Thin interbed longitudinal wave impedance prediction method based on deep learning

Abstract

The invention discloses a thin interbed longitudinal wave impedance prediction method based on deep learning, and the method comprises the steps: carrying out the manual collection and processing, and obtaining seismic data; performing well-seismic calibration and depth-time conversion to obtain longitudinal wave impedance and seismic data of a thin interbed section at a well point; performing time-frequency analysis and sliding a time window to obtain a time-frequency spectrum image in a local time window of a seismic channel beside a well, and taking the image as a deep learning network input sample; acquiring a longitudinal wave impedance value of the thin interbed in the local time window as a sample label; finishing the training of a thin interbed longitudinal wave impedance prediction network model, and obtaining optimal network parameters; traversing all seismic traces by using the optimal network parameters to obtain thin interbed longitudinal wave impedance prediction results in all time windows; and processing results in the overlapped time windows, and determining final spatial distribution of the longitudinal wave impedance. Compared with an existing method, the method combines logging and seismic data, can break through seismic resolution, does not need wavelet information, and is high in prediction precision and convenient to popularize.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir prediction in artificial seismic exploration, and further to the field of thin interlayer reservoir prediction, specifically a method for predicting thin interlayer longitudinal wave impedance based on deep learning. Background Art

[0002] Artificial seismic exploration refers to the artificial stimulation of seismic waves on the surface. When propagating underground, the seismic waves will be reflected and transmitted when encountering the interface of rock layers with different medium properties. The seismic waves are received by detectors on the surface to obtain seismic records. The characteristics of seismic records are related to the properties and structure of underground rock layers. By processing and interpreting seismic records, the properties and morphology of underground rock layers can be inferred. Since artificial seismic exploration is superior to other geophysical exploration methods in terms of exploration precision and accuracy, it plays an important role in oil and gas reservoir prediction.

[0003] In recent decades, with the continuous deepening of oil and gas exploration, large-scale and easy-to-find structural traps have been explored, and the oil and gas exploration targets have gradually shifted to lithologic and complex traps with small thickness and fast lithologic changes. Thin interbedded reservoirs are commonly developed in oil and gas basins in my country. Large-area thin interbedded lithologic oil and gas reservoirs are the focus of increasing reserves and production in the field of conventional oil and gas exploration at this stage. Thin interbedded strata have small single-layer thickness, frequent interbedding in the vertical direction, and fast lithologic and thickness changes in the lateral direction. It is difficult to accurately predict their longitudinal wave impedance. Therefore, conducting research on the prediction of thin interbedded longitudinal wave impedance is a very meaningful task in oil and gas exploration.

[0004] In actual artificial seismic exploration, a single-layer stratum with a thickness less than 1 / 4 of the main wavelength of the seismic wavelet is usually considered to be a thin layer, and multiple thin layers are vertically stacked to form a thin interlayer. Due to the limitation of the vertical resolution of seismic data, the P-wave impedance of thin interlayers is difficult to accurately predict. To predict P-wave impedance using seismic data, a seismic inversion method is usually used. The seismic inversion method can integrate well data, effectively improve the resolution of the results, and provide useful help for thin interlayer identification, but the inversion method requires seismic wavelet information, and the accuracy of the wavelet directly affects the accuracy of the inversion. The present invention provides a method for predicting P-wave impedance of thin interlayers based on deep learning, which uses the two-dimensional time-frequency spectrum of seismic data as input, automatically extracts features related to the P-wave impedance of thin interlayers in the time-frequency spectrum with the help of deep learning, establishes a mapping relationship between the local time-frequency spectrum and the P-wave impedance, and realizes the accurate prediction of the P-wave impedance of thin interlayers. Compared with the existing methods, this method combines logging and seismic data, breaks through the resolution of traditional post-stack seismic inversion methods, does not require seismic wavelet information, has high prediction accuracy, and is easy to promote and use. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the longitudinal wave impedance of thin interlayers based on deep learning, so as to improve the recognition and prediction capabilities of thin interlayers.

[0006] To achieve the above object, the present invention provides a method for predicting the longitudinal wave impedance of thin interlayers based on deep learning, and the present invention includes:

[0007] Collect artificial seismic data in the field and obtain post-stack seismic data through indoor processing;

[0008] Perform well-seismic calibration and deep-time conversion to obtain P-wave impedance data and seismic data of thin interbedded sections at the well point;

[0009] Perform time-frequency analysis on the seismic data of the target thin interbedded section to obtain a two-dimensional time-frequency spectrum of the seismic trace;

[0010] The two-dimensional time-frequency spectrum images in the local time window are obtained from the seismic traces near the well, and are used as input samples of the deep learning network after preprocessing;

[0011] The continuous value of the P-wave impedance of the thin interlayer in the local time window is obtained from the logging data and used as the label of the sample after normalization.

[0012] Build a deep learning network architecture, define the loss function, select the network model evaluation method, complete the thin interlayer longitudinal wave impedance prediction network model training, and obtain the optimal network parameters;

[0013] Traverse the two-dimensional time-frequency spectrum of all seismic traces to obtain the entire set of two-dimensional time-frequency spectrum images within the local time window;

[0014] Using the optimal network parameters, all sets of preprocessed two-dimensional time-frequency spectrum images are converted into the prediction results of the thin interlayer longitudinal wave impedance;

[0015] The P-wave impedance prediction results in the overlapping time window are processed to determine the final spatial distribution of the P-wave impedance of the thin interlayer.

[0016] The beneficial effect of the present invention is that, compared with the existing methods, the method proposed by the present invention combines seismic and well data, breaks through the resolution of traditional post-stack seismic inversion methods, does not require seismic wavelet information, obtains absolute longitudinal wave impedance information, has high prediction accuracy, and is easy to promote and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. It is obvious that the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0018] Figure 1 This is a flow chart of a method for inverting thin interlayer longitudinal wave impedance based on deep learning according to an embodiment of the present invention;

[0019] Figure 2 The seismic data including the thin interbedded section according to the embodiment of the present invention;

[0020] Figure 3 A seismic trace and a two-dimensional time-frequency spectrum image thereof according to an embodiment of the present invention;

[0021] Figure 4 The sample and label making method of the embodiment of the present invention;

[0022] Figure 5 A deep learning network architecture according to an embodiment of the present invention;

[0023] Figure 6 A diagram showing a ratio of data set division according to an embodiment of the present invention;

[0024] Figure 7 A network hyperparameter table of an embodiment of the present invention;

[0025] Figure 8 is a loss curve of an embodiment of the present invention;

[0026] Fig. 9 The inversion result of the thin interlayer longitudinal wave impedance obtained by commercial software in this embodiment;

[0027] Fig.10 This is the prediction result of the thin interlayer longitudinal wave impedance based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] Figure 1 FIG. 4 is a flow chart of a method for predicting the longitudinal wave impedance of thin interlayers based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0030] S101. Collect artificial seismic data on the surface, process seismic data indoors, and obtain post-stack seismic data. Specifically, in step S101, artificial earthquake refers to an earthquake that is artificially stimulated for the purpose of exploration. When conducting artificial seismic acquisition in the field, it can be three-dimensional acquisition along a surface on the surface, or it can be two-dimensional acquisition along a line, and the post-stack seismic data obtained can be three-dimensional data or two-dimensional data. An underground exploration target in a certain work area in a certain place has a two-dimensional line of wells connected with three-dimensional seismic data obtained by three-dimensional acquisition, and the application of an embodiment of the present invention is demonstrated. The underground exploration target in the area includes a set of thin interbedded sections, which are delta front deposits, sand and mudstone lithology, and a multi-layer sand and mudstone interbedded structure developed from bottom to top. The spatial relationship of the reservoir sand body is complex, the lateral changes are fast, and the heterogeneity is strong. Intercept the seismic data of the target thin interbedded section, Figure 2 This is the seismic data containing thin interbedded sections in an embodiment of the present invention. The arrows above indicate the locations of the wells, that is, there are 5 wells drilled in the area, and the well data are complete.

[0031] S102, using logging data and wellside seismic traces to perform well seismic calibration and deep-time conversion on the target layer segment, and obtain the P-wave impedance data and seismic data of the thin interlayer segment at the well point; specifically, in step S102, for the embodiment of the present invention, the well logging curves used are acoustic time difference curves and density curves. In this example, well seismic calibration adopts the method of making artificial synthetic seismic records. First, the acoustic velocity curve is obtained by reciprocal operation using the well logging acoustic time difference curve, and the P-wave impedance curve is obtained by multiplying the acoustic velocity curve with the density curve. The P-wave velocity curve is subjected to point-by-point reflection coefficient operation to obtain a discrete P-wave reflection coefficient sequence, and a sub-wave matching the main frequency of the seismic data is given to perform convolution operation with the reflection coefficient sequence to obtain a synthetic seismic record, and the well or wellside seismic traces are extracted from the actual seismic data, and the wellside seismic traces are matched and calibrated with the synthetic seismic record to determine the time-depth relationship at the well. When calibrating well seismic, focus on the matching degree of the target thin interlayer segment. The same calibration operation is performed on all the wells in the work area. Using the obtained time-depth relationship, the P-wave impedance data calculated on each well is converted into P-wave impedance data with time coordinates. The converted P-wave impedance data is consistent with the time coordinates of the seismic data and has the same time sampling interval. In this example, the time sampling interval of the seismic data is 1 ms.

[0032] S103, perform time-frequency analysis on the target thin interlayer segment seismic data, and convert the one-dimensional waveform data of the seismic trace into a two-dimensional time-frequency spectrum of the seismic trace. The time-frequency analysis in step S103 can be a linear time-frequency analysis, such as short-time Fourier transform, continuous wavelet transform and S transform, or a nonlinear time-frequency analysis, such as Wigner-Ville distribution, pseudo-Wigner-Ville distribution and Cohen-type time-frequency distribution. In this example, the wavelet transform time-frequency analysis method is used.

[0033] In step S103, the frequency range of the time-frequency analysis is based on the main frequency f of the seismic data of the target thin interlayer segment. m Its effective frequency band [f 1 , f h ] OK, low frequency starts at 0hz, high frequency takes 2f m +(f h -f i ) / 2, the frequency domain sampling interval is 1 Hz. In this example, the main frequency of the seismic data is f m It is about 30 Hz, the effective frequency band ranges from 8 Hz to 88 Hz, the frequency range of time-frequency analysis is 0 Hz to 100 Hz, and the frequency domain sampling interval is 1 Hz. Figure 3 A seismic trace and a two-dimensional time-frequency spectrum image of an embodiment of the present invention are shown, wherein the left side is a seismic trace waveform diagram, and the right side is a two-dimensional time-frequency spectrum image obtained after the seismic trace is transformed by wavelet.

[0034] S104, using a sliding time window to obtain a two-dimensional time-frequency spectrum image in a local time window from the wellbore seismic trace, and preprocessing it as an input sample for the deep learning network. Specifically, in step S104, the sliding time window has a time window length based on the main frequency f of the target thin interlayer segment seismic data. m OK, minimum is 1 / f m twice, the maximum is 1 / f m 4 times of the time sampling interval and an integer multiple of the time sampling interval; the sliding step length of the sliding window is at least one time sampling interval and at most one time window length. In this example, the main frequency of the seismic data f m It is about 30hz, the length of the sliding time window is determined to be 100ms, and the sliding step size adopts the minimum step size, that is, 1 time sampling interval is 1ms.

[0035] The preprocessing in step S104 is specifically to perform normalization and resampling on the two-dimensional time-frequency spectrum image in the local time window, wherein normalization is necessary, and resampling is selectively performed according to the requirements of recognition accuracy and computational efficiency. Resampling can be performed in the vertical and horizontal directions of the two-dimensional time-frequency spectrum image in the local time window, or only in one direction. In this example, the preprocessing is performed by normalization and resampling. Normalization is based on the mean and standard deviation of pixel values, mapping pixels to the range of [-1, 1]. After normalization, the data distribution is centered on 0, which is conducive to the optimization of gradient descent. The formula is:

[0036]

[0037] Where x is the pixel value of the image; μ is the mean used for normalization; σ is the standard deviation used for normalization; x norm is the normalized pixel value.

[0038] In this embodiment, the resampling process uses a linear interpolation method to improve the resolution of the time-frequency spectrum. The specific operation is to perform three rounds of linear interpolation on each line of time-frequency data. Suppose the original data is x=[x 1 , x 2 ,...,x n ], after one interpolation, the interpolated data is:

[0039]

[0040] Repeat the above steps three times to obtain the final interpolated data.

[0041] S105, using the same sliding time window as in step S104, obtain the continuous value of the longitudinal wave impedance in the local time window from the logging data, and perform minimum-maximum normalization processing on it as the label of the sample. The formula for minimum-maximum normalization processing is:

[0042]

[0043] Where, I is the original longitudinal wave impedance value; I max is the global maximum value of the longitudinal wave impedance, that is, the maximum value of all longitudinal wave impedance values ​​in the sample set; I min is the global minimum value of the longitudinal wave impedance, that is, the minimum value of all longitudinal wave impedance values ​​in the sample set; I norm is the normalized P-wave impedance value, which ranges from [0, 1]. The normalized P-wave impedance label is a one-dimensional vector whose values ​​are distributed in the interval [0, 1], and the length of the vector is the same as the number of time sampling points of the sample. In step S105, the continuous value of the P-wave impedance in the local time window is the P-wave impedance data with the same time sampling interval and number of samples as the seismic data in the local time window. The sliding time window is the same as that in step S104, with the same time window length and the same sliding step size. Figure 4 The sample and label making method of the embodiment of the present invention are shown. In this example, the time window length and sliding step length are 100ms and 1ms respectively. max and I min 6600m / s*g / cm 3 and 6380m / s*g / cm 3 .

[0044] In step S105, the sample label is a normalized one-dimensional vector, and the length of the vector is the same as the number of time sampling points of the sample. In this example, the label of each sample is a continuous numerical one-dimensional vector normalized to 0-1 after minimum-maximum normalization, and the vector length is 100. In this example, a single label indicates the longitudinal wave impedance value at each time sampling point position in the corresponding 100ms local time window. As the time window slides, the value change of adjacent labels indicates the vertical change of the longitudinal wave impedance.

[0045] S106, match samples and labels, divide training sets and validation sets, build deep learning network architecture, define loss function, select network model evaluation method, test network hyperparameters, complete thin interlayer longitudinal wave impedance prediction network model training, and obtain optimal network parameters. A deep learning network refers to a deep network with three or more feature extraction layers, and its feature extraction layer uses two-dimensional convolution operations. In this example, a deep network with a U-shaped structure is used. Figure 5 This is a deep learning network architecture diagram of an embodiment of the present invention. The network is an encoder-decoder structure. The encoder consists of four step-by-step downsampling modules, each of which contains two consecutive 3×3 convolution operations and one 2×2 maximum pooling operation. The decoder consists of four step-by-step upsampling modules, each of which contains one 2×2 deconvolution and two 3×3 convolution operations. The output of each layer in the encoder is spliced ​​with the upsampling result of the corresponding layer in the decoder using jump connections to achieve the fusion of image features of different scales.

[0046] Figure 6 This is a diagram showing the proportion of data sets divided according to an embodiment of the present invention. In this example, there are 5 wells with a total of 3840 labeled samples. Samples from 4 wells are used as training sets. Figure 2 The middle well was used as a blind well for verification, and its samples were used as the verification set. The training set and verification set accounted for 80% and 20% respectively.

[0047] In step S106, the loss function adopts the mean square error function (MSE):

[0048]

[0049] Among them, N is the number of samples; Y represents the label data; y represents the predicted data.

[0050] In step S106, the network model evaluation method adopts a loss curve evaluation method, and the loss curve is generated by a loss function. The loss curve is generated by the change of the value of the loss function with the number of training iterations, reflecting the optimization process of the model. Figure 7This is a network hyperparameter table of an embodiment of the present invention. In this example, the Adam optimization method is used. The ideal batch size obtained through testing is 1, the epoch is 150, and the learning rate is 0.001. Figure 8 This is the loss curve of the embodiment of the present invention. The loss curves of the training set and the validation set can converge normally, indicating that the network training has achieved the expected effect, the training of the thin interlayer longitudinal wave impedance prediction network model has been completed, and the optimal network parameters have been obtained.

[0051] S107. Use a sliding time window of the same length as in step S104 to traverse the two-dimensional time-frequency spectrum of all seismic traces, obtain the entire set of two-dimensional time-frequency spectrum images in the local time window, and preprocess all images in the set. In step S107, the length of the sliding time window is the same as the sliding time window length in step S104; the sliding step size can be the same as the sliding step size in step S104, or it can be different. In this example, a sliding time window exactly the same as that in step S104 is used, and the time window length and sliding step size are 100ms and 1ms respectively. In step S107, the same preprocessing method as in step S104 is used. In this example, normalization and resampling preprocessing are performed.

[0052] S108, using the optimal network parameters, converting all sets of preprocessed two-dimensional time-frequency spectrum images in the local time window into the prediction results of the P-wave impedance of the thin interlayer in the local time window. Specifically, in step S108, the optimal network is the deep network with the optimal weight obtained in step S106, and the conversion process refers to inputting each image in the preprocessed two-dimensional time-frequency spectrum image set in the local time window into the optimal network, and its output is the prediction result of the P-wave impedance of the thin interlayer in the corresponding local time window.

[0053] S109, according to the vertical position of the center point of the sliding time window of each seismic trace, the P-wave impedance prediction results in the overlapping time window are averaged and denormalized to determine the final spatial distribution of the P-wave impedance of the target thin interlayer. Specifically, in step S109, the sliding time window is the same sliding time window as that in step S107, with the same time window length and sliding step length; the overlapping time window refers to the overlapping part of the time window existing when the sliding step length of adjacent sliding time windows is less than the time window length, and the averaging method of the P-wave impedance prediction results in the overlapping time window is:

[0054]

[0055] Wherein, M is the number of overlapping time windows, which is related to the length of the sliding time window and the sliding step; N is the number of overlapping sampling points, which is related to the time length of the seismic trace and the length of the sliding time window; I ij is the predicted value of the longitudinal wave impedance of the i-th sampling point in the j-th overlapping time window; I iis the average P-wave impedance prediction value of the ith sampling point, which is the value in the normalized state and has a value range of [0, 1]. The existence of overlapping time windows enables multiple P-wave impedance predictions at the same underground location. The results of multiple P-wave impedance predictions are integrated to determine the final P-wave impedance value at the location, avoiding the uncertainty of a single prediction result and improving the accuracy and robustness of the prediction. In this example, M = 101 and N = 450.

[0056] Furthermore, in step S109, it is necessary to i Perform denormalization to restore to the actual longitudinal wave impedance range and obtain the final predicted longitudinal wave impedance value. The denormalization method is:

[0057] I f =I i ×(I max -I min )+I min

[0058] Among them, I f is the final longitudinal wave impedance value after denormalization; I i is the normalized average longitudinal wave impedance prediction value of the i-th sampling point; I max ,I min is the global maximum and global minimum of the longitudinal wave impedance, which is consistent with the normalization in step S105. In this example, I max and I mm 6600m / s*g / cm 3 and 6380m / s*g / cm 3 In step S109, the final spatial distribution of the P-wave impedance of the thin interlayer refers to the P-wave impedance prediction result obtained by averaging and denormalizing the P-wave impedance prediction values ​​of each sampling point in the space, which is consistent with the P-wave impedance data range at the well point. In this example, it means converting the one-dimensional vector predicted by the network with a value range between 0 and 1 into the actual P-wave impedance value, which is in the range of 6600m / s*g / cm 3 Up to 6380m / s*g / cm 3 In step S109, the obtained P-wave impedance spatial distribution has the same dimension as the post-stack seismic data in step S101, that is, when the post-stack seismic data is three-dimensional data, the obtained final spatial distribution of the P-wave impedance of the thin interlayer is three-dimensional; when the post-stack seismic data is two-dimensional data, the obtained final spatial distribution of the P-wave impedance of the thin interlayer is two-dimensional. In this example, Fig.10 The inversion result of the thin interlayer longitudinal wave impedance based on deep learning in an embodiment of the present invention shows a well line. Fig. 9 The longitudinal wave impedance inversion results obtained by the traditional method using commercial software are compared Fig. 9 and Fig.10 It can be seen that the method proposed in the present invention has obvious advantages in the identification of thin interbeds and the prediction of longitudinal wave impedance values. The thin interbed sand body structure is clearly separable, the longitudinal superposition relationship and the lateral annihilation point of the sand body are clearly revealed, and the longitudinal wave impedance value prediction accuracy is high.

[0059] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made on the basis of the present invention. Therefore, these modifications or improvements made on the basis of not departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for predicting the longitudinal wave impedance of thin interlayers based on deep learning, characterized in that The steps include: (1) Collect and process artificial seismic data on the surface to obtain post-stack seismic data; (2) Use well logging data and wellside seismic traces to perform well-seismic calibration and deep-time conversion on the target layer to obtain P-wave impedance data and seismic data of the thin interlayer at the well point; (3) Perform time-frequency analysis on the seismic data of the target thin interlayer segment and convert the one-dimensional waveform data of the seismic trace into a two-dimensional time-frequency spectrum of the seismic trace. The frequency range of the time-frequency analysis is based on the main frequency f of the seismic data of the target thin interlayer segment. m Its effective frequency band [f l ,f h ] OK, low frequency starts at 0hz, high frequency takes 2f m +(f h -f l ) / 2 or above, the frequency domain sampling interval is not greater than 1 Hz; (4) Use a sliding time window to obtain a two-dimensional time-frequency spectrum image in a local time window from the seismic trace near the well, and preprocess it as an input sample for the deep learning network. The time window length of the sliding time window is based on the main frequency f of the seismic data of the target thin interlayer segment. m OK, minimum is 1 / f m twice, the maximum is 1 / f m 4 times of the sliding step length; the minimum sliding step length is a time sampling interval and the maximum sliding step length is a time window length; (5) Using the same sliding time window as in step (4), obtain the value of the P-wave impedance in the local time window from the logging data, and perform minimum-maximum normalization processing on it as the label of the sample. The formula for minimum-maximum normalization processing is: Where, I is the original longitudinal wave impedance value; I max is the global maximum value of the longitudinal wave impedance, that is, the maximum value of all longitudinal wave impedance values ​​in the sample set; I min is the global minimum value of the longitudinal wave impedance, that is, the minimum value of all longitudinal wave impedance values ​​in the sample set; I norm is the normalized longitudinal wave impedance value, which ranges from [0,1]. The normalized longitudinal wave impedance label is a one-dimensional vector whose value is distributed in the interval [0,1]. The length of the vector is the same as the number of time sampling points of the sample; (6) Match samples and labels, divide training sets and validation sets, select network model evaluation methods, build deep learning network architecture, test network hyperparameters, complete thin interlayer longitudinal wave impedance prediction network model training, and obtain optimal network parameters; (7) using a sliding time window of the same length as in step (4) to traverse the two-dimensional time-frequency spectrum of all seismic traces, obtain a complete set of two-dimensional time-frequency spectrum images within the local time window, and preprocess all images in the set; (8) using the optimal network parameters, converting the entire set of preprocessed two-dimensional time-frequency spectrum images in the local time window into the prediction results of the longitudinal wave impedance of the thin interlayer in the local time window; (9) According to the vertical position of the center point of each seismic trace sliding time window, the P-wave impedance prediction results in the overlapping time window are averaged and denormalized to determine the final spatial distribution of the P-wave impedance of the target thin interlayer. The overlapping time window refers to the overlapping part of the adjacent sliding time windows when the sliding step length is less than the time window length. The averaging method of the P-wave impedance prediction results in the overlapping time window is: Wherein, M is the number of overlapping time windows, which is related to the length of the sliding time window and the sliding step; N is the number of overlapping sampling points, which is related to the time length of the seismic trace and the length of the sliding time window; I ij is the predicted value of the longitudinal wave impedance of the i-th sampling point in the j-th overlapping time window; I i is the predicted value of the average longitudinal wave impedance of the i-th sampling point, which is the value in the normalized state and its value range is between [0,1]; to I i Perform denormalization to restore to the actual longitudinal wave impedance range and obtain the final predicted longitudinal wave impedance value. The denormalization method is: I f =I i ×(I max -I min )+I min Among them, I f is the final longitudinal wave impedance value after denormalization; I i is the normalized average longitudinal wave impedance prediction value of the i-th sampling point; I max ,I min are the global maximum and global minimum values ​​of the longitudinal wave impedance, which are consistent with the normalization in step (5).

2. The method according to claim 1, characterized in that the artificial earthquake in step (1) is an earthquake artificially stimulated for the purpose of exploration; When conducting artificial seismic acquisition in the field, three-dimensional acquisition can be performed along a surface on the surface, or two-dimensional acquisition can be performed along a line. The post-stack seismic data obtained can be either three-dimensional data or two-dimensional data.

3. The method according to claim 1, characterized in that the logging data in step (2) includes an acoustic time difference curve and a density curve; The P-wave impedance data of the thin interlayer section at the well point refers to the P-wave impedance curve obtained by dividing the well logging density curve by the acoustic wave time difference curve at the well point, and the P-wave impedance data in the time domain obtained after deep-time conversion is consistent with the time coordinate of the seismic data and has the same time sampling interval.

4. The method according to claim 1 is characterized in that the time-frequency analysis described in step (3) can be linear time-frequency analysis or nonlinear time-frequency analysis.

5. The method according to claim 1, characterized in that the sliding time window in steps (4) and (5) is the same, having the same time window length and the same sliding step length; The method according to claim 1 is characterized in that the preprocessing described in step (4) refers to normalization and resampling of the two-dimensional time-frequency spectrum image within the local time window, and the resampling can be performed in both the vertical and horizontal directions of the two-dimensional time-frequency spectrum image within the local time window, or in only one direction.

6. The method according to claim 1 is characterized in that the continuous value of the P-wave impedance of the thin interlayer in the local time window described in step (5) is the P-wave impedance data with the same time sampling interval and number of samples as the seismic data in the local time window.

7. The method according to claim 1, characterized in that the deep learning network described in step (6) refers to a deep network with three or more feature extraction layers, and the feature extraction layer is performed using a two-dimensional convolution operation; The loss function adopts the mean square error function; The network model evaluation method adopts a loss curve evaluation method, and the loss curve is generated by a loss function.

8. The method according to claim 1 is characterized in that the length of the sliding time window described in step (7) is the same as the length of the sliding time window in step (4), and the sliding step length is the same as or different from the sliding step length in step (4); the preprocessing adopts the same method as the preprocessing described in step (4).

9. The method according to claim 1, characterized in that the optimal network in step (8) is a deep network with optimal weights obtained in step (6); The conversion refers to inputting each image in the preprocessed two-dimensional time-frequency spectrum image set in the local time window into the optimal network, and the output is the prediction result of the longitudinal wave impedance of the thin interlayer in the corresponding local time window.

10. The method according to claim 1 is characterized in that the existence of the overlapping time windows in step (9) realizes multiple P-wave impedance predictions for the same underground location, and the multiple P-wave impedance prediction results are integrated to determine the final P-wave impedance value of the location, thereby avoiding the uncertainty of a single prediction result and improving the accuracy and robustness of the prediction; The final spatial distribution of the thin interlayer longitudinal wave impedance refers to the result obtained by averaging and denormalizing the final predicted values ​​of each sampling point in space according to step (9); The spatial distribution has the same dimension as the post-stack seismic data described in step (1), that is, when the post-stack seismic data is three-dimensional data, the final spatial distribution of the thin interlayer longitudinal wave impedance obtained is three-dimensional; when the post-stack seismic data is two-dimensional data, the final spatial distribution of the thin interlayer longitudinal wave impedance obtained is two-dimensional.

Images