Intelligent seismic inversion, multi-task intelligent seismic inversion result prediction method and device
By combining deep learning networks and implicit label objective functions, a multi-task intelligent seismic inversion method is constructed, which solves the problem of seismic inversion under the condition of limited drilling and logging data, and achieves higher resolution and stability, making it suitable for oil and gas exploration of complex geological targets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA NAT PETROLEUM CORP
- Filing Date
- 2022-08-09
- Publication Date
- 2026-05-19
AI Technical Summary
In oil and gas exploration, conventional seismic inversion methods are insufficient to meet the detailed exploration needs of complex geological targets such as deep, thin interbedded layers, and tight/low permeability layers. In particular, intelligent seismic inversion results are poor when drilling and logging data are limited.
A deep learning network was used to train a seismic inversion learning neural network. By utilizing an implicit label objective function and a multi-task learning method, a neural network was constructed to output P-wave impedance and reservoir parameters through a combination of shared network units and task network units, thus overcoming the disadvantage of having few logging labels.
It improves the resolution and stability of seismic inversion, enhances the universality of the inversion method, effectively handles unlabeled non-wellpoint seismic trace data, and improves the accuracy and consistency of the inversion results.
Smart Images
Figure CN117631018B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysical oil and gas exploration technology, and in particular to an intelligent seismic inversion method, a multi-task intelligent seismic inversion result prediction method and device. Background Technology
[0002] Seismic exploration is a crucial method for predicting and identifying oil and gas geological targets in the field of geophysical oil and gas exploration. Its main tasks include seismic data acquisition, processing, and interpretation. In the acquisition phase, seismic waves generated at the surface propagate to subsurface reflecting interfaces or geological targets, then reflect back to the surface and are received by geophones. The received seismic data thus contains important information about these interfaces or targets. In the processing phase, the data received by the surface geophones undergoes static correction, denoising, and migration to obtain the post-stack seismic data volume. In the interpretation phase, the seismic phase axes of the post-stack seismic data volume represent stratigraphic interfaces from the same geological period. Oil and gas exploration work generally revolves around the target stratigraphic system between two stratigraphic interfaces. However, the seismic data of the target stratigraphic system is still reflection information from geological bodies or oil and gas reservoir interfaces. Seismic inversion technology is needed to convert this information into seismic impedance information with stratigraphic significance, thereby identifying oil and gas geological targets. Therefore, seismic inversion is a key technology for oil and gas reservoir prediction and fluid detection.
[0003] Over the past two decades, various methods have been developed for seismic inversion, including recursive inversion, sparse pulse inversion, model-based inversion, statistical seismic inversion, waveform indicator inversion, and intelligent seismic inversion. With the expansion of oil and gas exploration and development targets in my country, deep, thin interbedded layers, tight / low-permeability reservoirs, stratigraphy / lithology, and unconventional oil and gas reservoirs have gradually become the focus of exploration and development. Oil and gas geological targets are becoming more refined and concealed, and seismic responses are becoming more complex and subtle. Due to limitations in seismic data resolution, conventional seismic inversion methods such as recursive inversion, sparse pulse inversion, and model-based inversion also have relatively limited resolution, making it difficult to meet the current needs of refined geological target exploration and development. While statistical seismic inversion and waveform indicator inversion have improved the resolution of seismic inversion to some extent, they also have certain theoretical and application limitations. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide an intelligent seismic inversion method, a multi-task intelligent seismic inversion result prediction method and apparatus that overcome or at least partially solve the above problems.
[0005] In a first aspect, embodiments of the present invention provide an intelligent seismic inversion method, comprising:
[0006] Randomly select a predetermined number of sample data from the earthquake dataset; the sample data is an earthquake trace dataset.
[0007] The sample data is used as input to a preset seismic inversion learning neural network to train the seismic inversion learning neural network, and the parameters of the neural network are updated according to a preset objective function; the objective function is an implicit label objective function.
[0008] Repeat the above steps until the objective function reaches the preset convergence condition to obtain the earthquake inversion learning model.
[0009] The earthquake inversion learning neural network is constructed in the following manner:
[0010] A neural network is constructed, with the seismic trace dataset as input and the P-wave impedance inversion result corresponding to the seismic trace dataset as output.
[0011] The neural network also outputs elasticity parameters or reservoir parameters.
[0012] The neural network includes:
[0013] A shared network unit and multiple task network units connected to the shared network unit;
[0014] The output of the shared network unit serves as the input of multiple task network units.
[0015] The shared network unit comprises multiple convolutional network units in series; each task network unit comprises multiple convolutional network units in series, and the number is one less than the number of convolutional network units in the shared network unit; each convolutional network unit comprises a convolutional layer and a ReLU activation function.
[0016] The objective function is constructed in the following manner:
[0017] The neural network output for constructing the well bypass channel Actual label data of the well bypass A function of the squared distance of the L2 norm;
[0018] Constructing an implicit objective function for forward modeling of non-well-side channel seismic data, i.e., seismic data... With forward modeling data The squared distance of the second norm; where f() is the forward modeling operator function, w is the seismic wavelet, Δ is the difference operator, and log() is the logarithmic operator function;
[0019] Construct an implicit objective function to determine the relationship between output parameters of non-well bypass channels;
[0020] The neural network output of the constructed well bypass will be... Actual label data of the well bypass The objective function is the sum of the L2 norm distance squared function, the implicit objective function of forward modeling of non-well bypass seismic data, and the implicit objective function of the relationship between the output parameters of the non-well bypass.
[0021] The sample data is used as input to a preset seismic inversion learning neural network. The network is trained, and its parameters are updated according to a preset objective function. This process is repeated until the objective function reaches a preset convergence condition, resulting in the seismic inversion learning model. Specifically, this includes:
[0022] The stochastic gradient descent method is used to randomly select N samples from the entire earthquake dataset each time, input them into the neural network, and optimize the neural network parameters according to the objective function to complete the update of the neural network parameters;
[0023] Repeat the above process until the preset number of training iterations are completed, so that the objective function converges to the minimum value.
[0024] The number of training iterations k = K, and K = 5 * round(N) all / N);
[0025] Where N all This represents the total number of seismic traces in the entire seismic dataset, and round() is a function for rounding to the nearest integer.
[0026] Secondly, embodiments of the present invention provide a method for predicting multi-task intelligent seismic inversion results, including:
[0027] Input the seismic trace dataset to be predicted into the already trained seismic inversion learning model;
[0028] The seismic inversion results, as well as optional elastic parameters or reservoir parameters, are obtained through the output of the seismic inversion learning model.
[0029] The earthquake inversion learning model is obtained using the intelligent earthquake inversion method described above.
[0030] Thirdly, embodiments of the present invention provide an intelligent seismic inversion device, comprising:
[0031] The sample generation module is used to randomly extract a preset number of sample data from the earthquake dataset; the sample is an earthquake trace dataset.
[0032] The training module is used to train the earthquake inversion learning neural network by taking the sample data as input and updating the parameters of the neural network according to the preset objective function. The above steps are repeated until the objective function reaches the preset convergence condition to obtain the earthquake inversion learning model.
[0033] Fourthly, embodiments of the present invention provide a prediction device for multi-task intelligent seismic inversion results, comprising:
[0034] The input module is used to input the seismic trace dataset to be predicted into the trained seismic inversion learning model;
[0035] The prediction module is used to obtain seismic inversion results, as well as optional elastic parameters or reservoir parameters, through the output of the seismic inversion learning model.
[0036] The earthquake inversion learning model is obtained using the intelligent earthquake inversion method described above.
[0037] Fifthly, embodiments of the present invention provide a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned intelligent seismic inversion method or the aforementioned method for predicting multi-task intelligent seismic inversion results.
[0038] Sixthly, a non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by a processor, is capable of implementing the aforementioned intelligent seismic inversion method, or the aforementioned method for predicting multi-task intelligent seismic inversion results.
[0039] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:
[0040] This invention utilizes a seismic dataset to train a deep learning network. An implicitly labeled objective function is used as the target function to update the parameters of the neural network. Training terminates when the objective function reaches a preset convergence condition, resulting in a seismic inversion learning model. This invention combines deep learning with an implicitly labeled objective function to establish a theoretical relationship between the output and input of the neural network. This allows unlabeled non-wellpoint seismic traces to participate in neural network training, overcoming the disadvantage of insufficient well logging labels. It effectively solves the bottleneck problem of insufficient labels in intelligent seismic inversion, enhancing the stability of the inversion network and the universality of the inversion method.
[0041] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.
[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0044] Figure 1 This is a flowchart of the intelligent seismic inversion method in an embodiment of the present invention;
[0045] Figure 2 This is a schematic diagram of the neural network in an embodiment of the present invention;
[0046] Figure 3 This is a schematic diagram of a shared network unit in an embodiment of the present invention;
[0047] Figure 4 A flowchart illustrating the construction and training method of the seismic inversion learning model in an example provided by an embodiment of the present invention;
[0048] Figure 5 A flowchart illustrating the prediction method for multi-task intelligent seismic inversion results provided in an embodiment of the present invention;
[0049] Figure 6A and 6B The images show P-wave impedance well profiles obtained using conventional seismic inversion methods and prediction methods using multi-task intelligent seismic inversion results, respectively.
[0050] Figure 7 This is a porosity well profile obtained by the prediction method using multi-task intelligent seismic inversion results provided in an embodiment of the present invention.
[0051] Figure 8 This is a well profile diagram of dolomite content obtained by the prediction method using multi-task intelligent seismic inversion results provided in an embodiment of the present invention.
[0052] Figure 9A and 9B The images show the mean plane diagrams of the target layer P-wave impedance obtained using conventional seismic inversion methods and prediction methods using multi-task intelligent seismic inversion results, respectively.
[0053] Figure 10This is a porosity target layer mean plane map obtained by the prediction method of multi-task intelligent seismic inversion results provided in this embodiment of the invention;
[0054] Figure 11 This is a planar map of the target layer mean value of dolomite content obtained by the prediction method of multi-task intelligent seismic inversion results provided in this embodiment of the invention. Detailed Implementation
[0055] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0056] The inventors of this invention have discovered that, in the prior art, intelligent seismic inversion using big data and artificial intelligence technologies has great potential in improving efficiency, accuracy, and resolution. However, the quantity and completeness of the well logging label dataset directly determine the quality of the intelligent seismic inversion results. Existing intelligent seismic inversion methods often employ commonly used sample-label pair patterns for neural network training and application. For work areas with a large number of wells and abundant well logging information, this type of intelligent seismic inversion often yields good results. However, in the early stages of oil and gas exploration, there is often a limited amount of drilling and well logging data. For work areas with fewer wells and limited well logging information, implementing effective intelligent seismic inversion remains a significant challenge.
[0057] An intelligent seismic inversion method provided in this embodiment of the invention, referring to... Figure 1 As shown, it includes:
[0058] S11. Randomly select a preset number of sample data from the earthquake dataset; the sample data is an earthquake trace dataset;
[0059] S12. The sample data is used as input to a preset seismic inversion learning neural network to train the seismic inversion learning neural network, and the parameters of the neural network are updated according to a preset objective function; wherein, the objective function is an implicit label objective function.
[0060] S13. Repeat the above steps to determine if the objective function has reached the preset convergence condition.
[0061] S14. When the objective function reaches the preset convergence condition, the earthquake inversion learning model is obtained.
[0062] This invention utilizes a seismic dataset to train a deep learning network. An implicitly labeled objective function is used as the target function to update the parameters of the neural network. Training terminates when the objective function reaches a preset convergence condition, resulting in a seismic inversion learning model. This invention combines deep learning with an implicitly labeled objective function to establish a theoretical relationship between the output and input of the neural network. This allows unlabeled non-wellpoint seismic traces to participate in neural network training, overcoming the disadvantage of insufficient well logging labels. It effectively solves the bottleneck problem of insufficient labels in intelligent seismic inversion, enhancing the stability of the inversion network and the universality of the inversion method.
[0063] Furthermore, the earthquake inversion learning neural network is constructed in the following manner:
[0064] A neural network is constructed, with the seismic trace dataset as input and the P-wave impedance inversion result corresponding to the seismic trace dataset as output.
[0065] Preferably, the seismic inversion learning model provided in this embodiment of the invention can be a multi-task learning model, that is, in addition to outputting P-wave impedance inversion results, it can also optionally output elastic parameters or reservoir parameters. By putting multiple related elastic parameters or reservoir parameters together for neural network learning and training, multiple tasks influence and promote each other, improving the network's generalization ability and stability, and overcoming the disadvantage of having few well logging labels.
[0066] Reference Figure 2 As shown, the aforementioned neural network may include, for example:
[0067] A shared network unit and multiple task network units connected to the shared network unit;
[0068] The output of the shared network unit serves as the input to multiple task network units.
[0069] Reference Figure 3 As shown, the shared network unit contains multiple convolutional network units in series; each task network unit contains multiple convolutional network units in series, and the number is one less than the number of convolutional network units in the shared network unit; each convolutional network unit contains a convolutional layer and a ReLU activation function.
[0070] The objective function described above can be constructed from the sum of the following three terms:
[0071] The first step is to construct the neural network output for the well bypass. Actual label data of the well bypass A function of the squared distance of the L2 norm;
[0072] The second step is to construct an implicit objective function for forward modeling of non-well-side seismic data, i.e., seismic data... With forward modeling data The squared distance of the second norm; where f() is the forward modeling operator function, w is the seismic wavelet, Δ is the difference operator, and log() is the logarithmic operator function;
[0073] The third step is to construct an implicit objective function for the relationship between output parameters of non-well bypass channels;
[0074] The sum of the first, second, and third terms mentioned above is taken as the objective function.
[0075] Specifically, in the implementation of steps S12-S14 above, the stochastic gradient descent method can be used. Each time, N samples are randomly selected from the entire earthquake dataset, input into the neural network, and the neural network parameters are optimized according to the objective function to complete the update of the neural network parameters.
[0076] Repeat the above process until the preset number of training iterations are completed, so that the objective function converges to the minimum value.
[0077] The number of training iterations can be preset. Each time, N samples are randomly selected to train the neural network. The parameters of the neural network are updated using the objective function. Then, N samples are randomly selected again, and the above training process is repeated until the preset number of iterations is reached. At this point, the objective function converges to the minimum value.
[0078] The number of training iterations is k = K, and K = 5 * round(N) all / N);
[0079] Where N all This represents the total number of seismic traces in the entire seismic dataset, and round() is a function for rounding to the nearest integer.
[0080] The following example illustrates the intelligent seismic inversion method described above.
[0081] In this example, the construction and training methods of the seismic inversion learning model refer to... Figure 4 As shown, it includes the following steps:
[0082] Step 401: Construct a multi-task deep learning neural network for seismic inversion. The input to the deep learning neural network is a seismic trace dataset, and the output of the deep learning neural network is the P-wave impedance inversion result of the corresponding seismic trace dataset, as well as other elastic or reservoir parameters to be predicted. In this example, it includes three deep learning neural network outputs: the P-wave impedance inversion result, the reservoir porosity parameter, and the reservoir dolomite content parameter.
[0083] The input to the deep learning neural network is a seismic trace dataset, hereinafter referred to as {s}. i |i=1,2,3,…,N} represents, where s i A sample point, or a seismic trace, is a column vector whose length is the number of sample points in the seismic trace, hereinafter referred to as N. s In this embodiment, i represents the sample point number in the dataset, and N represents the number of sample points in the dataset, which is also the batch size of the deep learning neural network. For example, N = 1000 in this embodiment.
[0084] The outputs of the deep learning neural network in step 401, namely the P-wave impedance inversion result, reservoir porosity parameters, and reservoir dolomite content parameters, are all outputs of a multi-task deep learning neural network. However, the P-wave impedance inversion result has a special use in constructing the objective function of the deep learning neural network. Therefore, it will be used in the following... These represent the different types of output results, where p = 1, 2, ..., N p N represents the numbering of different types of output results. p N represents the number of types of output results from a deep learning neural network, including those related to longitudinal wave impedance inversion. When the deep learning neural network only outputs the longitudinal wave impedance inversion result, N represents the number of types of output results. p =1, in this embodiment of the invention, N p =3, Let N be a sample point of the p-th output result. It is a column vector, and the length of the vector is the number of sampling points in the seismic trace. N will be used in the following text. s express.
[0085] For ease of explanation, the following convention is used when p=1. The output of the longitudinal wave impedance inversion result of the deep learning neural network is defined as p = 2, ..., N. p time The output of the deep learning neural network is for other elasticity or reservoir parameters that it wants to predict, where i is the sample point number in the dataset. In this embodiment of the invention, For reservoir porosity parameters, The parameter for the dolomite content in the reservoir, the meaning and value of N are explained above;
[0086] The multi-task deep learning neural network for seismic inversion described in step 401 has a batch size of N. The deep learning neural network consists of one shared network unit and multiple task network units, with the number of task network units being N. pThe input to each task network unit is the output of the shared network unit, and the output of each task network unit is as described above. Where p = 1, 2, ..., N p This is used to assign numbers to different types of output results.
[0087] The aforementioned shared network unit can, for example, be composed of multiple convolutional network units connected in series, with the number of convolutional network units ranging from 3 to 10. Each convolutional network unit consists of a convolutional layer and a ReLU activation function. The core parameters of each convolutional layer are the kernel parameters, the kernel stride, and the zero-padding method. The kernel parameters include the kernel length, the number of input channels, and the number of output channels. Generally, the kernel length can be set to 3, the number of input channels is equal to the number of output channels of the previous convolutional network unit, the number of output channels is set to twice the number of input channels, the kernel stride is set to 1, and the zero-padding method is set to zero-padding. In this embodiment of the invention, the shared network unit can be composed of 6 convolutional network units connected in series.
[0088] Each task network unit described above is composed of multiple convolutional network units connected in series, with the number of convolutional network units being one less than the number of convolutional network units in the shared network unit. Each convolutional network unit consists of a convolutional layer and a ReLU activation function. The core parameters of each convolutional layer are the kernel parameters, the kernel stride, and the zero-padding method. The kernel parameters include the kernel length, the number of input channels, and the number of output channels. Generally, the kernel length is set to 3, the number of input channels is equal to the number of output channels of the previous convolutional network unit, the number of output channels is set to half the number of input channels, the kernel stride is set to 1, and the zero-padding method is set to zero-padding. In this embodiment of the invention, each task network unit may, for example, be composed of 5 convolutional network units connected in series.
[0089] Step 402: Construct the objective function of the deep learning neural network based on the input and output of the deep learning neural network in step 401.
[0090] The input to the deep learning neural network in step 402 is {s} i |i=1,2,3,…,N}, where the symbols have the same meaning as above. When constructing the objective function of the deep learning neural network, only the sample points of the well-side seismic traces have known labels. In this embodiment of the invention, the input / output sample points are selected based on whether they are well-side seismic traces, and the deep learning neural network is input to {s}. i |i=1,2,3,…,N} is divided into two subsets and in, For non-well bypass input subset, For the well-side input subset, In this embodiment of the invention, for example, there may be four well-side seismic trace sample points. When training the deep learning neural network, the well-side trace input subset contains these sample points.
[0091] The output of the deep learning neural network in step 402 is The symbols have the same meaning as described above. When constructing the objective function of the deep learning neural network, only the sample points of the well-side seismic traces have known labels. The following section defines the deep learning neural network output based on whether the input / output sample points are well-side seismic traces. Divided into two subsets and in, This is a subset of the output from the non-well bypass channel. Output a subset for the well bypass channel. In this embodiment of the invention, there are four well-side seismic trace sample points. When training the deep learning neural network, the well-side trace output subset contains the sample points.
[0092] The objective function of the deep learning neural network in step 402 can be, for example, the following formula:
[0093]
[0094] The number is the sum of three expressions, where the first term is the output of the deep learning neural network of the well bypass. Actual label data of the well bypass The second term is the L2 norm distance squared; the second term is the implicit objective function for forward modeling of non-wellside seismic data, i.e., seismic data. With forward modeling data The squared distance of the second norm, where f() is the forward modeling operator function, w is the seismic wavelet, Δ is the difference operator, and log() is the logarithmic operator function; the third term is the implicit objective function for the non-well bypass output parameter relationship, where... For output parameters and The theoretical or empirical relational objective function, The closer the value is to 0, the more it indicates and The more it conforms to theoretical or empirical relationships, the better. and When there is no theoretical or empirical relationship between them, It should also be set to 0.
[0095] In this embodiment of the invention, the longitudinal wave impedance inversion result output by the deep learning neural network With reservoir porosity parameters There is an empirical relation objective function
[0096] Step 403: Train the deep learning neural network.
[0097] In step 403, stochastic gradient descent can be used to train the deep learning neural network. Specifically, the steps could be as follows: Let the deep learning neural network be trained k = 1, 2, 3, ..., K times. Starting from k = 1, N samples {x} are randomly selected from the entire earthquake dataset each time. i The sequence |i=1,2,3,…,N} is input into the deep learning neural network, and the parameters of the deep learning neural network are optimized according to the objective function J, completing one parameter update of the deep learning neural network. This process is repeated until k=K. For example, K=5*round(N) all The objective function J will converge to its minimum value if N / N is constant. all This represents the total number of seismic traces in the entire seismic dataset, and `round()` is the rounding function. In this embodiment of the invention, N = 1000, N... all =947*1263=1196061, K=59805.
[0098] This invention also provides a method for predicting multi-task intelligent seismic inversion results. Referring to Figure 5, the method includes:
[0099] S51. Input the seismic trace dataset to be predicted into the trained seismic inversion learning model;
[0100] S52. Through the output of the earthquake inversion learning model, obtain the earthquake inversion results, as well as optional elastic parameters or reservoir parameters;
[0101] The earthquake inversion learning model described above was obtained using the intelligent earthquake inversion method mentioned earlier.
[0102] Taking S401-S403 as an example again, the first inversion result can be predicted using the trained seismic inversion learning model, and elastic parameters or reservoir parameters can be predicted optionally.
[0103] Specifically, the point vectors of seismic data from the entire seismic dataset can be sequentially input into a trained deep learning neural network to obtain the P-wave impedance inversion result, as well as other desired elastic or reservoir parameter outputs. This embodiment includes three deep learning neural network outputs: the P-wave impedance inversion result, reservoir porosity parameters, and reservoir dolomite content parameters.
[0104] Figure 6A It is a P-wave impedance well profile obtained using conventional seismic inversion methods in existing technologies. Figure 6BThis is a P-wave impedance well profile obtained by the prediction method using multi-task intelligent seismic inversion results provided in an embodiment of the present invention. (Comparison) Figure 6A and Figure 6B As can be seen, the P-wave impedance obtained by the prediction method using multi-task intelligent seismic inversion results proposed in this embodiment of the invention has higher longitudinal and lateral resolution, and it matches the actual well logging results better. Figure 7 This is a porosity well profile obtained by the prediction method using multi-task intelligent seismic inversion results provided in an embodiment of the present invention. Figure 8 This is a well profile diagram of dolomite content obtained by the prediction method using multi-task intelligent seismic inversion results provided in an embodiment of the present invention. Figure 7 and Figure 8 The output results are in excellent agreement with the corresponding actual logging results. Furthermore, the overall profile characteristics are consistent with the geological understanding of the dolomite karst reservoir in the target layer. Figure 9A It is a mean plane map of the target layer of P-wave impedance obtained using conventional seismic inversion methods in existing technology. Figure 9B This is a mean plane map of the target layer's P-wave impedance obtained by the prediction method using multi-task intelligent seismic inversion results provided in this embodiment of the invention. (Comparison) Figure 9A and Figure 9B As can be seen, the longitudinal wave impedance obtained by the embodiment of the present invention has a higher lateral resolution. Figure 10 This is a porosity target layer mean plane map obtained by the prediction method of multi-task intelligent seismic inversion results provided in this embodiment of the invention. Figure 11 This is a planar map of the target layer mean value of dolomite content obtained by the prediction method of multi-task intelligent seismic inversion results provided in this embodiment of the invention. Figure 9B , Figure 10 , Figure 11 The output results show strong heterogeneity in planar features, which is more consistent with the geological understanding of dolomite karst reservoirs in the target layer.
[0105] Based on the same inventive concept, embodiments of the present invention also provide an intelligent seismic inversion device and a prediction device for multi-task intelligent seismic inversion results. Since the principles by which these devices solve problems are similar to the aforementioned intelligent seismic inversion method and the prediction method for multi-task intelligent seismic inversion results, the implementation of this device can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.
[0106] An intelligent seismic inversion device provided in this embodiment of the invention includes:
[0107] The sample generation module is used to randomly extract a preset number of sample data from the earthquake dataset; the sample is an earthquake trace dataset.
[0108] The training module is used to train the earthquake inversion learning neural network by taking the sample data as input and updating the parameters of the neural network according to the preset objective function. The above steps are repeated until the objective function reaches the preset convergence condition to obtain the earthquake inversion learning model.
[0109] This invention provides a multi-task intelligent seismic inversion result prediction device, comprising:
[0110] The input module is used to input the seismic trace dataset to be predicted into the trained seismic inversion learning model;
[0111] The prediction module is used to obtain seismic inversion results, as well as optional elastic parameters or reservoir parameters, through the output of the seismic inversion learning model.
[0112] The earthquake inversion learning model is obtained using the intelligent earthquake inversion method described above.
[0113] This invention provides a server, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned intelligent seismic inversion method or the aforementioned method for predicting multi-task intelligent seismic inversion results.
[0114] This invention provides a non-transitory computer-readable storage medium that, when executed by a processor, enables the implementation of the aforementioned intelligent seismic inversion method or the aforementioned method for predicting multi-task intelligent seismic inversion results.
[0115] The intelligent seismic inversion method, multi-task intelligent seismic inversion result prediction method and device provided in the embodiments of the present invention introduce an implicit label objective function and establish a theoretical relationship between the output and input ends of the neural network, enabling unlabeled non-wellpoint seismic traces to participate in neural network training, thus overcoming the disadvantage of having few well logging labels. In addition, an implicit objective function for the relationship between the output parameters of the neural network is introduced. The theoretical or empirical relationship between these parameters constrains the training of the neural network, further improving the generalization ability and stability of the network. The embodiments of the present invention can effectively solve the bottleneck problem of few labels in intelligent seismic inversion and enhance the universality of intelligent inversion methods.
[0116] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0120] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A smart seismic inversion method, characterized in that, include: Randomly select a predetermined number of sample data from the earthquake dataset; The sample is a seismic trace dataset; The sample data is used as input to a preset seismic inversion learning neural network to train the seismic inversion learning neural network, and the parameters of the neural network are updated according to a preset objective function; the objective function is an implicit label objective function. Repeat the above steps until the objective function reaches the preset convergence condition to obtain the earthquake inversion learning model; The objective function is constructed in the following manner: The neural network output for constructing the well bypass channel Actual label data of the well bypass A function of the squared distance of the L2 norm; Constructing an implicit objective function for forward modeling of non-well-side channel seismic data, i.e., seismic data... With forward modeling data The squared distance of the second norm; where Forward operand functions, For seismic wavelets, For difference operators, It is a logarithmic operator function; Construct an implicit objective function to determine the relationship between output parameters of non-well bypass channels; The neural network output of the constructed well bypass will be... Actual label data of the well bypass The objective function is the sum of the L2 norm distance squared function, the implicit objective function of forward modeling of non-well bypass seismic data, and the implicit objective function of the relationship between the output parameters of the non-well bypass. Number the sample points in the dataset; This is used to assign numbers to different types of output results.
2. The method as described in claim 1, characterized in that, The earthquake inversion learning neural network is constructed in the following manner: A neural network is constructed, with the seismic trace dataset as input and the P-wave impedance inversion result corresponding to the seismic trace dataset as output.
3. The method as described in claim 2, characterized in that, The neural network also outputs elasticity parameters or reservoir parameters.
4. The method as described in claim 2, characterized in that, The neural network includes: A shared network unit and multiple task network units connected to the shared network unit; The output of the shared network unit serves as the input of multiple task network units.
5. The method as described in claim 4, characterized in that, The shared network unit comprises multiple convolutional network units in series; each task network unit comprises multiple convolutional network units in series, and the number is one less than the number of convolutional network units in the shared network unit; each convolutional network unit comprises a convolutional layer and a ReLU activation function.
6. The method as described in claim 1, characterized in that, The sample data is used as input to a preset seismic inversion learning neural network. The network is trained, and its parameters are updated according to a preset objective function. This process is repeated until the objective function reaches a preset convergence condition, resulting in the seismic inversion learning model. Specifically, this includes: The stochastic gradient descent method is used to randomly sample data from the entire earthquake dataset each time. Each sample is input into the neural network, and the neural network parameters are optimized according to the objective function to complete the update of the neural network parameters; Repeat the above process until the preset number of training iterations are completed, so that the objective function converges to the minimum value.
7. The method as described in claim 6, characterized in that, The number of training sessions ,and ; in The total number of seismic traces in the entire seismic dataset. This is a function for rounding to the nearest integer.
8. A method for predicting multi-task intelligent seismic inversion results, characterized in that, include: Input the seismic trace dataset to be predicted into the already trained seismic inversion learning model; The earthquake inversion results are obtained by using the output of the earthquake inversion learning model. The earthquake inversion learning model is obtained using the intelligent earthquake inversion method as described in any one of claims 1-7.
9. An intelligent seismic inversion device, characterized in that, include: The sample generation module is used to randomly extract a preset number of sample data from the earthquake dataset; The sample is a seismic trace dataset; The training module is used to train the earthquake inversion learning neural network by taking the sample data as input and updating the parameters of the neural network according to the preset objective function. The above steps are repeated until the objective function reaches the preset convergence condition to obtain the earthquake inversion learning model. The objective function is constructed in the following manner: The neural network output for constructing the well bypass channel Actual label data of the well bypass A function of the squared distance of the L2 norm; Constructing an implicit objective function for forward modeling of non-well-side channel seismic data, i.e., seismic data... With forward modeling data The squared distance of the second norm; where Forward operand functions, For seismic wavelets, For difference operators, It is a logarithmic operator function; Construct an implicit objective function to determine the relationship between output parameters of non-well bypass channels; The neural network output of the constructed well bypass will be... Actual label data of the well bypass The objective function is the sum of the L2 norm distance squared function, the implicit objective function of forward modeling of non-well bypass seismic data, and the implicit objective function of the relationship between the output parameters of the non-well bypass. Number the sample points in the dataset; This is used to assign numbers to different types of output results.
10. A device for predicting multi-task intelligent seismic inversion results, characterized in that, include: The input module is used to input the seismic trace dataset to be predicted into the trained seismic inversion learning model; The prediction module is used to obtain earthquake inversion results from the output of the earthquake inversion learning model; The earthquake inversion learning model is obtained using the intelligent earthquake inversion method as described in any one of claims 1-7.
11. A server, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the intelligent seismic inversion method as described in any one of claims 1-7, or the prediction method for multi-task intelligent seismic inversion results as described in claim 8.
12. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor, the intelligent seismic inversion method as described in any one of claims 1-7, or the prediction method for multi-task intelligent seismic inversion results as described in claim 8, can be implemented.