An intelligent lithology identification method based on frequency-phase characteristics

Abstract

The present invention belongs to the technical field of petroleum exploration, and specifically relates to an intelligent lithology identification method based on frequency-phase characteristics. The method comprises: performing time-frequency analysis based on inversion on seismic data to extract its frequency-phase characteristics; making data-label pairs; carrying out deep network algorithm design and optimization, selecting a network suitable for frequency-phase intelligent lithology identification; training the preferred network with frequency-phase characteristic data and lithology labels; performing time-frequency analysis on three-dimensional seismic data, extracting the frequency-phase characteristics of the seismic data, and inputting them into the trained deep network to obtain output three-dimensional lithology information, and finally realizing intelligent identification of lithology. The method of the present invention effectively improves the accuracy and reliability of facies identification in reservoir prediction, and has important economic and social benefits.

Description

Technical Field

[0001] The invention belongs to the technical field of petroleum exploration, and in particular relates to an intelligent lithology identification method based on frequency phase characteristics. Background Art

[0002] Time-frequency analysis is widely used in geophysical exploration. Currently, the commonly used time-frequency analysis methods include short-time Fourier transform, wavelet transform, S transform, Hilbert-Huang transform, quadratic time-frequency distribution, matching pursuit decomposition, etc. The short-time Fourier transform is equivalent to the cross-correlation of the seismic trace with the sine basis in the moving time window, and the wavelet transform is the cross-correlation of the seismic trace with the wavelet dictionary. These two methods usually require a trade-off between time and frequency resolution, which makes their application effect unsatisfactory in many cases; the S transform is the inheritance and development of the short-time Fourier transform and the wavelet transform. The S transform eliminates the selection of window function and improves the defect of fixed window width, and has some improvements in time-frequency resolution; the Hilbert-Huang transform can decompose the signal into a limited number of intrinsic mode functions, and then obtain the time-frequency spectrum of the signal. Its disadvantage is that it is sensitive to noise; the quadratic time-frequency distribution has very serious cross terms in the time-frequency plane, which leads to its not widespread application in seismic exploration; matching pursuit decomposition is to decompose the seismic signal into a weighted sum of many so-called atomic signals, and the time-frequency spectrum of the seismic trace can be obtained by superimposing the spectrum of each atomic signal.

[0003] There are many deep network models. Common deep networks (DNN) are suitable for data classification and fitting, convolutional neural networks (CNN) are suitable for image classification and fitting, U-Net networks are suitable for image semantic segmentation, recurrent neural networks are suitable for time series analysis, and deep belief neural networks are suitable for classification processing. The core of deep learning is the BP algorithm. The key to the BP algorithm is the gradient descent method. The gradient descent method starts from the loss function, and the activation function plays an important role in it.

[0004] Generally speaking, the deeper the network, the more information can be obtained and the richer the features. However, according to experiments, as the network deepens, the optimization effect becomes worse, and the accuracy of test data and training data decreases. This is because the deepening of the network will cause gradient explosion and gradient disappearance. At present, there is a solution to this phenomenon: normalize the input data and the data in the middle layer, but this method is only useful for networks with dozens of layers. When the network goes deeper, this method is useless.

[0005] Chinese patent application CN109919184A discloses a method for intelligent identification of complex lithologies in multiple wells based on logging data. The method first determines the target logging data file and performs format conversion and normalization preprocessing. Then, according to the known lithology of the key coring wells in the entire area in the coring section, feature screening and / or feature combination expansion are performed on the logging curve data to obtain logging curve data sensitive to lithology. The logging curve data sensitive to lithology response are then labeled and calibrated to form a sample database. At the same time, the unlabeled logging curve data in the entire area are formed into a database to be tested. Then, the data in the sample database is used in combination with several machine learning algorithms for machine learning training to automatically establish several lithology identification models. The optimal model is selected through classification performance evaluation rules, and the optimal model is used to perform lithology prediction on the data in the database to be tested to achieve intelligent identification of complex lithologies in multiple wells in the entire area.

[0006] Chinese patent application CN112801035A discloses a method for intelligent lithology identification based on knowledge and data dual drive, which includes inputting multiple images of the same surrounding rock at different angles into corresponding parallel target detection and positioning models, performing non-maximum suppression operation on the lithology characteristics and surrounding rock position information corresponding to the surrounding rock images output by the target detection and positioning model; obtaining element content test information at the surrounding rock position after the non-maximum suppression operation, and then outputting the predicted lithology characteristics through the lithology identification model based on data and knowledge drive, performing dot multiplication between the predicted lithology characteristics and the lithology characteristics after the non-maximum suppression operation, and outputting the lithology category and surrounding rock position candidate information through a fully connected layer; and performing non-maximum suppression operation on the lithology category and position candidate information to obtain the final lithology category and surrounding rock position.

[0007] Chinese invention patent CN110031493B discloses a lithology intelligent identification system and method based on image and spectral technology, which includes a rock shape analysis system, an image recognition system, a sample processing system, a spectral analysis system, and a central analysis control system; the central analysis control system determines the final lithology of the sample based on the rock identification results transmitted by the image recognition system and the analysis results transmitted by the spectral analysis system. The invention uses spectral technology to identify the mineral content and type, and finally fuses the spectral analysis results and image recognition results to give the rock lithology, which greatly improves the accuracy of rock lithology identification.

[0008] Currently, there is no method that combines time-frequency analysis with deep learning to achieve intelligent identification and classification of lithology. Summary of the invention

[0009] The main purpose of the present invention is to provide an intelligent lithology identification method based on frequency-phase characteristics, extract the frequency-phase characteristics of seismic data based on inversion high-resolution time-frequency analysis, preferably a depth network suitable for intelligent lithology division, and combine the frequency-phase with the depth network, thereby effectively improving the accuracy and reliability of lithology identification in reservoir prediction, which has important economic and social benefits.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] The present invention provides an intelligent lithology identification method based on frequency phase characteristics, which comprises the following steps:

[0012] Conduct inversion-based time-frequency analysis on seismic data to extract its frequency and phase characteristics;

[0013] Make data-label pairs;

[0014] Carry out deep network algorithm design and optimization, and select a network suitable for frequency-phase intelligent lithology identification;

[0015] The optimized network is trained using the frequency phase characteristic data and lithology labels;

[0016] The three-dimensional seismic data is subjected to time-frequency analysis to extract the frequency and phase characteristics of the seismic data and input them into the trained deep network to obtain the output three-dimensional lithology information, ultimately achieving intelligent identification of lithology.

[0017] Furthermore, based on the inversion time-frequency analysis, the method for extracting its frequency-phase characteristics is as follows: taking trigonometric functions as basis functions, a forward equation describing the relationship between the spectrum, trigonometric functions and seismic signals is established; based on the least squares principle, the spectrum analysis is transformed into an inverse problem; the spectrum of the signal can be obtained by iteratively solving the matrix equation; the seismic spectrum signal is analyzed point by point to obtain the least squares time-frequency analysis result of the seismic trace.

[0018] Furthermore, based on the inversion time-frequency analysis, extracting its frequency phase characteristics specifically includes the following steps:

[0019] The forward modeling process of calculating the signal from the desired time-frequency spectrum and trigonometric functions of different frequencies is defined as follows:

[0020] S=E·f (1)

[0021] Where S represents the signal to be analyzed, E represents the matrix composed of trigonometric functions of different frequencies, and f represents the spectrum vector to be calculated;

[0022] Formula (1) is solved by the L2 norm regularized constrained least squares method. In order to improve the stability and uniqueness of the solution, it is necessary to add constraints to the solution process. Finally, the time-frequency analysis result can be obtained:

[0023] fe =E'(E E'+aΩ) -1 S (2)

[0024] Among them, f e It is the high-resolution time-frequency analysis result based on inversion, Ω represents the constraint diagonal matrix, and a is a weight coefficient used to adjust the size of the constraint;

[0025] The frequency and phase characteristics of seismic data are extracted using the following formula:

[0026]

[0027] Among them, A(ω,t) represents the amplitude spectrum of the frequency, P(ω,t) represents the phase spectrum of the frequency, ω represents the frequency, t represents the time, arctan() represents the inverse tangent function, and I(f e ) represents the imaginary part of the spectrum, R(f e ) represents the real part of the spectrum. The amplitude spectrum calculated by this formula is greater than or equal to 0, and the phase range is between -180 degrees and 180 degrees.

[0028] Furthermore, the method of making data-label pairs comprises the following steps:

[0029] By making synthetic seismic records, the seismic and logging data are matched in time and depth or an accurate time-depth relationship is obtained, the logging data in the depth domain is converted into the time domain, and the data after well-seismic matching are sampled accordingly to form a training data set, and a point-point strategy or a point-window strategy is used to generate data-label pairs.

[0030] Furthermore, a point-window strategy is adopted to select the sampling time step and time window size, and the seismic data in the time window are matched with the well logging data corresponding to the window center time. The seismic data in time are subjected to time-frequency analysis to obtain a two-dimensional frequency-phase diagram, which together with the well logging lithology data form a data-label pair.

[0031] Furthermore, training the selected network with the frequency-phase characteristic data and lithology labels includes the following steps: taking the extracted high-resolution frequency-phase map as input and the corresponding labels as output, training the selected deep network, and continuously debugging the key parameters of the deep network during the training process.

[0032] Furthermore, the key parameters of the deep network include: activation function, optimization algorithm, batch size, and learning rate.

[0033] Furthermore, the residual network is selected as the network for frequency-phase intelligent lithology identification.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] The method of the present invention effectively combines frequency-phase characteristics and deep learning to achieve lithology identification. The extracted high-resolution frequency-phase map is used as input data, and the corresponding lithology label is used as output to train the selected deep network. During the training process, it is necessary to continuously debug the key parameters such as the activation function and optimization method of the deep network to further improve the generalization ability of the deep network, and finally obtain a trained deep network. The method of the present invention combines frequency-phase with deep network for the first time, effectively improving the accuracy and reliability of reservoir prediction, and has important economic and social benefits.

[0036] The method of the present invention adopts time-frequency analysis based on inversion to extract frequency-phase features. The method is characterized by high time-frequency resolution and stable extracted frequency-phase features, which is conducive to the next step of analysis and application.

[0037] The method of the present invention not only uses the amplitude spectrum characteristics, or only uses the phase spectrum characteristics, but comprehensively utilizes the frequency-phase characteristics of seismic data, namely the amplitude spectrum and phase spectrum characteristics, thereby greatly reducing the ambiguity of lithology identification and improving the objective accuracy of lithology identification.

[0038] Based on the analysis of the existing network, the present invention comprehensively considers two main factors: one is whether there is a better strategy for gradient vanishing; the other is whether there are other strategies besides Dropout that can better solve the problem of overfitting in frequency-phase feature extraction. The residual network suitable for frequency-phase feature extraction is selected, which can not only obtain a deep network, but also has a very good final lithology classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0040] Figure 1 This is a flow chart of the intelligent lithology identification method based on frequency phase characteristics described in Example 1 of the present invention;

[0041] Figure 2 The frequency characteristic diagram of the frequency phase characteristic described in Example 1 of the present invention;

[0042] Figure 3 It is a phase characteristic diagram in the frequency-phase characteristic described in Example 1 of the present invention;

[0043] Figure 4 This is the lithology identification result of the method described in Example 1 of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0046] In order to enable those skilled in the art to more clearly understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below in conjunction with specific embodiments.

[0047] Example 1

[0048] like Figure 1 As shown, the intelligent lithology identification method based on frequency phase characteristics includes the following steps:

[0049] Step 1. Perform inversion-based time-frequency analysis on seismic data to extract its frequency phase characteristics;

[0050] At present, most commonly used time-frequency analysis methods are based on Fourier transform. Relatively speaking, time-frequency analysis based on inversion has higher resolution and better stability, and can extract high-resolution frequency-phase characteristics of seismic data. In order to obtain a solution with higher resolution, it is necessary to improve the existing time-frequency analysis method and derive the objective function based on inversion time-frequency analysis.

[0051] 1) Define the forward modeling process of calculating the signal from the time-frequency spectrum to be determined and the trigonometric functions of different frequencies as follows:

[0052] S=E·f (1)

[0053] Among them, S represents the signal to be analyzed, E represents the matrix composed of trigonometric functions of different frequencies, and f represents the spectrum vector to be calculated.

[0054] Formula (1) is solved by the L2 norm regularized constrained least squares method. In order to improve the stability and uniqueness of the solution, it is necessary to add constraints to the solution process. Finally, the time-frequency analysis result can be obtained:

[0055] f e =E'(E E'+aΩ) -1 S (2)

[0056] Among them, f eIt is the high-resolution time-frequency analysis result based on inversion, Ω represents the constraint diagonal matrix, and a is a weight coefficient used to adjust the size of the constraint.

[0057] 2). Solving the objective function equations

[0058] There are two main methods for solving the objective function equation group. One is the completely nonlinear solution, such as simulated annealing, genetic algorithm, Monte Carlo, bee colony algorithm, etc.; the other is the linearization of nonlinear problems, such as gradient descent method, conjugate gradient method, LSQR method, etc.

[0059] The conjugate gradient method (CG) is a method between the steepest descent method and the Newton method. It only uses the first-order derivative information, but overcomes the disadvantage of the slow convergence of the steepest descent method and avoids the disadvantage of the Newton method that it needs to store and calculate the Hessian and matrix and invert them. The CG method is used to solve formula (2) and finally obtain the time-frequency analysis results of the seismic data.

[0060] 3) Frequency and phase characteristics of seismic data

[0061] After obtaining the high-resolution time-frequency analysis results based on inversion, e After that, it is necessary to further extract the frequency and phase information, namely the frequency-phase characteristics, and the calculation formula is as follows:

[0062]

[0063] Among them, A(ω,t) represents the amplitude spectrum of the frequency, P(ω,t) represents the phase spectrum of the frequency, ω represents the frequency, t represents the time, arctan() represents the inverse tangent function, and I(f e ) represents the imaginary part of the spectrum, R(f e ) represents the real part of the spectrum. The amplitude spectrum calculated by this formula is greater than or equal to 0, and the phase range is between -180 degrees and 180 degrees.

[0064] Step 2. Make data-label pairs;

[0065] Input logging data mainly includes velocity, density, gamma, natural potential and other curves, but is not limited to these curves. Velocity and density curves are used to generate synthetic seismic records and calibrate seismic data, while gamma and natural potential curves are used to classify lithology.

[0066] By making synthetic seismic records, seismic and well logging data are matched in time and depth or an accurate time-depth relationship is obtained, and the well logging data in the depth domain is converted to the time domain. The matching degree between well logging and seismic is a key factor affecting the accuracy of the results. Well logging data and seismic data often have different sampling rates. As the input and output of the deep learning model, well logging data and seismic data also need to be matched and sampled, such as 0.5ms, 1ms, 2ms, etc.

[0067] According to the point-window strategy, appropriate sampling time step and time window size are selected, the seismic data in the time window are matched with the well logging data corresponding to the window center time, and the seismic data in the time are subjected to time-frequency analysis to obtain a two-dimensional frequency phase diagram, which together with the well logging lithology data forms a data-label pair.

[0068] Step 3. Design and optimize the deep network algorithm and select a network suitable for frequency-phase intelligent lithology identification;

[0069] Based on the analysis of the existing network, two main factors are comprehensively considered: one is whether there is a better strategy for gradient vanishing; the other is whether there are other strategies besides Dropout that can better solve the problem of overfitting in frequency-phase feature extraction; and select a network suitable for frequency-phase feature extraction.

[0070] In order to solve the problem of gradient explosion and gradient disappearance caused by the deepening of the network, the residual network (R-Net) was adopted. The idea of ​​the R-Net network structure mainly comes from the idea of ​​residual and jump connection. R-Net introduced the residual network structure. This sub-network can form a very deep network by stacking. Through this residual network structure, the network layer can be made very deep, and the final lithology classification effect is also very good.

[0071] Step 4. Train the selected network using the frequency phase characteristic data and lithology labels:

[0072] The frequency-phase characteristics and deep learning are effectively combined to realize the identification of lithology. The extracted high-resolution frequency-phase map is used as input and the corresponding lithology label is used as output to train the selected deep network. During the training process, it is necessary to continuously debug the key parameters of the deep network, such as the activation function and optimization method, to further improve the generalization ability of the deep network and finally obtain a well-trained deep network.

[0073] In this step, the following key parameters are mainly considered for selection and debugging:

[0074] (1) Selection and debugging of optimization algorithms

[0075] By selecting different optimization algorithms such as SGD, SGDM, AdaGrad, Adam, etc., the learning efficiency and training accuracy of the training process can be improved;

[0076] (2) Batch size selection and debugging

[0077] The batch size is the number of samples required to calculate the gradient. If it is too small, it will lead to inefficiency and failure to converge. If it is too large, it will cause the memory to be unable to support it. When the batch size increases to a certain extent, its descent direction changes very little, so the batch size is a very important parameter, because the batch size needs to be continuously adjusted during the training process to find the optimal batch size suitable for frequency-phase deep learning.

[0078] (3) Learning rate selection and debugging

[0079] The learning rate is a key factor in training deep learning models. Gradients have direction and magnitude. In back propagation, the gradient descent algorithm multiplies a variable called the learning rate to determine the location of the next point. If the learning rate is set too small, the convergence speed will be very slow. If the learning rate is set too large, it will pass the lowest point and cannot reach the lowest point. Because the learning rate needs to be constantly adjusted during the training process to find the optimal learning rate.

[0080] Step 5. Perform time-frequency analysis on the three-dimensional seismic data, extract the frequency and phase characteristics of the seismic data, input them into the trained deep network, obtain the output three-dimensional lithology information, and finally realize the intelligent identification of lithology.

[0081] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An intelligent lithology identification method based on frequency phase characteristics, characterized in that: The following steps are involved: Conduct inversion-based time-frequency analysis on seismic data to extract its frequency and phase characteristics; Make data-label pairs; Carry out deep network algorithm design and optimization, and select a network suitable for frequency-phase intelligent lithology identification; The optimized network is trained using the frequency phase characteristic data and lithology labels; Perform time-frequency analysis on three-dimensional seismic data, extract the frequency and phase characteristics of the seismic data, input them into the trained deep network, and output three-dimensional lithology information, ultimately achieving intelligent identification of lithology; Based on the inversion time-frequency analysis, the method of extracting its frequency-phase characteristics is as follows: taking the trigonometric function as the basis function, a forward equation describing the relationship between the spectrum, trigonometric function and seismic signal is established. Based on the least squares principle, the spectrum analysis is transformed into an inverse problem. The spectrum of the signal can be obtained by iteratively solving the matrix equation. The seismic spectrum signal is analyzed point by point to obtain the least squares time-frequency analysis result of the seismic trace. Based on the inversion time-frequency analysis, extracting its frequency phase characteristics specifically includes the following steps: The forward modeling process of calculating the signal from the desired time-frequency spectrum and trigonometric functions of different frequencies is defined as follows: S=E·f(1) Where S represents the signal to be analyzed, E represents the matrix composed of trigonometric functions of different frequencies, and f represents the spectrum vector to be calculated; Formula (1) is solved by the L2 norm regularized constrained least squares method. In order to improve the stability and uniqueness of the solution, it is necessary to add constraints to the solution process. Finally, the time-frequency analysis result can be obtained: in e =E'(EE'+aΩ) -1 S(2) Among them, f e It is the high-resolution time-frequency analysis result based on inversion, Ω represents the constraint diagonal matrix, and a is a weight coefficient used to adjust the size of the constraint; The frequency and phase characteristics of seismic data are extracted using the following formula: Among them, A(ω,t) represents the amplitude spectrum of the frequency, P(ω,t) represents the phase spectrum of the frequency, ω represents the frequency, t represents the time, arctan() represents the inverse tangent function, and I(f e ) represents the imaginary part of the spectrum, R(f e ) represents the real part of the spectrum. The amplitude spectrum calculated by this formula is greater than or equal to 0, and the phase range is between -180 degrees and 180 degrees.

2. The method according to claim 1, characterized in that The method of making data-label pairs comprises the following steps: By making synthetic seismic records, the seismic and logging data are matched in time and depth to obtain an accurate time-depth relationship, the logging data in the depth domain is converted into the time domain, and the data after well-seismic matching are sampled accordingly to form a training data set, and a point-point strategy or point-window strategy is used to generate data-label pairs.

3. The method according to claim 2, characterized in that A point-window strategy is adopted to select the sampling time step and time window size. The seismic data in the time window are matched with the well logging data corresponding to the window center time. The seismic data in the time are subjected to time-frequency analysis to obtain a two-dimensional frequency-phase diagram, which together with the well logging lithology data forms a data-label pair.

4. The method according to claim 1, characterized in that: Training the selected network with frequency-phase characteristic data and lithology labels includes the following steps: using the extracted high-resolution frequency-phase map as input and the corresponding labels as output, training the selected deep network, and continuously debugging the key parameters of the deep network during the training process.

5. The method according to claim 4, characterized in that The key parameters of the deep network include: activation function, optimization algorithm, batch size, and learning rate.

6. The method according to claim 1 or 5, characterized in that: The residual network is selected as the network for frequency-phase intelligent lithology identification.

Images