

[0013] The acoustic array waveform data is processed to obtain a dispersion energy spectrum. The dispersion points in the dispersion energy spectrum are extracted to form a dispersion point cloud. The CLASSIX clustering algorithm is used to cluster the dispersion point cloud to obtain clusters of different dispersion modes. The dispersion point set representing the main mode of the target dipole shear wave and bending wave is screened from the clusters. The region in the dispersion energy spectrum with energy values ​​greater than the set energy threshold is extracted and processed into a dispersion spectrum weight matrix.

[0014] Determining formation and wellbore parameters to be inverted within a set range, including formation shear wave velocity, and inputting the formation and wellbore parameters into the constructed dispersion curve forward model to obtain a theoretical dispersion curve;

[0015] For each depth point in the target logging section, a joint error function is calculated based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve. The formation and wellbore parameters that minimize the joint error function are obtained, and the shear wave velocity contained in the formation and wellbore parameters is used as the shear wave velocity at the current depth point.

[0016] The shear wave velocity corresponding to each depth point of the target logging section is arranged according to depth to obtain a continuous formation shear wave velocity profile.

[0017] In some embodiments, the method further includes: before clustering the dispersive point cloud, normalizing the slowness / speed and frequency values ​​of the dispersive point cloud.

[0018] In some embodiments, a method for selecting a set of dispersion points representing a main mode of a target dipole shear-wave bending wave from a cluster is as follows: based on the number of dispersion points, frequency coverage, and slowness / speed distribution range in each cluster, combined with the energy concentration of each cluster on the dispersion energy spectrum, a set of dispersion points representing a main mode of a target dipole shear-wave bending wave is selected from the cluster.

[0019] In some embodiments, the method of processing the extracted region into a dispersion spectrum weight matrix is:

[0020] The extracted area is formed into a binary mask matrix;

[0021] Performing convolution processing on the mask matrix to obtain a continuous weight matrix;

[0022] The continuous weight matrix is ​​quantized to obtain a dispersion spectrum weight matrix.

[0023] In some embodiments, the dispersion curve forward model includes a fully connected layer, a bidirectional gated recurrent unit layer, and a one-dimensional convolutional neural network layer connected in series in sequence; the fully connected layer performs nonlinear feature transformation on the input parameters, the bidirectional gated recurrent unit layer is used to capture the long-range dependencies and contextual information in the output features of the fully connected layer, and the one-dimensional convolutional neural network layer is used to extract the local frequency variation features of the dispersion curve from the output of the bidirectional gated recurrent unit layer.

[0024] In some embodiments, the method for constructing the dispersion curve forward model is:

[0025] An initial dispersion curve forward model is constructed based on a hybrid neural network, with formation and wellbore parameters as input, theoretical dispersion curves as output, and a composite loss function containing multiple constraints as the loss function. The composite loss function is expressed as:

[0026]

[0027] Where, is the composite loss function, is the mean square error loss function, is the monotonic loss function, is the smoothness loss function, is the frequency range weighted loss function, is the weight coefficient, determined by experiment;

[0028] The mean square error loss function is expressed as:

[0029]

[0030] Where, N is the sample size, For the The predicted value of the sample, For the The true value of the samples;

[0031] The monotonicity loss function is expressed as:

[0032]

[0033] Where, is the total number of frequency points, For the The predicted value of the frequency point, For the The predicted value of each frequency point;

[0034] The smoothness loss function is expressed as:

[0035]

[0036] Where, For the The predicted value of each frequency point;

[0037] The frequency range weighted loss function is expressed as:

[0038]

[0039] Where, is the frequency band, is an optional emphasis weight, For the The true value of the frequency point;

[0040] The dispersion curve forward model is obtained by training the initial dispersion curve forward model with a pre-synthesized data set.

[0041] In some embodiments, the data set is synthesized by:

[0042] Set the value range and sampling step of formation and wellbore parameters;

[0043] Fix at least one formation and wellbore parameter, and perform combined sampling of other formation and wellbore parameters within their typical geological ranges to generate different sets of parameter combinations;

[0044] For each parameter combination, the dispersion equation in borehole acoustics theory is numerically solved to obtain the corresponding theoretical bending wave dispersion curve.

[0045] The theoretical bending wave dispersion curves corresponding to each parameter combination are collected to obtain a dispersion curve data set;

[0046] The data set is obtained by performing data cleaning on a dispersion curve data set.

[0047] In some embodiments, a method for cleaning a dispersion curve dataset to obtain the dataset is:

[0048] Eliminate dispersion curves from the dispersion curve dataset that fail to be calculated or do not conform to physical laws;

[0049] Smooth or correct the numerical perturbation of the dispersion curve in the low frequency band of 1-3kHz in the dispersion curve data set;

[0050] The data set is obtained by correcting the non-physical lift of the tail of the dispersion curve in the high frequency band above 7 kHz in the dispersion curve data set.

[0051] In a second aspect of the present application, a dipole shear wave velocity extraction system is provided, which is used to implement the dipole shear wave velocity extraction method described in the first aspect of the present application, comprising:

[0052] A data acquisition module is used to obtain acoustic array waveform data of the well logging record of the target well section on site;

[0053] The data processing module processes the acoustic array waveform data to obtain a dispersion energy spectrum; extracts dispersion points from the dispersion energy spectrum to form a dispersion point cloud; clusters the dispersion point cloud using the CLASSIX clustering algorithm to obtain clusters of different dispersion modes; and selects a dispersion point set representing the main mode of the target dipole shear wave and bending wave from the cluster; extracts regions in the dispersion energy spectrum where the energy value is greater than a set energy threshold and processes the regions into a dispersion spectrum weight matrix;

[0054] Model building module, used to build dispersion curve forward model;

[0055] A determination module, for determining formation and wellbore parameters to be inverted including formation shear wave velocity within a set range;

[0056] A forward modeling module, inputting the formation and wellbore parameters into a dispersion curve forward model to obtain a theoretical dispersion curve;

[0057] The inversion module calculates a joint error function for each depth point in the target logging section based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve, and obtains the formation and wellbore parameters that minimize the joint error function. The shear wave velocity contained in the formation and wellbore parameters is used as the shear wave velocity at the current depth point.

[0058] The sorting module arranges the shear wave velocity corresponding to each depth point of the target logging section according to the depth to obtain a continuous formation shear wave velocity profile.

[0059] In some embodiments, the data processing module includes:

[0060] The first processing module is used to process the acoustic wave array waveform data to obtain a dispersion energy spectrum;

[0061] The first extraction module extracts the dispersion points in the dispersion energy spectrum to form a dispersion point cloud;

[0062] Clustering module, clustering the scattered point cloud to obtain clusters with different dispersion patterns;

[0063] A screening module selects the frequency dispersion point set representing the main mode of the target dipole shear wave and bending wave from the cluster;

[0064] The second extraction module extracts the region in the dispersion energy spectrum where the energy value is greater than a set energy threshold;

[0065] The second processing module processes the region into a dispersion spectrum weight matrix.

[0066] Compared with the prior art, the advantages and positive effects of this application are:

[0067] (1) The dipole shear wave velocity extraction method and system provided in this application obtains a dispersion energy spectrum by processing the acoustic array waveform data. On the one hand, the dispersion points in the dispersion energy spectrum are extracted to form a dispersion point cloud, and the dispersion point cloud is clustered and the main mode is screened. After clustering, the continuous coverage rate of the dispersion point set in the target frequency band (such as 1-10kHz) is as high as more than 90%, with small errors and high main mode identification accuracy (greater than 95%). It realizes high-precision, high-robustness, and fully automated extraction of the main mode of the dipole shear wave bending wave, significantly improving the efficiency and reliability of dispersion analysis, and providing key data support for subsequent formation shear wave velocity inversion, anisotropy assessment and geological guidance. On the other hand, the area with energy value greater than the set energy threshold in the dispersion energy spectrum is extracted and the area is processed into a dispersion spectrum weight matrix, which can reflect the energy intensity of different frequency and velocity positions and is used to constrain the inversion process in the future.

[0068] (2) The dipole shear wave velocity extraction method and system provided in this application adopts the CLASSIX clustering algorithm to extract and optimize the scattered points. Compared with traditional algorithms such as DBSCAN, the method and system have lower parameter sensitivity, higher computational efficiency, and more stable clustering effect, thereby improving the automation level of scattered point picking and adaptability to different data qualities.

[0069] (3) The dipole shear wave velocity extraction method and system provided in this application uses a constructed dispersion curve forward model to perform forward modeling to obtain a theoretical dispersion curve. Compared with the traditional time-consuming numerical simulation forward modeling, the calculation speed of a single dispersion curve is shortened from several minutes to milliseconds, and the overall inversion efficiency is significantly improved, with good real-time or quasi-real-time performance.

[0070] (4) The dipole shear wave velocity extraction method and system provided in this application, the dispersion curve forward model is constructed based on a hybrid neural network. On the one hand, the dispersion curve forward model adopts a structural design including a fully connected layer, a bidirectional gated recurrent unit layer and a one-dimensional convolutional neural network layer connected in series, which can more accurately learn the complex nonlinear mapping relationship between formation and wellbore parameters and dispersion curves, while maintaining physical consistency. The forward modeling accuracy, computational efficiency and generalization ability are significantly improved. On the other hand, the dispersion curve forward model adopts a composite loss function including multiple constraints as the loss function, introduces physical constraints (including monotonicity and smoothness), and uses the overall energy distribution of the frequency spectrum in the inversion, ensuring that the inversion results better follow the laws of geophysics.

[0071] (5) The dipole shear wave velocity extraction method and system provided in this application innovatively introduces frequency spectrum energy weights as joint constraints, which effectively supplements the problem of insufficient information when using discrete frequency dispersion points (especially in low-frequency and low signal-to-noise ratio areas), suppresses noise interference and multi-solutions of inversion, and improves the accuracy and geological complexity of the inversion effect (especially in the low-frequency band).

[0072] (6) The dipole shear wave velocity extraction method and system provided in this application adopts a joint driving mechanism of data and physical information, which makes it more stable and reliable when processing data under complex actual working conditions such as wellbore collapse and noise interference.

[0073] (7) The dipole shear wave velocity extraction method and system provided in this application provide more effective shear wave velocity information for the detailed evaluation of complex formations, low porosity and low permeability oil and gas reservoirs by improving efficiency, accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a flow chart of the dipole shear wave velocity extraction method described in an embodiment of the present application;

[0075] Figure 2 A flow chart of a method for processing an extracted region into a dispersion spectrum weight matrix according to an embodiment of the present application;

[0076] Figure 3 Schematic diagram of the structure of the dispersion curve forward model described in the embodiment of the present application;

[0077] Figure 4 This is a flow chart of a method for constructing a dispersion curve forward model according to an embodiment of the present application;

[0078] Figure 5 This is a flow chart of the method for synthesizing the data set described in the embodiments of this application;

[0079] Figure 6 A flow chart of a method for cleaning a dispersion curve dataset to obtain the dataset according to an embodiment of the present application;

[0080] Figure 7 This is a structural block diagram of the dipole shear wave velocity extraction system described in an embodiment of the present application;

[0081] Figure 8 This is a structural block diagram of the data processing module described in the embodiment of the present application;

[0082] Figure 9 This is a structural block diagram of the model building module described in the embodiment of the present application;

[0083] Figure 10 Schematic diagram of the original dispersion spectrum and dispersion points of the embodiment of the present application;

[0084] Figure 11 This is a schematic diagram of the frequency dispersion points after clustering in an embodiment of the present application;

[0085] Figure 12 This is a schematic diagram of the frequency dispersion points after screening in the embodiment of the present application;

[0086] Figure 13 This is a schematic diagram of the binarization spectrum described in the embodiment of the present application;

[0087] Figure 14 This is a schematic diagram of the convolution spectrum described in the embodiment of the present application;

[0088] Figure 15 This is a schematic diagram of the quantization spectrum described in the embodiment of the present application;

[0089] Figure 16 A comparison chart of inversion results using the dipole shear wave velocity extraction method and system of the present application and the traditional method;

[0090] Figure 17 This is a schematic diagram of the inversion results of the caliper logging curve;

[0091] Figure 18This is a schematic diagram of the inversion results of the P-wave velocity logging curve;

[0092] Figure 19 Schematic diagram of the inversion results of natural gamma ray logging curves.

[0093] In the figure, 1. data acquisition module, 2. data processing module, 21. first processing module, 22. first extraction module, 23. clustering module, 24. screening module, 25. second extraction module, 26. second processing module, 3. model building module, 31. synthesis module, 32. building module, 33. training module, 4. determination module, 5. forward modeling module, 6. inversion module, 7. sorting module. DETAILED DESCRIPTION

[0094] The present application will be described in detail below with reference to exemplary embodiments in conjunction with the accompanying drawings. However, it should be understood that elements, structures, and features in one embodiment may also be beneficially combined in other embodiments without further description.

[0095] See also Figure 1 The first embodiment of the present application provides a method for extracting dipole shear wave velocity, the steps of which are:

[0096] S1. Acquire acoustic array waveform data from well logging records of a target well section.

[0097] S2. Process the acoustic array waveform data to obtain a dispersion energy spectrum; extract the dispersion points in the dispersion energy spectrum to form a dispersion point cloud, use the CLASSIX clustering algorithm to cluster the dispersion point cloud to obtain clusters of different dispersion modes, and screen out the dispersion point set representing the main mode of the target dipole shear wave and bending wave from the cluster; extract the area in the dispersion energy spectrum where the energy value is greater than the set energy threshold, and process the extracted area into a dispersion spectrum weight matrix.

[0098] In an embodiment of the present application, a weighted spectral coherence method or a similar time-frequency analysis technique is used to process the acoustic array waveform data to obtain a dispersion energy spectrum in the time-frequency domain.

[0099] It should be noted that the extracted dispersion point is essentially the point at the position of the maximum vertical coordinate on each horizontal coordinate in the dispersion energy spectrum. The horizontal coordinate is frequency and the vertical coordinate is velocity, that is, the point with the maximum velocity at a certain frequency. In the embodiment of the present application, the CLASSIX clustering algorithm is used to cluster the dispersion point cloud composed of the extracted dispersion points. The CLASSIX clustering algorithm can effectively identify clusters of different dispersion modes (such as: bending wave main mode, leakage P wave, Stoneley wave, etc.) and separate noise points through data sorting and greedy aggregation. Compared with the DBSCAN clustering algorithm, the CLASSIX clustering algorithm is insensitive to input parameters (such as domain radius), has high computational efficiency, and can improve the efficiency of dipole shear wave velocity extraction.

[0100] In some embodiments of the present application, the method for screening out the set of scattered points representing the main mode of the target dipole shear wave bending wave from the cluster is: based on the number of scattered points in each cluster, the frequency coverage range, the slowness / speed distribution range, and the energy concentration of each cluster on the scattered energy spectrum, the set of scattered points representing the main mode of the target dipole shear wave bending wave is screened out from the cluster.

[0101] In this embodiment, dual constraints on the number of frequency dispersion points and energy concentration within a cluster are used to effectively suppress the infiltration of false modes caused by noise interference, thereby improving the accuracy of primary mode identification. Frequency coverage evaluation ensures that the screening results fully retain the effective frequency band of the primary mode (e.g., 2-8 kHz), avoiding frequency band truncation caused by over-screening and reducing errors in subsequent shear wave velocity inversion. Combined with the clustering characteristics of the slowness / velocity distribution range (e.g., the slowness of the primary flexural wave mode is typically 15-20% lower than that of pseudo-Rayleigh waves), modes with near-overlapping velocity regions can be distinguished.

[0102] In some embodiments of the present application, the method further includes: before clustering the dispersive point cloud, normalizing the slowness / speed and frequency values ​​of the dispersive point cloud.

[0103] In the embodiment of the present application, the slowness / speed and frequency of the dispersion point cloud are normalized to eliminate the dimension effect.

[0104] In some embodiments of this application, see Figure 2 , the method of processing the extracted area into a dispersion spectrum weight matrix is:

[0105] S21. Form a binary mask matrix from the extracted area.

[0106] S22. Perform convolution processing on the mask matrix to obtain a continuous weight matrix.

[0107] Specifically, a two-dimensional Gaussian convolution is applied to the binary mask matrix (for example, the convolution kernel size is 5×5, the standard deviation is ), so that the weight of the extracted area is close to 1, gradually and smoothly decays outward, and the weight of noise or other areas is close to 0.

[0108] S23 , quantizing (for example, converting to 8-bit integers) the continuous weight matrix to obtain a dispersion spectrum weight matrix.

[0109] In an embodiment of the present application, the hard boundaries of the binary mask are smoothed into continuous weights through a convolution operation, effectively reducing the discontinuity at the boundaries, and the continuous weights are discretized into a finite level (such as 8-bit or a specific threshold) through a quantization operation, reducing the amount of data to improve subsequent calculation efficiency.

[0110] S3. Determine formation and wellbore parameters to be inverted within a set range, including formation shear wave velocity, and input the formation and wellbore parameters into the constructed dispersion curve forward model to obtain a theoretical dispersion curve.

[0111] In the embodiment of the present application, the formation and wellbore parameters to be inverted primarily include formation shear wave velocity, and may also include wellbore radius, fluid velocity, fluid density, formation density, etc. The formation and wellbore parameters within a set range include multiple parameter combinations. Each parameter combination is input into the constructed dispersion curve forward model to obtain a theoretical dispersion curve corresponding to the parameter combination.

[0112] In some embodiments of this application, see Figure 3 The dispersion curve forward model includes a fully connected layer 100, a bidirectional gated recurrent unit layer 200 and a one-dimensional convolutional neural network layer 300 connected in series in sequence. The fully connected layer 100 performs nonlinear feature transformation on the input parameters. The bidirectional gated recurrent unit layer 200 is used to capture the long-range dependency and contextual information in the output features of the fully connected layer. The one-dimensional convolutional neural network layer 300 is used to extract the local frequency variation features of the dispersion curve from the output of the bidirectional gated recurrent unit layer.

[0113] The fully connected layer 100 may be one or more fully connected layers. The fully connected layer 100 maps the input formation and wellbore parameters to a higher-dimensional feature space (e.g., 512 dimensions). The activation function may be ReLU. Mathematically, it can be expressed as:

[0114]

[0115] Where, is the output of the fully connected layer, is the weight of the fully connected layer, To input formation and wellbore parameters, is the bias of the fully connected layer.

[0116] The bidirectional gated recurrent unit layer 200 treats the features output by the fully connected layer 100 as a sequence (if the output of the fully connected layer 100 is a fixed-dimensional vector, it can be expanded into a sequence form by reshaping or copying, or directly applied to a variable-length sequence), and uses the bidirectional gated recurrent unit layer 200 to capture the long-range dependencies and contextual information in the sequence. The internal structure of the bidirectional gated recurrent unit layer 200 includes the update gate and reset gate .

[0117]

[0118]

[0119]

[0120]

[0121] Where, for The hidden state of the moment, for Input at the moment, is the weight matrix of the update gate, is the bias vector of the update gate, To reset the gate weight matrix, is the bias vector for the reset gate, for The candidate hidden state at time t, is the weight matrix of the candidate hidden state, is the bias vector of the candidate hidden state, for The final hidden state at the moment.

[0122] Output of bidirectional gated recurrent unit layer 200 Integrates forward and reverse information flows.

[0123] The one-dimensional convolutional neural network layer 300 may include multiple convolutional layers and pooling layers, acting on the output of the bidirectional gated recurrent unit layer 200. , extracting local frequency-varying features from the dispersion curve. For example, using multiple convolution kernels of different sizes.

[0124]

[0125] Where, is the output of the one-dimensional convolutional neural network layer.

[0126] In this embodiment, the dispersion curve forward model utilizes a cascaded structure consisting of a fully connected layer, a bidirectional gated recurrent unit layer, and a one-dimensional convolutional neural network layer, achieving multi-scale feature fusion and physically consistent modeling. The fully connected layer can learn the implicit influence of complex physical property combinations (such as low-velocity interlayers and high Poisson's ratio media) on dispersion, thereby improving the parameter sensitivity of the inversion target. In the bidirectional gated recurrent unit layer, the forward GRU captures the causal sequence dependencies of the frequency-phase velocity curve (such as high-frequency asymptotic trends), while the reverse GRU supplements the reverse context (such as low-frequency cutoff characteristics), addressing the vanishing gradient problem of traditional RNNs in long sequences. Based on the global features output by the bidirectional gated recurrent unit layer, the instantaneous slope change or inflection point features of the dispersion curve are extracted using a local receptive field (such as a 3×1 convolution kernel). The cascaded design of the dispersion curve forward model achieves end-to-end, high-precision forward modeling from physical property parameters to dispersion curves.

[0127] Specifically, see Figure 3The dispersion curve forward model also includes an input layer 400 connected to the fully connected layer 100 and an output layer 500 connected to the one-dimensional convolutional neural network layer 300.

[0128] The input layer 400 is used to receive input formation and wellbore parameter vectors.

[0129] The output layer 500 is a fully connected layer, which regresses the local frequency-varying features of the dispersion curve extracted by the one-dimensional convolutional neural network layer 300 to the phase velocity value at the predetermined sampling frequency point to obtain the ideal dispersion curve.

[0130] In some embodiments of this application, see Figure 4 , the method for constructing the dispersion curve forward model is:

[0131] S311. With formation and wellbore parameters as input, theoretical dispersion curve as output, and a composite loss function containing multiple constraints as loss function, an initial dispersion curve forward model is constructed based on a hybrid neural network.

[0132] The composite loss function is expressed as:

[0133]

[0134] Where, is the composite loss function, is the mean square error loss function, is the monotonic loss function, is the smoothness loss function, is the frequency range weighted loss function, is the weight coefficient, which is determined by experiments (such as grid search) and can be 0.2, 0.1, and 0.3 respectively.

[0135] The mean square error loss function is expressed as:

[0136]

[0137] Where, N is the sample size, For the The predicted value of the sample, For the The true value of the samples.

[0138] The monotonicity loss function is expressed as:

[0139]

[0140] Where, is the total number of frequency points, For the The predicted value of the frequency point, For the The predicted value of each frequency point.

[0141] The smoothness loss function is expressed as:

[0142]

[0143] Where, For the The predicted value of each frequency point.

[0144] The frequency range weighted loss function is expressed as:

[0145]

[0146] Where, is the frequency band, is an optional emphasis weight, For the The true value of the frequency point.

[0147] S312: Train the initial dispersion curve forward model using a pre-synthesized data set to obtain the dispersion curve forward model.

[0148] Specifically, use optimizers such as Adam for model training and set an appropriate learning rate (for example: ), batch size (e.g. 128), dropout rate (e.g. 0.32), and early stopping strategy to prevent overfitting.

[0149] In some embodiments of this application, see Figure 5 , the synthesis method of the data set is:

[0150] S321, setting the value range and sampling step of formation and wellbore parameters;

[0151] S322, fixing at least one formation and wellbore parameter, and performing combined sampling on other formation and wellbore parameters within their typical geological ranges to generate different sets of parameter combinations;

[0152] S323. For each parameter combination, numerically solve the dispersion equation in borehole acoustics theory to obtain the corresponding theoretical bending wave dispersion curve;

[0153] S324, collecting the theoretical bending wave dispersion curves corresponding to each parameter combination to obtain a dispersion curve data set;

[0154] S325 , cleaning the dispersion curve data set to obtain the data set.

[0155] For example, the formation and wellbore parameters include wellbore radius, longitudinal wave velocity of the fluid in the well, density of the fluid in the well, shear wave velocity of the original formation, density of the original formation, longitudinal wave velocity of the formation, instrument radius, and instrument modulus. The instrument radius and instrument modulus are fixed, and other parameters are sampled in combination within their typical geological range to generate hundreds of thousands to millions of parameter combinations. For each set of parameter combinations, the real axis integration method or modal search algorithm is used to numerically solve the dispersion equation in the wellbore acoustics theory to obtain the corresponding theoretical bending wave dispersion curve. Each theoretical bending wave dispersion curve contains 200 Equally spaced frequency points within a frequency range.

[0156] The dispersion equation can be expressed as:

[0157]

[0158] Where, is the wave number; is the angular frequency; is the wellbore parameter; is the formation parameter; is the instrument radius, unit: m; is the instrument modulus, unit: GPa.

[0159] In some embodiments of this application, see Figure 6 , the method for cleaning the dispersion curve data set to obtain the data set is:

[0160] S331. Eliminate dispersion curves from the dispersion curve dataset that fail to be calculated or do not conform to physical laws. For example, dispersion curves with drastic non-monotonic jumps in phase velocity or abnormal energy.

[0161] S332. Smoothing or correcting the numerical disturbance of the dispersion curve in the low frequency band of 1-3 kHz in the dispersion curve data set.

[0162] S333. Correct the non-physical lift of the tail of the dispersion curve in the high frequency band above 7 kHz in the dispersion curve dataset to obtain the dataset.

[0163] Data cleaning ensures that the data distribution follows physical laws, improves data quality, and makes the dispersion curve show a monotonically smooth decrease in phase velocity as the frequency increases.

[0164] S4. For each depth point of the target logging section, a joint error function is calculated based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve to obtain the formation and wellbore parameters that minimize the joint error function. The shear wave velocity contained in the formation and wellbore parameters is used as the shear wave velocity at the current depth point.

[0165] Specifically, for each depth point in the target logging section, reasonable search limits (i.e., set ranges) and initial search steps are set for formation and borehole parameters. A multi-round grid search strategy is employed. For example, in the first round, a coarse search is performed on all parameters to be inverted, using a larger step size to determine the general trend and sensitivity range of the parameters. In subsequent rounds, some insensitive or already determined parameters can be fixed, and for highly sensitive parameters (especially formation shear wave velocity), the search range and step size can be gradually narrowed to perform a refined search.

[0166] Specifically, the joint error function is expressed as:

[0167]

[0168]

[0169]

[0170] Where, is the joint error function, is the point error, is the spectrum matching cost; 、 is the weight coefficient, is the theoretical dispersion curve, is the measured frequency dispersion point, is the frequency corresponding to the measured frequency dispersion point, is the dispersion spectrum weight matrix, is the current candidate parameter combination, is the frequency corresponding to the theoretical dispersion curve, is the mid-frequency dispersion point of the theoretical dispersion curve Corresponding slowness / speed.

[0171] Specifically, the position error is the error between the theoretical frequency dispersion point and the measured frequency dispersion point. Calculate theoretical dispersion curves At the measured frequency dispersion point Corresponding frequency The sum of the absolute differences (L1 norm) on (or the mean squared error L2 norm).

[0172] Spectral matching cost is the error between the theoretical dispersion points and the dispersion spectrum matrix. Function-measured theoretical dispersion curve and the dispersion spectrum weight matrix For example, the degree of matching can be calculated for each point on the theoretical dispersion curve. In the dispersion spectrum weight matrix Corresponding position in If the weight is low, a penalty is imposed, or the negative weighted sum of the spectral weights on the curve path is calculated.

[0173] Weight coefficient 、 Used to balance the contribution of point error and spectral constraint. Weight coefficient 、 The value of can be: .

[0174] S5. Arrange the shear wave velocity corresponding to each depth point of the target logging section according to the depth to obtain a continuous formation shear wave velocity profile.

[0175] See also Figure 7 The second embodiment of the present application provides a dipole shear wave velocity extraction system for implementing the dipole shear wave velocity extraction method described in the first aspect of the present application, comprising:

[0176] Data acquisition module 1, used to obtain acoustic array waveform data of well logging records of the target well section on site;

[0177] Data processing module 2 processes the acoustic array waveform data to obtain a dispersion energy spectrum; extracts dispersion points from the dispersion energy spectrum to form a dispersion point cloud, clusters the dispersion point cloud using the CLASSIX clustering algorithm to obtain clusters of different dispersion modes, and selects a dispersion point set representing the main mode of the target dipole shear wave and bending wave from the cluster; extracts regions in the dispersion energy spectrum where the energy value is greater than a set energy threshold, and processes the regions into a dispersion spectrum weight matrix;

[0178] Model building module 3, used to build a dispersion curve forward model;

[0179] Determination module 4, determining the formation and wellbore parameters to be inverted including the formation shear wave velocity within a set range;

[0180] Forward modeling module 5, inputting the formation and wellbore parameters into a dispersion curve forward model to obtain a theoretical dispersion curve;

[0181] Inversion module 6 calculates a joint error function for each depth point in the target logging section based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve, obtains the formation and wellbore parameters that minimize the joint error function, and uses the shear wave velocity contained in the formation and wellbore parameters as the shear wave velocity at the current depth point;

[0182] The sorting module 7 arranges the shear wave velocity corresponding to each depth point of the target logging section according to the depth to obtain a continuous formation shear wave velocity profile.

[0183] In some embodiments of the application, see Figure 8 , the data processing module 2 includes:

[0184] The first processing module 21 is used to process the acoustic wave array waveform data to obtain a dispersion energy spectrum;

[0185] A first extraction module 22 extracts dispersion points from the dispersion energy spectrum to form a dispersion point cloud;

[0186] A clustering module 23 clusters the dispersion point cloud to obtain clusters of different dispersion patterns;

[0187] A screening module 24 selects a set of frequency dispersion points representing the main mode of the target dipole shear wave and bending wave from the cluster;

[0188] The second extraction module 25 extracts the region in the dispersion energy spectrum where the energy value is greater than a set energy threshold;

[0189] The second processing module 26 processes the region into a dispersion spectrum weight matrix.

[0190] In some embodiments of the application, see Figure 9 , the model building module 3 includes:

[0191] A synthesis module 31, for synthesizing a data set;

[0192] A construction module 32 is configured to construct an initial dispersion curve forward model based on a hybrid neural network, using formation and wellbore parameters as input, a theoretical dispersion curve as output, and a composite loss function containing multiple constraints as a loss function;

[0193] The training module 33 trains the initial dispersion curve forward model using the synthesized data set to obtain the dispersion curve forward model.

[0194] In order to verify the effectiveness of the dipole shear wave velocity extraction method and system described in the above embodiments of the present application, the following specific embodiments are used for illustration.

[0195] Example: Take the extraction of dipole shear wave velocity in a selected well section in a certain area as an example.

[0196] The original dispersion spectrum and dispersion points of the selected well section in this area are shown in Figure 10 The extracted dispersion points are clustered to obtain the clustered dispersion points. Figure 11 , filter the clustered dispersion points to obtain the filtered dispersion points. Figure 12 The extracted area is binarized to obtain the binary spectrum (i.e., the binary mask matrix). Figure 13 , after convolution processing on the binary spectrum, the convolution spectrum (i.e., continuous weight matrix) is obtained. Figure 14 , the convolution spectrum is quantized to obtain the quantized spectrum (i.e., the dispersion spectrum weight matrix) Figure 15 .

[0197] Figure 16 The inversion results of the dipole shear wave velocity extraction method and system described in this application and other different inversion strategies for the selected well section ([3500 m, 3670 m]) in this area are presented and compared with conventional well logging curves, including well diameter (see Figure 17 ), longitudinal wave velocity (see Figure 18 ) and natural gamma (see Figure 19 ), and conduct a comprehensive comparative analysis.

[0198] Figure 16 In the figure, the blue curve represents the reference benchmark, that is, the monopole shear wave logging velocity profile of the selected well section (abbreviated as: monopole); the red curve represents the inversion result using the dipole shear wave velocity extraction method and system proposed in this application; the green curve represents the inversion result using only the traditional frequency dispersion point constraint strategy (abbreviated as: traditional method).

[0199] From the overall trend analysis of the entire well section, it can be seen that both inversion methods (traditional method and this application) can effectively reveal the macroscopic variation characteristics of the formation shear wave velocity. For example, in the depth range of 3580–3600 m, the inversion results of both methods show a high degree of consistency with the monopole logging curve. However, in several other key well sections (such as 3530–3570 m), the inversion results of this application (red curve) are significantly closer to the monopole logging benchmark, and its prediction accuracy is significantly better than the inversion results of the traditional method (green curve). The advantages of this application are particularly prominent in layers with relatively poor dispersion information quality (for example, sparse low-frequency data points, unstable dispersion point distribution, or noise interference), proving that this application can improve the robustness and accuracy of inversion.

[0200] Specifically, we introduce the specific contributions to the inversion performance, select several depth points with typical dispersion characteristics, and conduct a detailed comparative analysis of their inversion results. The details are as follows:

[0201] 3503.70m: At this depth, the measured dispersion points (black scattered points in the image) after clustering and screening are continuously distributed and morphologically stable. The corresponding dispersion spectrum energy also exhibits good focusing. Comparing the inversion results, the dispersion curve based on the traditional method (blue) and the dispersion curve of this application (orange) both show a high degree of match with the measured dispersion points. This demonstrates that, under conditions of high-quality dispersion information, both methods can effectively characterize the dispersion characteristics of the formation at this depth. The red dot on the left y-axis represents the shear wave velocity value measured by the monopole. The three are almost identical, indicating that the inversion results are comparable.

[0202] 3513.43m: The dispersion spectrum at this depth shows significant multimodal characteristics, especially in the 4.5–5.5kHz frequency band, where multiple side energy peaks are observed. If only dispersion point constraints are used, filtering to avoid interference from these abnormal frequency bands may result in the loss of effective high-frequency information. Its inversion result (blue curve) deviates in the high-frequency region and may lead to an overestimation of the low-frequency velocity prediction value through the coupling effect of the inversion process. In contrast, the present application (orange curve) incorporates the energy distribution (weight) information of the dispersion spectrum, and its inversion process is guided by the high-energy main mode region, so that the inverted dispersion curve is effectively corrected in the high-frequency part, and the final shear wave velocity is closer to the reference monopole shear wave logging result.

[0203] 3555.49m: The measured dispersion points at this depth exhibit significant up-and-down fluctuations in the low-frequency region. Constrained inversion based solely on these dispersion points (blue curve) is susceptible to outliers and tends to fit the average trend, resulting in an underestimation of the overall velocity prediction. However, this application (orange curve) introduces dispersion spectrum energy weighting to enhance the dominant role of high-energy-density regions (i.e., more reliable dispersion information) in the inversion process. This significantly corrects the inversion results, making them more consistent with the expected formation characteristics.

[0204] 3600.00m: At this depth, the inversion results (blue and orange curves) obtained by the two different constraint strategies (traditional method and application) are highly consistent. This phenomenon indicates that the dispersion information at this depth is of excellent quality: the measured dispersion points are distributed stably, and the dispersion spectrum energy is highly concentrated. Under these ideal conditions, the two methods are equivalent.

[0205] 3628.34m: The measured dispersion points at this depth are relatively flat. However, the inversion results (orange curve) from this application show a slight overall increase compared to the traditional inversion results (blue curve). This reflects a certain degree of energy unevenness in the dispersion spectrum in this region. This application effectively integrates this spectral information through weighting, fine-tuning the inversion results.

[0206] 3650.00 m: The dispersion points are distributed smoothly and the dispersion spectrum is highly concentrated. The blue and orange curves are highly consistent and completely overlap with the red shear wave velocity point, once again verifying that under conditions of good information quality, the inversion results of the two methods tend to be consistent.

[0207] The above embodiments are used to explain the present application rather than to limit the present application. Any modifications and changes made to the present application within the spirit of the present application and the protection scope of the claims shall fall within the protection scope of the present application.

Claims

1. A dipole shear wave velocity extraction method, characterized in that: The steps are: Acquire acoustic array waveform data from well logging records of the target well section on site; The acoustic array waveform data is processed to obtain a dispersion energy spectrum. The dispersion points in the dispersion energy spectrum are extracted to form a dispersion point cloud. The CLASSIX clustering algorithm is used to cluster the dispersion point cloud to obtain clusters of different dispersion modes. The dispersion point set representing the main mode of the target dipole shear wave and bending wave is screened from the clusters. The region in the dispersion energy spectrum with energy values ​​greater than the set energy threshold is extracted and processed into a dispersion spectrum weight matrix. Determining formation and wellbore parameters to be inverted within a set range, including formation shear wave velocity, and inputting the formation and wellbore parameters into the constructed dispersion curve forward model to obtain a theoretical dispersion curve; For each depth point in the target logging section, a joint error function is calculated based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve. The formation and wellbore parameters that minimize the joint error function are obtained, and the shear wave velocity contained in the formation and wellbore parameters is used as the shear wave velocity at the current depth point. Arrange the shear wave velocity corresponding to each depth point of the target logging section according to the depth to obtain a continuous formation shear wave velocity profile; The method of processing the extracted region into a dispersion spectrum weight matrix is ​​as follows: forming a binary mask matrix of the extracted region; The mask matrix is ​​subjected to convolution processing to obtain a continuous weight matrix; and the continuous weight matrix is ​​subjected to quantization processing to obtain a dispersion spectrum weight matrix.

2. The dipole shear wave velocity extraction method according to claim 1, wherein: The method further includes normalizing the slowness / speed and frequency values ​​of the dispersion point cloud before clustering the dispersion point cloud.

3. The dipole shear wave velocity extraction method according to claim 1, wherein: The method for selecting the dispersion point set representing the main mode of the target dipole shear-bending wave from the cluster is as follows: based on the number of dispersion points, frequency coverage, slowness / velocity distribution range in each cluster, combined with the energy concentration of each cluster on the dispersion energy spectrum, the dispersion point set representing the main mode of the target dipole shear-bending wave is selected from the cluster.

4. The dipole shear wave velocity extraction method according to claim 1, wherein: The dispersion curve forward model includes a fully connected layer, a bidirectional gated recurrent unit layer and a one-dimensional convolutional neural network layer connected in series in sequence. The fully connected layer performs nonlinear feature transformation on the input parameters. The bidirectional gated recurrent unit layer is used to capture the long-range dependency and contextual information in the output features of the fully connected layer. The one-dimensional convolutional neural network layer is used to extract the local frequency variation features of the dispersion curve from the output of the bidirectional gated recurrent unit layer.

5. The dipole shear wave velocity extraction method according to claim 1, wherein: The method for constructing the dispersion curve forward model is: An initial dispersion curve forward model is constructed based on a hybrid neural network, with formation and wellbore parameters as input, theoretical dispersion curves as output, and a composite loss function containing multiple constraints as the loss function. The composite loss function is expressed as: Where, is the composite loss function, is the mean square error loss function, is the monotonic loss function, is the smoothness loss function, is the frequency range weighted loss function, is the weight coefficient, determined by experiment; The mean square error loss function is expressed as: Where, N is the sample size, For the The predicted value of the sample, For the The true value of the samples; The monotonicity loss function is expressed as: Where, is the total number of frequency points, For the The predicted value of the frequency point, For the The predicted value of each frequency point; The smoothness loss function is expressed as: Where, For the The predicted value of each frequency point; The frequency range weighted loss function is expressed as: Where, is the frequency band, is an optional emphasis weight, For the The true value of the frequency point; The dispersion curve forward model is obtained by training the initial dispersion curve forward model with a pre-synthesized data set.

6. The dipole shear wave velocity extraction method according to claim 5, characterized in that: The synthesis method of the dataset is: Set the value range and sampling step of formation and wellbore parameters; Fix at least one formation and wellbore parameter, and perform combined sampling of other formation and wellbore parameters within their typical geological ranges to generate different sets of parameter combinations; For each parameter combination, the dispersion equation in borehole acoustics theory is numerically solved to obtain the corresponding theoretical bending wave dispersion curve. The theoretical bending wave dispersion curves corresponding to each parameter combination are collected to obtain a dispersion curve data set; The data set is obtained by performing data cleaning on a dispersion curve data set.

7. The dipole shear wave velocity extraction method according to claim 6, characterized in that: The method for cleaning the dispersion curve data set to obtain the data set is: Eliminate dispersion curves from the dispersion curve dataset that fail to be calculated or do not conform to physical laws; Smooth or correct the numerical perturbation of the dispersion curve in the low frequency band of 1-3kHz in the dispersion curve data set; The data set is obtained by correcting the non-physical lift of the tail of the dispersion curve in the high frequency band above 7 kHz in the dispersion curve data set.

8. A dipole shear wave velocity extraction system, used to implement the dipole shear wave velocity extraction method according to any one of claims 1 to 7, characterized in that: include: A data acquisition module is used to obtain acoustic array waveform data of the well logging record of the target well section on site; A data processing module processes the acoustic array waveform data to obtain a dispersion energy spectrum; extracts dispersion points from the dispersion energy spectrum to form a dispersion point cloud, clusters the dispersion point cloud to obtain clusters of different dispersion modes, and selects dispersion point sets representing the main modes of the target dipole shear wave and bending wave from the clusters; extracts regions in the dispersion energy spectrum where the energy value is greater than a set energy threshold, and processes the regions into a dispersion spectrum weight matrix; Model building module, used to build dispersion curve forward model; A determination module, for determining formation and wellbore parameters to be inverted including formation shear wave velocity within a set range; A forward modeling module, inputting the formation and wellbore parameters into a dispersion curve forward model to obtain a theoretical dispersion curve; The inversion module calculates a joint error function for each depth point in the target logging section based on the dispersion point set, the dispersion spectrum weight matrix, and the theoretical dispersion curve, and obtains the formation and wellbore parameters that minimize the joint error function. The shear wave velocity contained in the formation and wellbore parameters is used as the shear wave velocity at the current depth point. The sorting module arranges the shear wave velocity corresponding to each depth point of the target logging section according to the depth to obtain a continuous formation shear wave velocity profile.

9. The dipole shear wave velocity extraction system according to claim 8, characterized in that: The data processing module includes: The first processing module is used to process the acoustic wave array waveform data to obtain a dispersion energy spectrum; The first extraction module extracts the dispersion points in the dispersion energy spectrum to form a dispersion point cloud; Clustering module, clustering the scattered point cloud to obtain clusters with different dispersion patterns; A screening module selects the frequency dispersion point set representing the main mode of the target dipole shear wave and bending wave from the cluster; The second extraction module extracts the region in the dispersion energy spectrum where the energy value is greater than a set energy threshold; The second processing module processes the region into a dispersion spectrum weight matrix.

Images