Well logging seismic closed loop depth network joint inversion method and device based on dtw loss

Abstract

The application provides a logging and seismic closed-loop depth network joint inversion method and device based on a DTW loss, relates to the technical field of seismic inversion, and comprises the following steps: acquiring seismic data of a region to be studied; inputting the seismic data into a pre-trained seismic wave impedance inversion model, so that the pre-trained seismic wave impedance inversion model outputs wave impedance data without time delay from the seismic data; wherein the seismic wave impedance inversion model is a convolutional neural network model. The technical scheme provided by the application can obtain logging data matched with seismic data, thereby greatly improving the seismic inversion precision.

Description

Technical Field

[0001] This invention relates to the field of seismic inversion technology, and in particular to a joint inversion method and apparatus for well logging seismic closed-loop depth networks based on DTW loss. Background Technology

[0002] In geophysical exploration, seismic inversion uses observed seismic data to infer subsurface physical parameters and reconstruct the Earth's internal structure. The inferred parameters include velocity, density, and wave impedance, which are crucial for locating and evaluating oil and gas deposits and determining drilling locations. Due to the nonlinear mapping between actual seismic data and well logging data, deep learning-based methods are generally used to build seismic inversion models. Existing techniques propose a seismic inversion method based on one-dimensional closed-loop learning, which can incorporate unlabeled seismic data, reducing the need for labeled samples. Existing techniques also improve inversion accuracy by designing an uncertainty backpropagation network to backpropagate the uncertainty of unlabeled seismic data. Furthermore, a high-dimensional closed-loop network inversion method has been proposed to improve the lateral continuity of the inversion results.

[0003] The inversion performance of deep learning-based seismic inversion methods largely depends on the quality of the training data, one factor influencing this quality being the degree of matching between well logging curves and seismic data. Due to differences in acquisition methods, absorption and attenuation in seismic data, and interference from various noise sources, mismatches exist between well logging and seismic data at the same location. Therefore, well-seismic matching is a necessary step before joint well-seismic inversion. While existing techniques can estimate time-depth conversion curves even with significant time delays between seismic and well logging data, residual well-seismic mismatches still exist. These residual mismatches manifest as small-delay perturbations in the well logging data, which also affect the accuracy of seismic inversion. Summary of the Invention

[0004] To address the problems in the prior art, this application proposes a joint inversion method and apparatus for well logging seismic closed-loop depth network based on DTW loss, which can obtain well logging data that matches seismic data, thereby greatly improving the accuracy of seismic inversion.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] In a first aspect, embodiments of the present invention provide a joint inversion method for well logging seismic closed-loop depth networks based on DTW loss, the method comprising:

[0007] Acquire seismic data for the area to be studied;

[0008] The seismic data is input into a pre-trained seismic impedance inversion model so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model.

[0009] In some embodiments, the seismic wave impedance inversion model is trained in the following manner:

[0010] Acquire first earthquake data;

[0011] Acquire first wave impedance data without time delay from the first seismic data;

[0012] The first seismic data is used as the input to the seismic impedance inversion model, and the first wave impedance data is used as the output of the seismic impedance inversion model. The seismic impedance inversion model is trained using a first preset loss function.

[0013] The first preset loss function is:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] in, λ1, λ2, and λ3 are the first preset loss function; λ1, λ2, and λ3 are preset weights. This represents the seismic wave impedance inversion model; This represents a pre-established forward modeling network model, which is the inverse operation of the seismic impedance inversion model; AI h ' represents the first wave impedance data; S represents the first seismic data; This is unlabeled seismic data.

[0021] In some embodiments, acquiring first wave impedance data that is not delayed from the first seismic data includes:

[0022] Based on the well logging data, the second wave impedance data corresponding to the first seismic data is obtained;

[0023] Iteratively perform the following operations:

[0024] S1: The first seismic data is used as the input of the seismic impedance inversion model, and the second wave impedance data is used as the output of the seismic impedance inversion model. The second preset loss function is used to make the seismic impedance inversion model output the predicted wave impedance data.

[0025] S2: Calculate the minimum error value between the predicted wave impedance data and the second wave impedance data;

[0026] S3: Based on the minimum error value, the second wave impedance data is corrected in the time dimension to obtain the corrected second wave impedance data;

[0027] S4: Calculate the cross-correlation coefficient between the corrected second-wave impedance data and the corrected second-wave impedance data in the previous iteration;

[0028] S5: Calculate the absolute value of the difference between the cross-correlation coefficient and the cross-correlation coefficient in the previous iteration;

[0029] S6: Determine whether the absolute value of the difference is less than a preset threshold; when the absolute value of the difference is less than the preset threshold, obtain the corrected second wave impedance data as the first wave impedance data and stop the iteration operation; when the absolute value of the difference is not less than the preset threshold, go to S1.

[0030] In some embodiments, calculating the minimum error between the predicted wave impedance data and the second wave impedance data includes:

[0031] The DTW algorithm is used to calculate the path probability matrix between the predicted wave impedance data and the second wave impedance data;

[0032] The path backtracking algorithm is used to find an optimal path from the path probability matrix as the minimum error value.

[0033] In some embodiments, the second preset loss function is:

[0034]

[0035] in, The second preset loss function; This represents the seismic wave impedance inversion model; AI h S is the second wave impedance data; S is the first seismic data; γ is a preset smoothing factor, and γ > 0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0036] In some embodiments, the second preset loss function is:

[0037]

[0038]

[0039]

[0040]

[0041] in, λ1, λ2, and λ3 are the second preset loss function; λ1, λ2, and λ3 are the preset weights. This represents the seismic wave impedance inversion model; This represents the forward model; AI h S represents the second wave impedance data; S represents the first seismic data. The unlabeled seismic data is denoted as γ; γ is a preset smoothing factor, and γ>0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0042] In some embodiments, the seismic impedance inversion model is a U-Net model.

[0043] Secondly, embodiments of the present invention provide a joint inversion device for well logging seismic closed-loop depth networks based on DTW loss, the device comprising:

[0044] The seismic data acquisition unit is used to acquire seismic data of the area under study.

[0045] An input unit is used to input the seismic data into a pre-trained seismic impedance inversion model, so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model.

[0046] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing program code, which, when executed by a processor, implements the well logging seismic closed-loop depth network joint inversion method based on DTW loss as described in any of the above embodiments.

[0047] Fourthly, embodiments of the present invention provide an electronic device, the electronic device including a memory and a processor, wherein the memory stores program code that can run on the processor, and when the program code is executed by the processor, it implements the well logging seismic closed-loop depth network joint inversion method based on DTW loss as described in any of the above embodiments.

[0048] The well logging seismic closed-loop depth network joint inversion method and apparatus based on DTW loss provided in this invention inputs the acquired seismic data into a pre-trained seismic impedance inversion model, so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data, which is the well logging data. In other words, the technical solution provided in this invention can directly obtain time-delayed well logging data that matches the seismic data through a deep learning convolutional neural network, thereby greatly improving the accuracy of seismic inversion. Attached Figure Description

[0049] The scope of this invention can be better understood by reading the following detailed description of exemplary embodiments in conjunction with the accompanying drawings. The accompanying drawings are:

[0050] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the process for training the seismic wave impedance inversion model in an embodiment of the present invention;

[0052] Figure 3 This is a flowchart of the method for obtaining first wave impedance data in an embodiment of the present invention;

[0053] Figure 4A This is a graph showing the changes in the correlation coefficients between various parameters during the training of the seismic wave impedance inversion model according to an embodiment of the present invention.

[0054] Figure 4B This is a graph showing the variation of MSE between the reconstructed seismic signal and the real seismic signal during the training of the seismic impedance inversion model in this embodiment of the invention.

[0055] Figure 5A A comparison diagram of the true wave impedance and the disturbance wave impedance;

[0056] Figure 5B A comparison diagram of the actual wave impedance and the wave impedance inverted using the closed-loop method;

[0057] Figure 5C This is a comparison diagram of the actual wave impedance and the wave impedance inverted by the method in this embodiment of the invention;

[0058] Figure 6AA comparison chart showing the real seismic signal and the seismic signal reconstructed using the closed-loop method;

[0059] Figure 6B This is a comparison diagram of the real seismic signal and the seismic signal reconstructed using the method of this embodiment of the invention;

[0060] Figure 7A This is the actual wave impedance data;

[0061] Figure 7B The wave impedance obtained by inversion using the closed-loop method;

[0062] Figure 7C The wave impedance obtained by inversion using the method of the embodiments of the present invention;

[0063] Figure 8A This is a real earthquake signal;

[0064] Figure 8B The seismic signal was reconstructed using a closed-loop method;

[0065] Figure 8C The seismic signal reconstructed using the method of the embodiments of the present invention;

[0066] Figure 9 This is a structural diagram of the device according to an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the present invention clearer, the implementation method of the present invention will be described in detail below with reference to the accompanying drawings and embodiments, so that the process of how the present invention uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0068] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0069] Example 1

[0070] This invention is a joint inversion method for well logging and seismic closed-loop depth networks based on DTW loss. It is a technical solution that can improve the matching degree between well logging data and seismic data when there is a mismatch between well and seismic data, and ultimately improve the seismic impedance inversion.

[0071] This invention provides a joint inversion method for well logging seismic closed-loop depth networks based on DTW loss, such as... Figure 1As shown, the joint inversion method of well logging seismic closed-loop depth network based on DTW loss described in this embodiment includes steps S101 and S102. The specific contents of these steps are described in detail below:

[0072] Step S101: Obtain seismic data for the area to be studied;

[0073] In this embodiment, the required seismic data can be obtained from the seismic data of the area to be studied, and the seismic data can be one-dimensional seismic data.

[0074] Step S102: Input the seismic data into a pre-trained seismic impedance inversion model so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model.

[0075] In this embodiment, the wave impedance data that is not delayed with the acquired seismic data can be obtained based on the pre-trained seismic wave impedance inversion model. That is, the seismic wave impedance inversion model can output well logging data that matches the acquired seismic data.

[0076] In this embodiment, the seismic impedance inversion model is trained in the following manner: acquiring first seismic data; acquiring first wave impedance data with no time delay from the first seismic data; using the first seismic data as the input to the seismic impedance inversion model, using the first wave impedance data as the output of the seismic impedance inversion model, and training the seismic impedance inversion model using a first preset loss function; wherein, the first preset loss function is:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] in, λ1, λ2, and λ3 are the first preset loss function; λ1, λ2, and λ3 are preset weights. This represents the seismic wave impedance inversion model; This represents a pre-established forward modeling network model, which is the inverse operation of the seismic impedance inversion model; AI h' represents the first wave impedance data; S represents the first seismic data; This is unlabeled seismic data.

[0084] Specifically, the first seismic data can be obtained through relevant measurement data. The first preset loss function... For open-loop loss function, W is the cycle consistency impairment function. B W represents the weights of the seismic wave impedance inversion model. F The weights W represent the weights of the forward modeling network. During the training of the aforementioned seismic impedance inversion model, the weights W are adjusted using gradient descent algorithms such as Adam. B and W F To minimize the loss function Network optimization is implemented to obtain a well-trained seismic impedance inversion model. This well-trained seismic impedance inversion model can directly output impedance data that is time-delayed from the input seismic data, i.e., directly output well logging data that matches the seismic data. This significantly improves the accuracy of subsequent joint inversion based on seismic and well logging data.

[0085] To obtain the aforementioned first-wave impedance data more accurately, this embodiment describes obtaining the first-wave impedance data without time delay from the first seismic data, including: obtaining second-wave impedance data corresponding to the first seismic data based on well logging data; such as... Figure 2 and Figure 3 As shown, the following operations are performed iteratively:

[0086] S1: The first seismic data is used as the input of the seismic impedance inversion model, and the second wave impedance data is used as the output of the seismic impedance inversion model. The second preset loss function is used to make the seismic impedance inversion model output the predicted wave impedance data.

[0087] S2: Calculate the minimum error value between the predicted wave impedance data and the second wave impedance data;

[0088] S3: Based on the minimum error value, the second wave impedance data is corrected in the time dimension to obtain the corrected second wave impedance data;

[0089] S4: Calculate the cross-correlation coefficient between the corrected second-wave impedance data and the corrected second-wave impedance data in the previous iteration;

[0090] S5: Calculate the absolute value of the difference between the cross-correlation coefficient and the cross-correlation coefficient in the previous iteration;

[0091] S6: Determine whether the absolute value of the difference is less than a preset threshold; when the absolute value of the difference is less than the preset threshold, obtain the corrected second wave impedance data as the first wave impedance data and stop the iteration operation; when the absolute value of the difference is not less than the preset threshold, go to S1.

[0092] To more accurately obtain the minimum error value between the predicted wave impedance data and the second wave impedance data, the calculation of the minimum error value between the predicted wave impedance data and the second wave impedance data in this embodiment includes: using the DTW (Dynamic Time Warping) algorithm to calculate the path probability matrix between the predicted wave impedance data and the second wave impedance data; and using the path backtracking algorithm to find an optimal path from the path probability matrix as the minimum error value.

[0093] In order to ensure that the seismic impedance inversion model learns only the morphological similarity of the wave impedance, the second preset loss function described in this embodiment of the invention adopts the Soft-DTW-loss loss function:

[0094]

[0095] in, The second preset loss function; This represents the seismic wave impedance inversion model; AI h S is the second wave impedance data; S is the first seismic data; γ is a preset smoothing factor, and γ>0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0096] To obtain the first wave impedance data more accurately, the second preset loss function used in this embodiment is:

[0097]

[0098]

[0099]

[0100]

[0101] in, λ1, λ2, and λ3 are the second preset loss function; λ1, λ2, and λ3 are the preset weights. This represents the seismic wave impedance inversion model; This represents the forward model; AI h S represents the second wave impedance data; S represents the first seismic data. The unlabeled seismic data is denoted as γ; γ is a preset smoothing factor, and γ>0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0102] In this embodiment, the seismic impedance inversion model is the U-Net model. Currently, other end-to-end network models can also be selected according to actual needs; this embodiment does not impose specific restrictions on this.

[0103] The well-logging seismic closed-loop depth network joint inversion method based on DTW loss provided in this invention inputs the acquired seismic data into a pre-trained seismic impedance inversion model. This model outputs impedance data that is time-delayed with the seismic data, which is the well-logging data. In other words, the technical solution provided in this invention can directly obtain time-delayed well-logging data that matches the seismic data through a deep learning convolutional neural network, thereby significantly improving the accuracy of seismic inversion.

[0104] Example 2

[0105] The purpose of this invention is to realize a joint inversion method for well logging seismic closed-loop depth networks based on DTW loss. This method, based on a closed-loop network structure, uses DTW loss as a relaxation constraint to calculate the regular path distance between the inversion solution and well logging curves with residual well-seismic mismatch. Through a network learning process, the well-seismic matching relationship of multiple wells is determined. The corrected well logging curves are then used as new well logging labels for further closed-loop learning, effectively improving the inversion results.

[0106] The following uses a practical application as an example to further explain in detail the joint inversion method of well logging seismic closed-loop depth network based on DTW loss provided in the embodiments of the present invention and its application principle:

[0107] The joint inversion method for well logging seismic closed-loop depth networks based on DTW loss provided in this embodiment of the invention includes:

[0108] Step 1: Train the established seismic impedance inversion model to obtain a trained seismic impedance inversion model; wherein, the seismic impedance inversion model is a convolutional neural network model.

[0109] Step 2: Obtain seismic data for the area to be studied;

[0110] Step 3: Input the earthquake data into the pre-trained seismic impedance inversion model so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the earthquake data.

[0111] The training process described in step one includes the following two stages:

[0112] (1) Under the relaxed constraint of DTW-loss, the seismic wave impedance inversion model is allowed to learn iteratively, and the predicted wave impedance data and high-frequency wave impedance (i.e., the second wave impedance data) are calculated in each round. h The path probability matrix A between the two is used to find an optimal path A* in A through a path backtracking algorithm, and the high-frequency wave impedance (i.e., the second wave impedance data) AI is then used. h Matching is performed along the optimal path A*, and the wave impedance value after each round of matching is saved, but AI is not updated. h When the iteration termination condition is met, the high-frequency impedance AI is updated using the final matched result. h This allows the closed-loop network to simultaneously learn the well-seismic matching relationship of multiple wells.

[0113] (2) Using the final updated wave impedance as the label of the training network in this stage, the closed-loop network is learned again, and finally the seismic wave impedance inversion model under well-seismic matching is established.

[0114] The following describes the two stages in detail:

[0115] Phase 1: such as Figure 2 As shown, under the relaxed constraint of DTW loss, the seismic impedance inversion model (hereinafter referred to as the inversion network) is allowed to predict the wave impedance that is most similar in shape to the wave impedance without time delay disturbance, so that the closed-loop network can learn the well-seismic matching relationship.

[0116] The seismic data used in the training process is the first seismic data. For a certain trace of one-dimensional first seismic data, it is denoted as S, and the corresponding well logging impedance data is denoted as AI. This impedance data is the impedance with residual well-seismic mismatch after well-seismic matching, that is, there is still a time delay deviation between AI and the actual impedance. The high-frequency part is AI. h The low-frequency part is AI l This low-frequency signal can be reconstructed using methods provided by related technologies; therefore, the entire inversion process does not require a known low frequency. For example... Figure 2 As shown, an inversion network is first used to learn the high-frequency impedance data AI from the first seismic data S. h To ensure that the inversion network learns only the morphological similarity of wave impedance during mapping, the Open loss function used here is the Soft-DTW-loss function.

[0117]

[0118] in, W represents the seismic wave impedance inversion model. B This represents the weight parameters of the model; AI h S represents the second wave impedance data; S represents the first seismic data; γ is a preset smoothing factor, and γ>0, with 0.015 being the optimal value verified experimentally; A represents the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them. Where:

[0119]

[0120] δ is defined as with,AI h Similarity between them, for example, is defined by Euclidean distance.

[0121] During network training, due to Constraints of the loss function, predicted wave impedance data The solution space will be relatively large, meaning that solutions similar to both the pre- and post-disturbance wave impedances will exist simultaneously. To find the solution most similar to the pre-disturbance wave impedance, a forward modeling network, which is the inverse operation of the aforementioned inversion network, is introduced. Under the strong constraints of the forward modeling process, a wave impedance that can be reconstructed from multiple morphologically similar wave impedances is found. The loss function used here is the cycle consistency loss function.

[0122]

[0123] in, W represents the forward modeling network model. F This represents the weight parameters of the model.

[0124] In addition, to make full use of unlabeled data for training, another cycle consistency loss function is introduced.

[0125]

[0126] Based on the loss function described above, define the overall loss function for this training process. for:

[0127]

[0128] Where λ1, λ2, λ3 > 0, these values ​​are used to adjust the weights of the three loss functions. The network model used in this embodiment is U-Net or other end-to-end network models. During the training process of the network model, the weights W are adjusted using gradient descent algorithms such as Adam. B and W F To minimize the loss function Implement network optimization.

[0129] like Figure 2 As shown, the overall loss function is defined. Then, during the iterative training of the network, each iteration The path probability matrix A is calculated, and an optimal path A* is found in A using a path backtracking algorithm, which yields the predicted wave impedance data. With the second wave impedance data AI h The minimum error value between them. Based on this minimum error value, the AI ​​of high-frequency logging impedance (i.e., second wave impedance data) is analyzed. h Correction is performed along the time dimension to obtain the corrected second-wave impedance data in the i-th iteration. During each iteration of training, the corrected second-wave impedance data for that round is calculated. The second wave impedance data after the previous round of correction The cross-correlation coefficient (Corr) between them is updated, but the label values ​​are not updated; that is, the label values ​​are not updated. Figure 2 The second wave impedance data AI shown in the figure h .

[0130] from Figure 4A Showing the training process and AI h The changes in the correlation coefficients between them show that... and AI h The correlation coefficient between them will first increase and then decrease. Figure 4B The display shows the normalized mean square error (MSE) between the reconstructed earthquake and the actual earthquake. It can be seen that the MSE between the reconstructed earthquake signal and the actual earthquake signal shows a trend of first decreasing and then increasing, indicating that the network will first converge to the vicinity of the true value during the learning process, and then converge to the perturbation value. Figure 4A middle and The correlation coefficient change curve between them illustrates that and The correlation coefficient between them will first increase, then stabilize, and then decrease, and when the cross-correlation coefficient Corr in the (i+1)th round... i+1 Cross-correlation coefficient Corr from the previous round iWhen the absolute value of the difference is less than 0.01, the number of iterations corresponding to the curve is exactly the same as the number of iterations corresponding to the stationary point of the black solid line in the figure. That is, the number of iterations corresponding to the maximum value of the two curves is the same, indicating that the predicted value of the inversion network has converged to the vicinity of the true value at this time.

[0131] Therefore, the iteration termination condition for this stage in this embodiment is: the cross-correlation coefficient Corr in the (i+1)th round. i+1 Cross-correlation coefficient Corr from the previous round i The absolute value of the difference is less than 0.01, meaning that the time-depth relationship curve has essentially stopped changing at this point, and the second wave impedance data after the last round of correction is used. To update the initial second-wave impedance data AI h Since the task at this stage is simply to find structurally similar solutions to the wave impedance, the number of iterations is not required. Experiments have shown that a maximum of 100 iterations is optimal.

[0132] Phase Two: The training process for this phase is similar to Phase One, except that there is no label update process, and the final updated wave impedance from Phase One is used as the label for this phase. That is, in this phase, the second wave impedance data after the last round of correction in Phase One is used as the first wave impedance data without time delay from the first seismic data to train the inversion network, thus obtaining a trained seismic wave impedance inversion model. Furthermore, the open-loop loss function uses... The loss function consists of two loss terms:

[0133]

[0134]

[0135] A new cycle consistency loss term is added as follows:

[0136]

[0137] The other two cycle consistency loss functions are defined the same as in Phase 1, representing the overall loss function during this training process. for:

[0138]

[0139] That is, the loss function used in this stage is

[0140]

[0141]

[0142]

[0143]

[0144]

[0145]

[0146] Among them, λ1, λ2, and λ3 are pre-set weights, all of which are greater than 0; This represents the seismic wave impedance inversion model; This represents a pre-established forward modeling network model, which is the inverse operation of the seismic impedance inversion model; AI h ' represents the first wave impedance data; S represents the first seismic data; This is unlabeled seismic data.

[0147] The task at this stage is to find a solution that matches the corrected second-wave impedance data. Experiments have shown that 2500 iterations are optimal.

[0148] The following describes the experimental simulation results using the method provided in the embodiments of the present invention:

[0149] To verify the effectiveness and superiority of this invention, we applied the method proposed in this embodiment to synthetic seismic data. The inversion results of this embodiment are shown below with reference to Figures 5-8 and Table 1.

[0150] First of all Figure 7A The actual wave impedances shown are perturbed at four logging locations to produce perturbed wave impedances, as follows: Figures 7A-7C As shown, the logging positions of these four impedance channels are channel 50, channel 150, channel 250, and channel 400, respectively. Figures 5A-5C , Figures 6A-6B The results show a comparison between the one-dimensional inverted wave impedance and the reconstructed seismic data at one well location under two different methods. Figure 5B This indicates that the wave impedance obtained by the closed-loop method is similar in shape to the wave impedance of the disturbance, while from... Figure 5C As can be seen, the wave impedance obtained by inversion in this embodiment is closer to the true wave impedance. Through comparison... Figure 6A and Figure 6B The reconstructed seismic signal obtained by the method in this embodiment is closer to the real seismic signal, while the wave impedance obtained by the closed-loop method is basically consistent with the wave impedance of the disturbance, and the quality of the reconstructed seismic signal is also worse. Figures 7A-7C , Figures 8A-8CThe comparison diagram of the inverted 2D wave impedance profile and the reconstructed 2D seismic profile is shown. Referring to Table 1 below, the normalized mean square error (MSE) between the inverted wave impedance and the true value in this embodiment is 0.009, while the MSE between the wave impedance inverted by the closed-loop method and the true value is 0.0169. Similarly, the MSE between the reconstructed seismic data and the true value in this embodiment is 0.0680, while the MSE between the reconstructed seismic data and the true value using the closed-loop method is 0.0179. Comparatively, the MSE between the inverted wave impedance and the reconstructed seismic data in this embodiment is smaller. Therefore, it can be seen that the method in this embodiment can correct the remaining well-seismic mismatch even when there is still an incomplete match in the well-seismic data, ultimately improving the wave impedance inversion effect.

[0151] Table 1 Comparison of Inverted Wave Impedance and MSE of Reconstructed Seismic Signal

[0152]

[0153] This embodiment presents a joint inversion method for well logging seismic closed-loop depth networks based on DTW loss. Building upon the closed-loop network structure, this embodiment employs DTW loss as a new open-loop loss function. This relaxed constraint expands the solution space of the inversion problem. Furthermore, under the strong constraints of the forward modeling network, an inversion solution capable of reconstructing the original seismic event is found within this solution space. This inversion solution is the one most morphologically similar to the well logging curve without time delay perturbation. While inverting the error, the DTW loss also calculates the regularized path distance between the inversion solution and well logging curves with residual well-seismic mismatch. In other words, the well-seismic matching relationship of multiple wells is determined through the network learning process. Then, the well logging curves with residual well-seismic mismatch are corrected using this regularized path. The corrected well logging curves are used as new well logging labels for further closed-loop learning, effectively improving the inversion performance. The technical solution provided by this embodiment can improve seismic inversion performance while correcting residual well-seismic mismatch.

[0154] Example 3

[0155] Corresponding to the above-described method embodiments, the present invention also provides a joint inversion device for well logging seismic closed-loop depth networks based on DTW loss, such as... Figure 9 As shown, the apparatus described in this embodiment includes:

[0156] Seismic data acquisition unit 201 is used to acquire seismic data of the area to be studied.

[0157] The input unit 202 is used to input the seismic data into a pre-trained seismic impedance inversion model so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model.

[0158] The apparatus described in this embodiment is also used to train the seismic wave impedance inversion model in the following manner:

[0159] Acquire first earthquake data;

[0160] Acquire first wave impedance data without time delay from the first seismic data;

[0161] The first seismic data is used as the input to the seismic impedance inversion model, and the first wave impedance data is used as the output of the seismic impedance inversion model. The seismic impedance inversion model is trained using a first preset loss function.

[0162] The first preset loss function is:

[0163]

[0164]

[0165]

[0166]

[0167]

[0168]

[0169] in, λ1, λ2, and λ3 are the first preset loss function; λ1, λ2, and λ3 are preset weights. This represents the seismic wave impedance inversion model; This represents a pre-established forward modeling network model, which is the inverse operation of the seismic impedance inversion model; AI h ' represents the first wave impedance data; S represents the first seismic data; This is unlabeled seismic data.

[0170] The apparatus described in this embodiment is also used to acquire first wave impedance data that is not delayed from the first seismic data in the following manner:

[0171] Based on the well logging data, the second wave impedance data corresponding to the first seismic data is obtained;

[0172] Iteratively perform the following operations:

[0173] S1: The first seismic data is used as the input of the seismic impedance inversion model, and the second wave impedance data is used as the output of the seismic impedance inversion model. The second preset loss function is used to make the seismic impedance inversion model output the predicted wave impedance data.

[0174] S2: Calculate the minimum error value between the predicted wave impedance data and the second wave impedance data;

[0175] S3: Based on the minimum error value, the second wave impedance data is corrected in the time dimension to obtain the corrected second wave impedance data;

[0176] S4: Calculate the cross-correlation coefficient between the corrected second-wave impedance data and the corrected second-wave impedance data in the previous iteration;

[0177] S5: Calculate the absolute value of the difference between the cross-correlation coefficient and the cross-correlation coefficient in the previous iteration;

[0178] S6: Determine whether the absolute value of the difference is less than a preset threshold; when the absolute value of the difference is less than the preset threshold, obtain the corrected second wave impedance data as the first wave impedance data and stop the iteration operation; when the absolute value of the difference is not less than the preset threshold, go to S1.

[0179] The apparatus described in this embodiment is also used to calculate the minimum error value between the predicted wave impedance data and the second wave impedance data in the following manner:

[0180] The DTW algorithm is used to calculate the path probability matrix between the predicted wave impedance data and the second wave impedance data;

[0181] The path backtracking algorithm is used to find an optimal path from the path probability matrix as the minimum error value.

[0182] In this embodiment, the second preset loss function is:

[0183]

[0184] in, The second preset loss function; This represents the seismic wave impedance inversion model; AI h S is the second wave impedance data; S is the first seismic data; λ is a preset smoothing factor, and γ>0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0185] In this embodiment, the second preset loss function is:

[0186]

[0187]

[0188]

[0189]

[0190] in, λ1, λ2, and λ3 are the second preset loss function; λ1, λ2, and λ3 are the preset weights. This represents the seismic wave impedance inversion model; This represents the forward model; AI h S represents the second wave impedance data; S represents the first seismic data. The unlabeled seismic data is denoted as γ; γ is a preset smoothing factor, and γ>0; A is the path probability matrix to be searched; A k,k This represents all possible paths from point (0,0) to point (k,k); express and AI h The cost matrix between them.

[0191] In this embodiment, the seismic wave impedance inversion model is the U-Net model.

[0192] For details regarding the working principle, workflow, and specific implementation methods of the aforementioned device, please refer to the specific implementation method of the well logging seismic closed-loop depth network joint inversion method based on DTW loss provided by this invention. The same technical content will not be described in detail here.

[0193] The well-logging seismic closed-loop depth network joint inversion device based on DTW loss provided in this embodiment of the invention inputs the acquired seismic data into a pre-trained seismic impedance inversion model, so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data, and this impedance data is the well-logging data. In other words, the technical solution provided in this embodiment of the invention can directly obtain time-delayed well-logging data that matches the seismic data through a deep learning convolutional neural network, thereby greatly improving the accuracy of seismic inversion.

[0194] Example 4

[0195] According to an embodiment of the present invention, a computer-readable storage medium is also provided, wherein program code is stored on the computer-readable storage medium, and when the program code is executed by a processor, it implements the joint inversion method of well logging seismic closed-loop depth network based on DTW loss as described in any of the above embodiments.

[0196] Example 5

[0197] According to an embodiment of the present invention, an electronic device is also provided, the electronic device including a memory and a processor, wherein the memory stores program code that can run on the processor, and when the program code is executed by the processor, it implements the joint inversion method of well logging seismic closed-loop depth network based on DTW loss as described in any of the above embodiments.

[0198] The well logging seismic closed-loop depth network joint inversion method, apparatus, storage medium, and electronic equipment based on DTW loss provided in this invention input the acquired seismic data into a pre-trained seismic impedance inversion model, so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data, which is the well logging data. In other words, the technical solution provided in this invention can directly obtain time-delayed well logging data that matches the seismic data through a deep learning convolutional neural network, thereby greatly improving the accuracy of seismic inversion.

[0199] The present invention also has the advantage that, based on the closed-loop network results, the DTW loss is defined as a new open-loop loss function. During the closed-loop learning process of the network, not only are seismic inversion and wave impedance forward modeling established, but the well-seismic matching relationship of multiple wells is also learned. According to the application effect, the present invention can improve the seismic inversion performance while correcting the residual well-seismic mismatch.

[0200] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0201] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0202] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0203] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0204] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A joint inversion method for well logging seismic closed-loop depth networks based on DTW loss, characterized in that, The method includes: Acquire seismic data for the area to be studied; The seismic data is input into a pre-trained seismic impedance inversion model so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model. The seismic wave impedance inversion model is trained using the following method: Acquire first earthquake data; Acquire first wave impedance data without time delay from the first seismic data; The first seismic data is used as the input to the seismic impedance inversion model, and the first wave impedance data is used as the output of the seismic impedance inversion model. The seismic impedance inversion model is trained using a first preset loss function. The first preset loss function is: in, The first preset loss function; , , Pre-set weights; This represents the seismic wave impedance inversion model; This represents a pre-established forward modeling network model, which is the inverse operation of the seismic impedance inversion model; This refers to the impedance data of the first wave. This refers to the first earthquake data; This is unlabeled seismic data; The acquisition of first wave impedance data without time delay from the first seismic data includes: Based on the well logging data, the second wave impedance data corresponding to the first seismic data is obtained; Iteratively perform the following operations: S1: The first seismic data is used as the input of the seismic impedance inversion model, and the second wave impedance data is used as the output of the seismic impedance inversion model. The second preset loss function is used to make the seismic impedance inversion model output the predicted wave impedance data. S2: Calculate the minimum error value between the predicted wave impedance data and the second wave impedance data; S3: Based on the minimum error value, the second wave impedance data is corrected in the time dimension to obtain the corrected second wave impedance data; S4: Calculate the cross-correlation coefficient between the corrected second-wave impedance data and the corrected second-wave impedance data in the previous iteration; S5: Calculate the absolute value of the difference between the cross-correlation coefficient and the cross-correlation coefficient in the previous iteration; S6: Determine whether the absolute value of the difference is less than a preset threshold; when the absolute value of the difference is less than the preset threshold, obtain the corrected second wave impedance data as the first wave impedance data and stop the iteration operation; when the absolute value of the difference is not less than the preset threshold, go to S1.

2. The joint inversion method for well logging seismic closed-loop depth network based on DTW loss according to claim 1, characterized in that, The calculation of the minimum error value between the predicted wave impedance data and the second wave impedance data includes: The DTW algorithm is used to calculate the path probability matrix between the predicted wave impedance data and the second wave impedance data; The path backtracking algorithm is used to find an optimal path from the path probability matrix as the minimum error value.

3. The joint inversion method for well logging seismic closed-loop depth network based on DTW loss according to claim 2, characterized in that, The second preset loss function is: in, The second preset loss function; This represents the seismic wave impedance inversion model; This refers to the impedance data of the second wave. This refers to the first earthquake data; As a preset smoothing factor, and A is the path probability matrix to be searched; This represents all possible paths from point (0, 0) to point (k, k); express and The cost matrix between them.

4. The joint inversion method for well logging seismic closed-loop depth network based on DTW loss according to claim 2, characterized in that, The second preset loss function is: in, The second preset loss function; , , The pre-set weights; This represents the seismic wave impedance inversion model; This represents the forward model; This refers to the impedance data of the second wave. This refers to the first earthquake data; The unlabeled seismic data; As a preset smoothing factor, and A is the path probability matrix to be searched; This represents all possible paths from point (0, 0) to point (k, k); express and The cost matrix between them.

5. The joint inversion method for well logging seismic closed-loop depth network based on DTW loss according to claim 1, characterized in that, The seismic impedance inversion model is the U-Net model.

6. A joint inversion device for a well-logging seismic closed-loop depth network based on DTW loss for implementing the joint inversion method for well-logging seismic closed-loop depth networks based on DTW loss as described in any one of claims 1-5, characterized in that, The device includes: The seismic data acquisition unit is used to acquire seismic data of the area under study. An input unit is used to input the seismic data into a pre-trained seismic impedance inversion model, so that the pre-trained seismic impedance inversion model outputs impedance data that is time-delayed with the seismic data; wherein, the seismic impedance inversion model is a convolutional neural network model.

7. A computer-readable storage medium storing program code thereon, characterized in that, When the program code is executed by the processor, it implements the well logging seismic closed-loop depth network joint inversion method based on DTW loss as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores program code that can run on the processor. When the program code is executed by the processor, it implements the well logging seismic closed-loop depth network joint inversion method based on DTW loss as described in any one of claims 1 to 5.

Images