A seismic inversion model training method, device, equipment and medium

Abstract

Embodiments of the present application provide a seismic inversion model training method and device, equipment and medium, which belong to the technical field of seismic inversion. The target seismic inversion model is obtained by iteratively training the seismic inversion model to be trained based on a training sample set. In one iteration training process, the following operations are performed: multiple feature extractions are performed on the selected target training sample to obtain a sample seismic feature map. In one feature extraction process, the following operations are performed: depth separable convolution is performed on the input parameters of the current feature extraction to obtain a current channel feature map, and the current channel feature map is extracted through a large kernel attention mechanism to obtain a current seismic feature map, and the current seismic feature map is used as the input parameters of the next feature extraction. Based on the sample seismic feature map, the sample seismic inversion result of the target training sample is determined, and the parameter is adjusted based on the comprehensive loss value corresponding to the sample seismic inversion result, thereby improving the inversion accuracy and computational efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic inversion, and in particular to a seismic inversion model training method and device, equipment and a medium. BACKGROUND

[0002] Seismic inversion is an important means to obtain elastic parameters of underground medium and formation lithology. With the change of oil and gas exploration targets to complex lithology and unconventional reservoirs, the seismic inversion method faces challenges such as high computational complexity, strong nonlinearity, insufficient solution stability, and uncertainty quantification.

[0003] Seismic inversion methods based on deep learning have shown great potential in different fields such as seismic inversion and interpretation, but the existing seismic inversion methods based on deep learning have low inversion accuracy and computational efficiency, especially in the context of complex oil and gas reservoirs.

[0004] For example, traditional convolutional neural networks (CNNs) are limited by local receptive fields and are difficult to effectively capture macro geological structures such as faults; recurrent neural networks (RNNs) can model temporal dependencies, but have problems of gradient vanishing or explosion; traditional local attention mechanisms are limited by the receptive field range and are difficult to effectively model the long-range spatial correlation characteristics of underground medium. SUMMARY

[0005] The embodiments of the present application provide a seismic inversion model training method, device, equipment and medium to improve the inversion accuracy and computational efficiency of seismic inversion.

[0006] In a first aspect, the embodiments of the present application provide a seismic inversion model training method, which comprises:

[0007] Obtaining a training sample set, each training sample in the training sample set comprising multi-angle sample seismic data;

[0008] Iteratively training a seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, wherein in one iteration training process, the following operations are performed:

[0009] Performing multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, wherein in one feature extraction process, the following operations are performed: performing depth separable convolution on the input parameters of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as the input parameters of the next feature extraction;

[0010] Based on the sample seismic feature map, a sample seismic inversion result of the target training sample is determined, and parameter tuning is performed based on a comprehensive loss value corresponding to the sample seismic inversion result.

[0011] In an optional embodiment, based on the sample seismic feature map, the sample seismic inversion result of the target training sample is determined, including:

[0012] The sample seismic feature map is subjected to time sequence feature extraction of different scales to obtain a time sequence feature map.

[0013] Based on the time sequence feature map, the sample seismic inversion result of the target training sample is determined.

[0014] In an optional embodiment, based on the time sequence feature map, the sample seismic inversion result of the target training sample is determined, including:

[0015] The time sequence feature map is subjected to time sequence dependence modeling and time domain noise filtering to obtain a preliminary seismic inversion sequence.

[0016] The preliminary seismic inversion sequence is subjected to up-sampling, and the up-sampled data is mapped from a feature domain to a target domain through a regression layer to obtain the sample seismic inversion result.

[0017] In an optional embodiment, parameter tuning is performed based on a comprehensive loss value corresponding to the sample seismic inversion result, including:

[0018] Based on multiple loss functions, multiple sub-loss values of the sample seismic inversion result are calculated.

[0019] The multiple sub-loss values are subjected to weighted summation to obtain the comprehensive loss value, and parameter tuning is performed based on the comprehensive loss value.

[0020] In an optional embodiment, the multiple loss functions include an attribute loss function and other loss functions, and the other loss functions include a seismic loss function and / or a low-frequency loss function.

[0021] Then, based on the multiple loss functions, the multiple sub-loss values of the sample seismic inversion result are calculated, including:

[0022] Based on the attribute loss function, an error between a sample predicted value of a reservoir parameter included in the sample seismic inversion result and a sample true value of the reservoir parameter in the well logging data is calculated to obtain an attribute loss value.

[0023] The sample prediction value is input into a seismic forward model to obtain synthetic seismic data, and an error between the synthetic seismic data and sample seismic data of the target training sample is calculated based on a seismic loss function to obtain a seismic loss value; and / or, a low-frequency background model is generated based on the spatial distribution of the logging data, and the sample prediction value is low-frequency filtered to obtain a low-frequency sample prediction value, and an error between the low-frequency sample prediction value and a low-frequency sample true value in the low-frequency background model is calculated based on a low-frequency loss function to obtain a low-frequency loss value.

[0024] In an optional embodiment, the target seismic inversion model comprises a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module, and a regression layer module.

[0025] The local feature perception layer module comprises a plurality of depth separable large kernel attention modules, each of which is composed of a depth separable convolution and a large kernel attention mechanism.

[0026] The structural feature analysis layer module comprises a plurality of time sequence convolution networks.

[0027] The sequence modeling layer module comprises a plurality of bidirectional gated recurrent units.

[0028] In an optional embodiment, the method further comprises:

[0029] The multi-angle seismic data to be inverted is input into the target seismic inversion model as an input parameter to obtain a target prediction value of the corresponding reservoir parameter of the multi-angle seismic data to be inverted.

[0030] In a second aspect, the embodiments of the present application further provide a device for training a seismic inversion model, and the device comprises:

[0031] The acquisition module is configured to acquire a training sample set, each training sample in the training sample set comprising multi-angle sample seismic data.

[0032] The training module is configured to iteratively train a seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, wherein in one iteration training process, the following operations are performed:

[0033] The selected target training sample is subjected to multiple feature extractions to obtain a sample seismic feature map, wherein in one feature extraction process, the following operations are performed: the input parameter of the current feature extraction is subjected to depth separable convolution to obtain a current channel feature map, and the current channel feature map is subjected to feature extraction through a large kernel attention mechanism to obtain a current seismic feature map, and the current seismic feature map is taken as an input parameter for next feature extraction;

[0034] Based on the sample seismic feature map, a sample seismic inversion result of the target training sample is determined, and parameter adjustment is performed based on a comprehensive loss value corresponding to the sample seismic inversion result.

[0035] In an optional embodiment, when the sample seismic inversion result of the target training sample is determined based on the sample seismic feature map, the training module is further configured to:

[0036] perform time sequence feature extraction on the sample seismic feature map to obtain a time sequence feature map;

[0037] Based on the time sequence feature map, the sample seismic inversion result of the target training sample is determined.

[0038] In an optional embodiment, when the sample seismic inversion result of the target training sample is determined based on the time sequence feature map, the training module is further configured to:

[0039] perform time sequence dependence modeling and time domain noise filtering on the time sequence feature map to obtain a preliminary seismic inversion sequence;

[0040] perform upsampling on the preliminary seismic inversion sequence, and map the upsampled data from a feature domain to a target domain through a regression layer to obtain the sample seismic inversion result.

[0041] In an optional embodiment, when the parameter adjustment is performed based on the comprehensive loss value corresponding to the sample seismic inversion result, the training module is further configured to:

[0042] Based on a plurality of loss functions, a plurality of sub-loss values of the sample seismic inversion result are calculated;

[0043] The plurality of sub-loss values are weighted and summed to obtain the comprehensive loss value, and the parameter adjustment is performed based on the comprehensive loss value.

[0044] In an optional embodiment, the plurality of loss functions include an attribute loss function and other loss functions, and the other loss functions include a seismic loss function and / or a low-frequency loss function.

[0045] Then, based on the plurality of loss functions, the plurality of sub-loss values of the sample seismic inversion result are calculated, and the training module is further configured to:

[0046] Based on the attribute loss function, an error between a sample predicted value of a reservoir parameter included in the sample seismic inversion result and a sample true value of the reservoir parameter in the well logging data is calculated to obtain an attribute loss value.

[0047] The sample prediction value is input into a seismic forward model to obtain synthetic seismic data, and based on a seismic loss function, an error between the synthetic seismic data and sample seismic data of the target training sample is calculated to obtain a seismic loss value; and / or, a low-frequency background model is generated based on the spatial distribution of the logging data, and the sample prediction value is low-frequency filtered to obtain a low-frequency sample prediction value, and based on a low-frequency loss function, an error between the low-frequency sample prediction value and a low-frequency sample true value in the low-frequency background model is calculated to obtain a low-frequency loss value.

[0048] In an optional embodiment, the target seismic inversion model comprises: a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module, and a regression layer module.

[0049] The local feature perception layer module comprises: a plurality of depth separable large kernel attention modules, each of which is composed of a depth separable convolution and a large kernel attention mechanism.

[0050] The structural feature analysis layer module comprises: a plurality of time sequence convolution networks.

[0051] The sequence modeling layer module comprises: a plurality of bidirectional gated recurrent units.

[0052] In an optional embodiment, the device further comprises an inversion module configured to:

[0053] The multi-angle seismic data to be inverted is input into the target seismic inversion model as an input parameter to obtain a target prediction value of the corresponding reservoir parameter of the multi-angle seismic data to be inverted.

[0054] In a third aspect, the embodiments of the present application further provide an electronic device, comprising:

[0055] a processor; and

[0056] a memory storing a program,

[0057] wherein the program comprises instructions which, when executed by the processor, cause the processor to perform the training method of the seismic inversion model according to the first aspect.

[0058] In a fourth aspect, the embodiments of the present application further provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the training method of the seismic inversion model according to the first aspect.

[0059] In a fifth aspect, the present application provides a computer program product, which, when invoked by a computer, causes the computer to perform the training method steps of the seismic inversion model according to the first aspect.

[0060] The application has the following advantages:

[0061] In the method for training a seismic inversion model provided in the embodiments of the application, the seismic inversion model to be trained is iteratively trained based on a training sample set to obtain a target seismic inversion model. In one iteration training process, the following operations are performed: multiple feature extractions are performed on the selected target training sample to obtain a sample seismic feature map. In one feature extraction process, the following operations are performed: depth separable convolution is performed on input parameters of the current feature extraction to obtain a current channel feature map, and the current channel feature map is extracted through a large kernel attention mechanism to obtain a current seismic feature map, and the current seismic feature map is taken as input parameters of the next feature extraction. Based on the sample seismic feature map, a sample seismic inversion result of the target training sample is determined, and a comprehensive loss value corresponding to the sample seismic inversion result is adjusted. In this way, the depth separable convolution and the large kernel attention mechanism are combined to propose a depth separable large kernel attention mechanism. Global features are extracted through large kernel convolution, and key band information is adaptively enhanced in combination with a channel attention mechanism, so as to expand the effective receptive field range by expanding the convolution kernel dimension, realize long-range spatial correlation modeling of the lateral continuity features of seismic wave impedance, more efficiently extract multi-scale features, significantly reduce the computational complexity, and improve the calculation efficiency and inversion accuracy of the seismic inversion. Further, the comprehensive loss value corresponding to the sample seismic inversion result is adjusted, which can improve the accuracy of the target seismic inversion model. In addition, multiple feature extractions are performed on the selected target training sample to capture the dependency relationship in different spatial ranges, realize multi-scale feature coverage from micro to macro, and improve the adaptability to complex geological structures.

[0062] In addition, other features and advantages of the present application will be further described in the following description, and some will become apparent from the description, or will be learned through implementation of the present application. The purposes and other advantages of the present application can be achieved and obtained through the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings described here are used to provide further understanding of the application, and form a part of the application. They do not constitute an improper limitation on the application. In the drawings:

[0064] Figure 1 An optional system architecture schematic diagram applicable to the embodiments of the application;

[0065] Figure 2 An implementation flowchart of the method for training a seismic inversion model provided in the embodiments of the application;

[0066] Figure 3 The true EI and estimated EI comparison chart of Marmousi 2 model when the azimuth angle is 0°, wherein, Fig. (a) is the true EI, Fig. (b) is the estimated EI;

[0067] Figure 4 The true EI and estimated EI comparison chart of Marmousi 2 model when the azimuth angle is 10°, wherein, Fig. (a) is the true EI, Fig. (b) is the estimated EI;

[0068] Figure 5 The true EI and estimated EI comparison chart of Marmousi 2 model when the azimuth angle is 20°, wherein, Fig. (a) is the true EI, Fig. (b) is the estimated EI;

[0069] Figure 6 The true EI and estimated EI comparison chart of Marmousi 2 model when the azimuth angle is 30°, wherein, Fig. (a) is the true EI, Fig. (b) is the estimated EI;

[0070] Figure 7 The residual comparison chart of two methods when the azimuth angle is 0°, wherein, Fig. (a) is the residual comparison chart of the true EI and the estimated EI of the seismic inversion method proposed in the embodiment of the application, Fig. (b) is the residual comparison chart of the true EI and the estimated EI of the traditional method;

[0071] Figure 8 The residual comparison chart of two methods when the azimuth angle is 10°, wherein, Fig. (a) is the residual comparison chart of the true EI and the estimated EI of the seismic inversion method proposed in the embodiment of the application, Fig. (b) is the residual comparison chart of the true EI and the estimated EI of the traditional method;

[0072] Figure 9 The residual comparison chart of two methods when the azimuth angle is 20°, wherein, Fig. (a) is the residual comparison chart of the true EI and the estimated EI of the seismic inversion method proposed in the embodiment of the application, Fig. (b) is the residual comparison chart of the true EI and the estimated EI of the traditional method;

[0073] Figure 10 The residual comparison chart of two methods when the azimuth angle is 30°, wherein, Fig. (a) is the residual comparison chart of the true EI and the estimated EI of the seismic inversion method proposed in the embodiment of the application, Fig. (b) is the residual comparison chart of the true EI and the estimated EI of the traditional method;

[0074] Figure 11Figure 2 is a comparison chart of the estimated EIs of the two methods when the azimuth angle is 0°, wherein figure (a) is the estimated EIs obtained by the seismic inversion method proposed in the embodiments of the present application, and figure (b) is the estimated EIs obtained by the traditional method;

[0075] Figure 12 Figure 3 is a comparison chart of the estimated EIs of the two methods when the azimuth angle is 10°, wherein figure (a) is the estimated EIs obtained by the seismic inversion method proposed in the embodiments of the present application, and figure (b) is the estimated EIs obtained by the traditional method;

[0076] Figure 13 Figure 4 is a comparison chart of the estimated EIs of the two methods when the azimuth angle is 20°, wherein figure (a) is the estimated EIs obtained by the seismic inversion method proposed in the embodiments of the present application, and figure (b) is the estimated EIs obtained by the traditional method;

[0077] Figure 14 Figure 5 is a comparison chart of the estimated EIs of the two methods when the azimuth angle is 30°, wherein figure (a) is the estimated EIs obtained by the seismic inversion method proposed in the embodiments of the present application, and figure (b) is the estimated EIs obtained by the traditional method;

[0078] Figure 15 Figure 6 is a schematic diagram of the actual seismic data processing test, wherein figure (a) is a true post-stack seismic profile, figure (b) is a seismic low-frequency background, figure (c) is the actual data inversion result obtained by using the traditional semi-supervised sequence modeling seismic inversion method, and figure (d) is the actual data inversion result obtained by using the seismic inversion method proposed in the embodiments of the present application;

[0079] Figure 16 Figure 7 is a structural schematic diagram of a training device of a seismic inversion model provided by the embodiments of the present application;

[0080] Figure 17 Figure 8 is a structural schematic diagram of an electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0081] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided to make the present application more thorough and complete. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes, and are not intended to limit the scope of protection of the present application.

[0082] It should be understood that each of the steps recited in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present application is not limited in this regard.

[0083] The term "comprises" and variations thereof herein are open-ended, that is, "comprising but not limited to." The term "based on" is "based, at least in part, on." The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments." Related terms shall be construed accordingly. It should be noted that the use of "first", "second", or "third" etc. in the present application is merely to distinguish one concept from another, and does not imply the order or interdependence of the functions performed by these devices, modules or units.

[0084] It should be noted that the modification of "one" or "multiple" mentioned in the present application is illustrative but not limiting, and those skilled in the art should understand that unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0085] The names of the messages or information exchanged between the devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0086] The design idea of the embodiments of the present application is briefly introduced as follows:

[0087] Seismic inversion is an important means to obtain elastic parameters of underground medium and formation lithology. With the change of oil and gas exploration targets to complex lithology and unconventional reservoirs, the seismic inversion method faces the challenges of high computational complexity, strong nonlinearity, insufficient solution stability and uncertainty quantification. In recent years, intelligent seismic inversion and probabilistic seismic inversion method based on Bayesian inference help to mine underground oil and gas reservoir information from multiple types of data such as seismic, logging and geology, and gradually become a research hotspot and frontier in the field of oil and gas seismic exploration.

[0088] Seismic inversion methods based on deep learning show great potential in different fields such as seismic inversion and interpretation, but traditional CNN is limited by local receptive field and is difficult to effectively capture macro geological structures such as faults; RNN can model temporal dependence, but has the problem of gradient vanishing / explosion; and traditional local attention mechanism is limited by the range of receptive field and is difficult to effectively model the long-range spatial correlation characteristics of underground medium.

[0089] Therefore, it is urgent to innovate the deep learning network architecture to improve the inversion accuracy and computational efficiency of seismic inversion, and to provide a high-precision and high-efficiency solution for complex reservoir seismic inversion.

[0090] In an optional implementation, the embodiment of the present application provides a training method of a seismic inversion model, which can specifically include: obtaining a training sample set, each training sample in the training sample set including: multi-angle sample seismic data; iteratively training a seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, wherein in one iteration training process, the following operations are performed: performing multiple feature extractions on a selected target training sample to obtain a sample seismic feature map, wherein in one feature extraction process, the following operations are performed: performing depth separable convolution on input parameters of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as input parameters of the next feature extraction; determining a sample seismic inversion result of the target training sample based on the sample seismic feature map, and adjusting parameters based on a comprehensive loss value corresponding to the sample seismic inversion result.

[0091] In the above manner, first, depth separable convolution is used to decompose standard convolution into a cascade structure of depth convolution and point convolution, which not only significantly reduces the computational complexity, but also significantly reduces the parameter amount, and effectively captures the spatial correlation within the channel through depth convolution and enhances the feature expression capability through point convolution. At the same time, the depth separable convolution is combined with the large kernel attention mechanism to propose a depth separable large kernel attention mechanism, which extracts global features through large kernel convolution and adaptively enhances key band information through the channel attention mechanism, thereby expanding the effective receptive field range through the dimension of the convolution kernel, modeling the long-range spatial correlation of the lateral continuity feature of the seismic wave impedance, more efficiently extracting multi-scale features, fully considering the problem that the traditional local attention mechanism is limited by the receptive field range and is difficult to effectively model the long-range spatial correlation characteristics of the underground medium, improving the calculation efficiency and inversion accuracy of the seismic inversion, and the depth separable large kernel attention mechanism uses modularization, the depth convolution, point convolution and channel attention mechanism can be independently adjusted, has high flexibility, is suitable for complex tasks and multi-scale feature extraction, and the depth convolution is independently calculated on each channel to reduce noise interference, and the global average pooling is used to suppress noise and enhance the sensitivity to important frequency bands. Further, adjusting the parameters based on the comprehensive loss value corresponding to the sample seismic inversion result can improve the accuracy of the target seismic inversion model. In addition, multiple feature extractions are performed on the selected target training sample to capture the dependency relationship in different spatial ranges, realize multi-scale feature coverage from micro to macro, and improve the adaptability to complex geological structures.

[0092] In particular, the preferred embodiments of the present application are described below in conjunction with the accompanying drawings, it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0093] Referring to Figure 1 As shown in the figure, it is an optional system architecture diagram applicable to the embodiments of the present application, which can include acquisition device 101 and server 102. The acquisition device 101 and the server 102 can interact information through a communication network, wherein the communication network can adopt a communication mode including wireless communication mode and wired communication mode. For example, the acquisition device 101 can access the network through cellular mobile communication technology and communicate with the server 102. Wherein, the cellular mobile communication technology, such as, including the fifth generation mobile communication (5th generation mobile networks, 5G) technology or the next generation mobile communication technology. Optionally, the terminal device (101a, 101b) can access the network through short distance wireless communication mode and communicate with the server 102. Wherein, the short distance wireless communication mode, such as, including wireless fidelity (wireless fidelity, Wi-Fi) technology.

[0094] The embodiments of the present application do not make any limitation on the number of devices involved in the above system architecture, for example, the above system architecture can include more acquisition devices, or can also include other devices. For example, Figure 1 As shown in the figure, only the acquisition device 101 and the server 102 are described as an example, the following briefly introduces each communication device and its respective function.

[0095] The acquisition device 101 is a device that can collect seismic data. For example, the acquisition device 101 can include but not limited to: geophone, split type seismic data collector and microseismic data collector, etc.

[0096] The server 102 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (content delivery network, CDN), and basic cloud computing services such as big data and artificial intelligence platform.

[0097] It is worth mentioning that the server 102 in the embodiment of the present application can perform iterative training on the seismic inversion model to be trained based on the obtained training sample set to obtain a target seismic inversion model. In one iteration training process, the following operations are performed: performing multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, determining a sample seismic inversion result of the target training sample based on the sample seismic feature map, and adjusting parameters based on a comprehensive loss value corresponding to the sample seismic inversion result.

[0098] Optionally, a pre-trained target seismic inversion model can be deployed on the server 102. In this way, after obtaining the multi-angle seismic data to be inverted, the server 102 can input the multi-angle seismic data to be inverted into the target seismic inversion model to determine a target seismic inversion result of the multi-angle seismic data to be inverted.

[0099] Illustratively, the output target seismic inversion result can be a target predicted value of a reservoir parameter corresponding to the multi-angle seismic data to be inverted, and the reservoir parameter is elastic wave impedance, that is, the output target seismic inversion result can be a target predicted value of elastic wave impedance corresponding to the multi-angle seismic data to be inverted.

[0100] The training method of the seismic inversion model provided by the exemplary embodiment of the present application will be described below in combination with the above-mentioned system architecture and with reference to the accompanying drawings. It should be noted that the above-mentioned system architecture is only shown for the purpose of facilitating understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect.

[0101] Referring to Figure 2 FIG. 2 shows an implementation flowchart of a training method of a seismic inversion model provided by an embodiment of the present application, and the execution subject is taken as an example of a server. The specific implementation process of the method is as follows:

[0102] S20: Obtain a training sample set.

[0103] Each training sample in the training sample set includes multi-angle sample seismic data, which includes but is not limited to sample seismic data with incident angles of 0°, 10°, 20° and 30°, and each training sample in the training sample set further includes a sample true value of a reservoir parameter in well logging data.

[0104] In the embodiment of the present application, multi-angle sample seismic data is collected by a collection device to construct a training sample set that can be used for training of a target seismic inversion model.

[0105] Optionally, in the embodiment of the present application, the collected seismic data is subjected to data preprocessing, which includes any one or any combination of the following processing: collation, singular value processing and missing value processing.

[0106] S21: iteratively train the seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, wherein in one iteration training process: includes S210-S211.

[0107] In the embodiments of the present application, the target seismic inversion model includes a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module, and a regression layer module, which will be introduced below:

[0108] (1) The local feature perception layer module includes a plurality of depthwise separable large kernel attention (DS-LKA) modules, which are composed of depthwise separable convolution and large kernel attention mechanism, and the depthwise separable convolution is composed of depthwise convolution and point convolution.

[0109] Specifically, the local feature perception layer module is composed of three DS-LKA modules connected in series.

[0110] In the embodiments of the present application, the specific workflow of the DS-LKA module is as follows: first, the DS-LKA module decomposes the standard convolution into two steps of depthwise convolution and point convolution through depthwise separable convolution. Then the large kernel attention mechanism is introduced to expand the convolution kernel dimension to improve the effective receptive field range, realize the long-range spatial correlation modeling of the lateral continuity features of the seismic wave impedance, and the core idea is to extract global features through large kernel convolution and combine channel attention mechanism to adaptively enhance key band information.

[0111] Wherein, the depthwise convolution is independently convolved on each input channel, and the parameter quantity is CxK (wherein, C is the channel number, and K is the kernel size), and the mathematical expression is:

[0112]

[0113] Wherein, W dw is the weight matrix of the depthwise convolution, and x is the input parameter.

[0114] In addition, it is worth noting that in the embodiments of the present application, the channel number and the kernel size can be optimized based on the characteristics of the seismic data (such as signal-to-noise ratio, resolution) and hardware constraints, for example, the kernel size is larger to cover long-range space.

[0115] Wherein, the point convolution uses 1x1 convolution for channel mixing, and the parameter quantity is CxC, and the mathematical expression is:

[0116]

[0117] wherein W pw is a weight matrix of the point convolution.

[0118] In addition, it is worth noting that the total parameter amount of the depth separable convolution is C x K + C x C, which is much lower than C x C x K of the standard convolution.

[0119] wherein the mathematical expression of the large kernel attention mechanism is:

[0120]

[0121] wherein GAP represents the global average pooling, FC1 and FC2 are the fully connected layers, sigma is the Sigmoid function, and finally the residual connection is performed to retain the original input features and relieve the gradient vanishing problem.

[0122] In summary, the mathematical expression of the DS-LKA module is as follows:

[0123]

[0124] wherein represents the channel-wise multiplication.

[0125] In this way, compared with the large kernel attention (LKA), the DS-LKA has the following advantages: (1) the depth separable convolution is used to decompose the standard convolution into a cascade structure of depth convolution and point convolution, which significantly reduces the computational complexity and makes the parameter amount of the DS-LKA significantly lower than that of the LKA; (2) the depth convolution adopts the channel-independent operation mode to effectively capture the spatial correlation within the channel, and the point convolution performs channel mixing through the 1x1 convolution to enhance the feature expression capability; finally, the depth separable convolution structure and the dynamic channel attention mechanism are integrated to more efficiently extract multi-scale features; (3) the DS-LKA uses modular design, and the depth convolution, point convolution and channel attention mechanism can be independently adjusted, which has high flexibility and is suitable for complex tasks and multi-scale feature extraction; (4) the depth convolution is independently calculated on each channel to reduce noise interference, and the global average pooling is used to suppress noise and enhance the sensitivity to important frequency bands, so that the anti-noise performance of the DS-LKA is better than that of the LKA, especially in low signal-to-noise ratio data.

[0126] (2) The structural feature analysis layer module comprises a plurality of temporal convolutional networks (TCN).

[0127] Specifically, three different TCNs are connected in parallel, and the temporal causal constraint and residual learning mechanism are introduced. By using different dilation factors (such as d = 1, 3, 6), the model can extract time series features of different scales, thereby improving the modeling ability of the model for long-range dependence of seismic signals.

[0128] In addition, it is worth noting that TCN itself is also a kind of convolutional network, but it is specially designed for time series, emphasizing causality and long-range dependence. The core components include causal convolution, dilated convolution and residual connection. The causal convolution of TCN naturally satisfies the time series constraint and is more suitable for seismic signal processing than traditional CNN. Causal convolution ensures the causality in time and does not leak future information, which is very important for seismic data processing. The dilated convolution expands the receptive field and captures the long-range dependence, while the residual connection helps the deep network training more stable. TCN may be more efficient in processing long sequences due to the use of dilated convolution and residual connection, especially in parallel implementation.

[0129] (3) The sequence modeling layer module comprises a plurality of bidirectional gated recurrent units (B-GRU).

[0130] Specifically, the sequence modeling layer module is composed of three B-GRUs connected in series, and a B-GRU cascade network is constructed. The gating mechanism is used to filter noise interference, and a cross-time step association is established between the hidden layers by referring to the multi-scale feature self-attention mechanism, which is used to extract complex time series features and generate preliminary seismic inversion sequences. The three-layer B-GRU is equivalent to a three-layer Gated Recurrent Unit (GRU) network, which can model the complex nonlinear relationships that cannot be extracted by shallow networks.

[0131] (4) The up-sampling layer module comprises a plurality of deconvolution blocks.

[0132] Specifically, the up-sampling layer module is composed of two deconvolution blocks with different kernel strides. The purpose is to up-sample the seismic data so that the resolution of the seismic data matches the resolution of the logging data.

[0133] (5) The regression layer module comprises a GRU and a fully connected layer.

[0134] Specifically, the regression layer module is composed of a GRU and a fully connected layer, which can complete the mapping from the feature space to the target parameter. The GRU in the module is a simple single-layer B-GRU that integrates the idea of pre-stack inversion to establish a nonlinear mapping between angle gathers and elastic parameters, thereby enhancing the global temporal features of extracted information. The fully connected layer introduces F-norm constraint to ensure that parameters such as density and velocity conform to rock physics laws, and can map the output characteristics of the GRU to the same number of features in the target domain.

[0135] In the embodiments of the present application, in order to obtain the target seismic inversion model, the seismic inversion model to be trained is iteratively trained based on the training sample set until convergence.

[0136] The convergence condition includes but is not limited to that the loss value and the accuracy rate are stable or change very little during the training process.

[0137] Further, the detailed process of each iteration training can be referred to S210-S211, and is specifically as follows:

[0138] S210: performing multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, wherein in the process of each feature extraction, the following operations are performed: performing depth separable convolution on the input parameters of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as the input parameters of the next feature extraction.

[0139] In the embodiments of the present application, the local feature perception layer module is used to perform multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, and each DS-LKA module in the DS-LKA module is used to perform one feature extraction, wherein in the process of each feature extraction, the following operations are performed: performing convolution operation on the input parameters of the current feature extraction independently on each input channel to obtain deep convolution features, then performing channel mixing using 1x1 convolution to obtain a current channel feature map, then performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain an attention feature map, weighting the channels of the attention feature map to the current channel feature map through channel-by-channel multiplication, and finally performing residual connection between the current channel feature map and the original input features (i.e., the input parameters) to obtain a current seismic feature map, and taking the current seismic feature map as the input parameters of the next feature extraction.

[0140] The mathematical expression of the DS-LKA module is as follows:

[0141]

[0142] In this way, since the DS-LKA is usually used to capture long-range dependencies, it is helpful for continuous stratigraphic structures in seismic data, and by increasing the size of the convolution kernel to cover a larger spatial range, a longer range can be covered in the time dimension, the perception of the model to the context is enhanced, and geological structure features such as faults in the seismic profile can be effectively captured. Moreover, multiple DS-LKA modules are stacked, and the output of each module is used as the input of the next module, so that the features can be refined one by one, the feature expression ability is gradually enhanced, and the accuracy of the model is further improved, thereby improving the accuracy of the inversion.

[0143] In addition, it is worth noting that the feature map in the embodiments of the present application is a feature tensor.

[0144] S211: Based on the sample seismic feature map, a sample seismic inversion result of the target training sample is determined, and the sample seismic inversion result is adjusted based on the corresponding comprehensive loss value.

[0145] Optionally, in the embodiments of the present application, a possible embodiment is provided for determining the sample seismic inversion result of the target training sample based on the sample seismic feature map, and the following operations are performed:

[0146] S2110: Time sequence feature extraction of different scales is performed on the sample seismic feature map to obtain a time sequence feature map.

[0147] In the embodiments of the present application, a structural feature analysis layer module is used to perform time sequence feature extraction of different scales on the sample seismic feature map to obtain a time sequence feature map.

[0148] Specifically, three parallel TCNs are used to extract time sequence features of different scales, thereby improving the modeling capability of the model for long-range dependencies of seismic signals.

[0149] In this way, the combination of the depth separable large kernel attention mechanism and the time domain convolution network enables the network to have strong spatial and temporal long-range dependency modeling capabilities, and the network can simultaneously process long-range relationships of seismic data in space (such as between channels) and time (such as time series), thereby more comprehensively describing the seismic data and improving the inversion accuracy.

[0150] S2111: Based on the time sequence feature map, a sample seismic inversion result of the target training sample is determined.

[0151] Optionally, in the embodiments of the present application, a possible embodiment is provided for determining the sample seismic inversion result of the target training sample based on the time sequence feature map, and the following operations are performed:

[0152] S2111-1: Time sequence dependency modeling and time domain noise filtering are performed on the time sequence feature map to obtain a preliminary seismic inversion sequence.

[0153] In the embodiment of the application, the sequence modeling layer module is adopted to perform time sequence dependent modeling and time domain noise filtering on the time sequence feature map to obtain a preliminary seismic inversion sequence.

[0154] Specifically, three series of B-GRUs are adopted to perform time sequence dependent modeling and time domain noise filtering on the time sequence feature map to extract complex time sequence features and obtain a preliminary seismic inversion sequence.

[0155] S2111-2: The preliminary seismic inversion sequence is up-sampled, and the up-sampled data is mapped from the feature domain to the target domain through the regression layer to obtain a sample seismic inversion result.

[0156] In the embodiment of the application, the up-sampling layer module is adopted to up-sample the preliminary seismic inversion sequence, and the regression layer module is adopted to map the up-sampled data from the feature domain to the target domain to obtain a sample seismic inversion result.

[0157] Specifically, two deconvolution blocks with different kernel steps are adopted to up-sample the preliminary seismic inversion sequence to match the resolution of the seismic data with the resolution of the logging data, and then a GRU and a fully connected layer are adopted to map the up-sampled data from the feature domain to the target domain to obtain a sample seismic inversion result.

[0158] Optionally, in the embodiment of the application, a possible embodiment is provided for parameter adjustment based on a comprehensive loss value corresponding to the sample seismic inversion result, and the following operation is performed:

[0159] S2112: A plurality of sub-loss values of the sample seismic inversion result are calculated based on a plurality of loss functions.

[0160] In the embodiment of the application, the plurality of loss functions include an attribute loss function and other loss functions, and the other loss functions include a seismic loss function and / or a low-frequency loss function.

[0161] For example, the plurality of loss functions include the attribute loss function and the seismic loss function, for another example, the plurality of loss functions include the attribute loss function and the low-frequency loss function, and for another example, the plurality of loss functions include the attribute loss function, the seismic loss function and the low-frequency loss function.

[0162] Optionally, in the embodiment of the application, one loss function is adopted to calculate a sub-loss value corresponding to the loss function, including the following three ways:

[0163] Way 1: Based on the attribute loss function, an error between a sample predicted value of a reservoir parameter included in the sample seismic inversion result and a sample true value of the reservoir parameter in the logging data is calculated to obtain an attribute loss value.

[0164] In the embodiments of the present application, the mathematical expression of the attribute loss function is:

[0165]

[0166] wherein, L property (Θ) represents the attribute loss function, S represents a set of available attribute traces in the well logging data, m i represents the sample true value of the reservoir parameter in the well logging data, is the ith trace in S, D is a distance metric, d i represents the sample seismic data of the target training sample, represents the sample seismic response result of the sample seismic data obtained through the target seismic inversion model, that is, the sample predicted value of the reservoir parameter contained in the sample seismic inversion result.

[0167] Method 2: inputting the sample predicted value into the seismic forward model to obtain the synthetic seismic data, and calculating the error between the synthetic seismic data and the sample seismic data of the target training sample based on the seismic loss function to obtain the seismic loss value.

[0168] In the embodiments of the present application, the mathematical expression of the seismic loss function is:

[0169]

[0170] wherein, L seismic (Θ) represents the seismic loss function, d i represents the sample seismic data of the target training sample, represents the synthetic seismic data of the sample predicted value of the reservoir parameter contained in the sample seismic inversion result through the seismic forward model, and D is a distance metric, wherein the seismic forward model is a seismic forward model without learnable parameters.

[0171] Method 3: generating a low-frequency background model based on the spatial distribution of the well logging data, performing low-frequency filtering on the sample predicted value to obtain a low-frequency sample predicted value, and calculating the error between the low-frequency sample predicted value and the low-frequency sample true value in the low-frequency background model based on the low-frequency loss function to obtain a low-frequency loss value.

[0172] In the embodiments of the present application, the mathematical expression of the low-frequency loss function is:

[0173]

[0174] wherein, L low (Θ) represents the low-frequency model loss function, m low represents the low-frequency model of the reservoir parameter to be inverted, that is, the low-frequency sample true value in the low-frequency background model, represents a low-frequency filtering operation.

[0175] In this way, a low-frequency background model is generated, the trend accuracy of the inversion result is constrained, and the low-frequency model constraint can be closer to the parameter distribution of the underground medium, thereby increasing the geological and physical interpretability of the model.

[0176] S2113: Weighted sum of the plurality of sub-loss values to obtain a comprehensive loss value, and parameter tuning based on the comprehensive loss value.

[0177] In the embodiments of the present application, the plurality of sub-loss values are weighted and summed according to the weight coefficients corresponding to the plurality of loss functions to obtain a comprehensive loss value, and parameter tuning is performed based on the comprehensive loss value.

[0178] For example, when the plurality of loss functions include an attribute loss function, a seismic loss function and a low-frequency loss function, the mathematical expression of the comprehensive loss value is as follows:

[0179] L(Θ) = a1L property (Θ) + a2L seismic (Θ) + a3L low (Θ)

[0180] Wherein, a1 is the weight coefficient of the attribute loss function, L property (Θ) is the sub-loss value corresponding to the attribute loss function, a2 is the weight coefficient of the seismic loss function, L property (Θ) is the sub-loss value corresponding to the seismic loss function, and a3 is the weight coefficient of the low-frequency loss function. L low (Θ) is the sub-loss value corresponding to the low-frequency loss function.

[0181] In this way, the parameters of the seismic inversion model are adjusted by combining the three kinds of losses optimized by gradient descent in the above three formulas, and the process is repeated until convergence. The seismic loss ensures that the synthetic seismic data matches the input sample seismic data, and the waveform similarity is constrained. The attribute loss directly provides strong supervision between the sample predicted value and the sample true value of the reservoir parameter at the known well point, and the low-frequency model loss introduces geological and physical constraints, so that the result is more consistent with the actual distribution of the underground medium. By jointly optimizing seismic data matching, attribute matching and low-frequency trend constraint, end-to-end optimization is realized, so that the inversion result meets multiple conditions at the same time, thereby improving the accuracy. In addition, the balance of multiple loss terms can avoid the optimization process from falling into local minimum and accelerate the convergence.

[0182] Further, after obtaining the trained target seismic inversion model, the target seismic inversion model is used for seismic inversion. Specifically, the multi-angle seismic data to be inverted is input into the target seismic inversion model as an input parameter to obtain the target predicted value of the reservoir parameter corresponding to the multi-angle seismic data to be inverted.

[0183] The reservoir parameter is elastic impedance (EI).

[0184] In this way, the calculation efficiency and inversion accuracy of the seismic inversion can be improved by performing the seismic inversion based on the target seismic inversion model, and the seismic inversion method is suitable for a complex reservoir seismic inversion scene.

[0185] In the embodiments of the present application, in order to further verify the feasibility of the seismic inversion method (i.e., the time-domain seismic inversion method driven by the depth separable large kernel attention mechanism) proposed in the embodiments of the present application, the feasibility verification includes two steps:

[0186] First step: numerical experiment test the feasibility of the seismic inversion method, select Marmousi 2 model as a verification case, generate four incident angles of 0°, 10°, 20° and 30° elastic wave impedance data based on the elastic model of Marmousi 2. The full-area seismic trace data of Marmousi 2 model and part of the well-calibrated elastic wave impedance (EI) channel are jointly used as samples to construct a training sample set. The seismic inversion method proposed in the embodiments of the present application is used for inversion (i.e., the address inversion is performed using the target seismic inversion model described above), and the effectiveness of the seismic inversion method is tested by comparing the inversion results of the original inversion method and the seismic inversion method proposed in the embodiments of the present application.

[0187] Second step: test the effectiveness of the seismic inversion method by actual seismic data processing, select a two-dimensional seismic profile with complex geological structure, and use the seismic inversion method proposed in the embodiments of the present application to perform inversion. The stability and effectiveness of the inversion method are tested by comparing the inversion results obtained by the original inversion method.

[0188] To further illustrate the feasibility and effectiveness of the present application, two embodiments are listed as follows:

[0189] Embodiment 1: theoretical model test, see Figures 3-14 .

[0190] Figures 3-6 are the comparison diagrams of the real EI and the estimated EI of the Marmousi 2 model when the azimuth angle is 0°, 10°, 20° and 30°, respectively, wherein Fig. (a) is the real EI and Fig. (b) is the estimated EI. From the comparison diagram, it can be seen that the inversion result of the seismic inversion method proposed in the embodiments of the present application has high consistency with the actual model, which verifies the feasibility and stability of the inversion method.

[0191] Figures 7-10respectively, are residual comparison diagrams of two methods when the azimuth angle is 0°, 10°, 20° and 30° respectively, wherein Fig. (a) is a residual comparison diagram of the true EI and the estimated EI of the seismic inversion method proposed in the embodiment of the present application, and Fig. (b) is a residual comparison diagram of the true EI and the estimated EI of the traditional method. As can be seen from the comparison of the white rectangular boxes in the figures, the accuracy and lateral continuity of the inversion results of the seismic inversion method proposed in the embodiment of the present application are improved, proving the practicability and superiority of the seismic inversion method proposed in the embodiment of the present application.

[0192] Figures 11-14 respectively, are estimated EI comparison diagrams of two methods when the azimuth angle is 0°, 10°, 20° and 30° respectively, wherein Fig. (a) is the estimated EI obtained by the seismic inversion method proposed in the embodiment of the present application, and Fig. (b) is the estimated EI obtained by the traditional method. As can be seen from the comparison of the black boxes marked on the figures, the lateral continuity of the inversion results of the target seismic inversion model of the present application is obviously improved compared with the conventional semi-supervised sequential inversion network.

[0193] Embodiment 2: Actual data processing, see Figure 15 .

[0194] The seismic inversion method proposed in the embodiment of the present application is tested by inversion on actual data. The actual data used is the actual seismic profile collected in M working area of a certain place, Figure 15 , wherein (a) is the true post-stack seismic profile.

[0195] Figure 15 , wherein (b) is the seismic low-frequency background constructed according to the spatial distribution of the actual logging data by well interpolation technology, Figure 15 , wherein (c) and Figure 15 , wherein (d) are the actual data inversion results obtained by using the traditional semi-supervised sequential modeling seismic inversion method and the seismic inversion method proposed in the embodiment of the present application respectively. As can be seen from the comparison of the inversion results, the seismic inversion method proposed in the embodiment of the present application has a higher degree of agreement with the actual seismic data, and the resolution of the inversion results obtained by the seismic inversion method proposed in the embodiment of the present application is higher and the lateral continuity is better compared with the inversion results obtained by the traditional method. Therefore, the seismic inversion method proposed in the embodiment of the present application has strong advantages in intelligent seismic inversion, proving the practicability and effectiveness of the seismic inversion method proposed in the embodiment of the present application.

[0196] Further, based on the same technical concept, the embodiment of the present application provides a seismic inversion model training device for implementing the above method process of the embodiment of the present application. For example, as shown in Figure 16 , the seismic inversion model training device 1600 can include an acquisition module 1601, a training module 1602 and an inversion module 1603, wherein:

[0197] The acquisition module 1601 is configured to acquire a training sample set, each training sample in the training sample set comprising multi-angle sample seismic data.

[0198] The training module 1602 is configured to iteratively train a to-be-trained seismic inversion model based on the training sample set to obtain a target seismic inversion model, wherein in one iteration training process, the following operations are performed:

[0199] performing multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, wherein in one feature extraction process, the following operations are performed: performing depth separable convolution on input parameters of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as input parameters of the next feature extraction;

[0200] determining a sample seismic inversion result of the target training sample based on the sample seismic feature map, and adjusting parameters based on a comprehensive loss value corresponding to the sample seismic inversion result.

[0201] In an optional embodiment, when the sample seismic feature map is used to determine the sample seismic inversion result of the target training sample, the training module 1602 is further configured to:

[0202] performing time sequence feature extraction of different scales on the sample seismic feature map to obtain a time sequence feature map;

[0203] determining the sample seismic inversion result of the target training sample based on the time sequence feature map.

[0204] In an optional embodiment, when the time sequence feature map is used to determine the sample seismic inversion result of the target training sample, the training module 1602 is further configured to:

[0205] performing time sequence dependence modeling and time domain noise filtering on the time sequence feature map to obtain a preliminary seismic inversion sequence;

[0206] performing up-sampling on the preliminary seismic inversion sequence, and mapping the up-sampled data from a feature domain to a target domain through a regression layer to obtain the sample seismic inversion result.

[0207] In an optional embodiment, when the comprehensive loss value corresponding to the sample seismic inversion result is used to adjust parameters, the training module 1602 is further configured to:

[0208] calculating a plurality of sub-loss values of the sample seismic inversion result based on a plurality of loss functions;

[0209] performing weighted summation on the plurality of sub-loss values to obtain the comprehensive loss value, and adjusting parameters based on the comprehensive loss value.

[0210] In an optional embodiment, the plurality of loss functions comprises an attribute loss function and other loss functions, the other loss functions comprising a seismic loss function and / or a low-frequency loss function;

[0211] Then, based on the plurality of loss functions, the training module 1602 is further configured to calculate a plurality of sub-loss values of the sample seismic inversion result.

[0212] Based on the attribute loss function, an error between a sample predicted value of a reservoir parameter contained in the sample seismic inversion result and a sample true value of the reservoir parameter in the well logging data is calculated to obtain an attribute loss value.

[0213] The sample predicted value is input into a seismic forward model to obtain synthetic seismic data, and based on the seismic loss function, an error between the synthetic seismic data and the sample seismic data of the target training sample is calculated to obtain a seismic loss value; and / or, a low-frequency background model is generated based on the spatial distribution of the well logging data, and the sample predicted value is low-frequency filtered to obtain a low-frequency sample predicted value, and based on the low-frequency loss function, an error between the low-frequency sample predicted value and a low-frequency sample true value in the low-frequency background model is calculated to obtain a low-frequency loss value.

[0214] In an optional embodiment, the target seismic inversion model comprises a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module, and a regression layer module.

[0215] The local feature perception layer module comprises a plurality of depth separable large kernel attention modules, each depth separable large kernel attention module being composed of a depth separable convolution and a large kernel attention mechanism.

[0216] The structural feature analysis layer module comprises a plurality of time series convolution networks.

[0217] The sequence modeling layer module comprises a plurality of bidirectional gated recurrent units.

[0218] In an optional embodiment, the device further comprises an inversion module 1603, which is configured to:

[0219] Input the multi-angle seismic data to be inverted as an input parameter into the target seismic inversion model to obtain a target predicted value of the corresponding reservoir parameter of the multi-angle seismic data to be inverted.

[0220] Based on the description of the method embodiments and the device embodiments, the exemplary embodiments of the present application further provide an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication. The memory stores a computer program capable of being executed by the at least one processor, and the computer program, when executed by the at least one processor, is configured to cause the electronic device to perform the method according to the embodiments of the present application.

[0221] The embodiments of the present application further provide a non-transitory computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is configured to cause the computer to perform the method according to the embodiments of the present application.

[0222] The embodiments of the present application further provide a computer program product comprising a computer program, wherein the computer program, when executed by a processor of a computer, is configured to cause the computer to perform the method according to the embodiments of the present application.

[0223] Referring to Figure 17 A block diagram of the structure of an electronic device 1700 that can be a server or a client of the present application will now be described, which is an example of a hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent a variety of forms of digital electronic computing devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent a variety of forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application described and / or claimed in this document.

[0224] As Figure 17 shown, the electronic device 1700 includes a computing unit 1701 that can perform various appropriate actions and processes in accordance with a computer program stored in a read only memory (ROM) 1702 or a computer program loaded from a storage unit 1708 into a random access memory (RAM) 1703. Various programs and data required for the operation of the device 1700 can also be stored in the RAM 1703. The computing unit 1701, the ROM 1702, and the RAM 1703 are connected to each other through a bus 1704. An input / output (I / O) interface 1705 is also connected to the bus 1704.

[0225] The plurality of components in the electronic device 1700 are connected to the I / O interface 1705, including an input unit 1706, an output unit 1707, a storage unit 1708, and a communication unit 1709. The input unit 1706 can be any type of device that can input information to the electronic device 1700, and can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the electronic device. The output unit 1707 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1708 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 1709 allows the electronic device 1700 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth device, a WiFi device, a worldwide interoperability for microwave access (WiMax) device, a cellular communication device, and / or the like.

[0226] The computing unit 1701 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, and the like. The computing unit 1701 performs various methods and processes described above. For example, in some embodiments, the training method of the seismic inversion model described above can be implemented as a computer software program, which is tangibly embodied in a machine-readable medium, such as the storage unit 1708. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 1700 via the ROM 1702 and / or the communication unit 1709. In some embodiments, the computing unit 1701 can be configured to perform the training method of the seismic inversion model described above by any other appropriate means, such as by means of firmware.

[0227] Program code for carrying out methods of the present application can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, causes the functions / acts specified in the flowcharts and / or block diagrams to be implemented. The program code can be entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine or entirely on a remote machine or server.

[0228] In the context of the present application, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium would include one or more lines of electrical conductors, a portable computer disk, a hard disk, RAM, ROM, erasable programmable read-only memory (EPROM) or flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0229] As used in the present application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0230] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0231] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0232] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0233] And, it should be understood that all the disclosed herein are only preferred embodiments of the application, and cannot be used to limit the scope of the application, therefore, the equivalent changes made according to the claims of the application, still belong to the scope of the application.

Claims

1. A method of training a seismic inversion model, characterized in that, The method comprises the following steps: obtaining a training sample set, each training sample in the training sample set comprising multi-angle sample seismic data; iteratively training a seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, the target seismic inversion model comprising a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module and a regression layer module; the local feature perception layer module comprises a plurality of depth separable large kernel attention modules, each depth separable large kernel attention module being composed of a depth separable convolution and a large kernel attention mechanism; the structural feature analysis layer module comprises a plurality of time convolution networks; the sequence modeling layer module comprises a plurality of bidirectional gated recurrent units; wherein, in one iteration training process, the following operations are performed: performing multiple feature extractions on a selected target training sample to obtain a sample seismic feature map, wherein, in one feature extraction process, the following operations are performed: performing depth separable convolution on input parameters of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as input parameters of the next feature extraction; performing time sequence feature extraction of different scales on the sample seismic feature map to obtain a time sequence feature map; performing time sequence dependent modeling and time domain noise filtering on the time sequence feature map to obtain a preliminary seismic inversion sequence; performing up-sampling on the preliminary seismic inversion sequence, and mapping the up-sampled data from a feature domain to a target domain through a regression layer to obtain a sample seismic inversion result of the target training sample; based on the comprehensive loss value corresponding to the sample seismic inversion result, the parameters are adjusted.

2. The method of claim 1, wherein, The adjusting of the parameters based on the comprehensive loss value corresponding to the sample seismic inversion result comprises: based on a plurality of loss functions, calculating a plurality of sub-loss values of the sample seismic inversion result; performing weighted summation on the plurality of sub-loss values to obtain the comprehensive loss value, and adjusting the parameters based on the comprehensive loss value.

3. The method of claim 2, wherein, The plurality of loss functions comprise an attribute loss function and other loss functions, the other loss functions comprising a seismic loss function and / or a low-frequency loss function; then the calculating of the plurality of sub-loss values of the sample seismic inversion result based on the plurality of loss functions comprises: based on the attribute loss function, calculating an error between a sample predicted value of a reservoir parameter contained in the sample seismic inversion result and a sample true value of the reservoir parameter in the logging data to obtain an attribute loss value; inputting the sample prediction value into a seismic forward model to obtain synthetic seismic data, and calculating an error between the synthetic seismic data and sample seismic data of the target training sample based on the seismic loss function to obtain a seismic loss value; and / or, generating a low-frequency background model based on a spatial distribution of the logging data, and performing low-frequency filtering on the sample prediction value to obtain a low-frequency sample prediction value, and calculating an error between the low-frequency sample prediction value and a low-frequency sample true value in the low-frequency background model based on the low-frequency loss function to obtain a low-frequency loss value.

4. The method according to any one of claims 1 to 3, characterized in that, The method further comprises: inputting multi-angle seismic data to be inverted as an input parameter into the target seismic inversion model to obtain a target prediction value of a corresponding reservoir parameter of the multi-angle seismic data to be inverted.

5. A device for training a seismic inversion model, characterized in that, comprise: an acquisition module, configured to acquire a training sample set, each training sample in the training sample set comprising multi-angle sample seismic data; a training module, configured to iteratively train a seismic inversion model to be trained based on the training sample set to obtain a target seismic inversion model, the target seismic inversion model comprising a local feature perception layer module, a structural feature analysis layer module, a sequence modeling layer module, an up-sampling layer module, and a regression layer module; the local feature perception layer module comprising a plurality of depth separable large kernel attention modules, each depth separable large kernel attention module being composed of a depth separable convolution and a large kernel attention mechanism; the structural feature analysis layer module comprising a plurality of time series convolution networks; the sequence modeling layer module comprising a plurality of bidirectional gated recurrent units; wherein, in one iteration training process, the following operations are performed: performing multiple feature extractions on the selected target training sample to obtain a sample seismic feature map, wherein, in one feature extraction process, the following operations are performed: performing depth separable convolution on an input parameter of the current feature extraction to obtain a current channel feature map, and performing feature extraction on the current channel feature map through a large kernel attention mechanism to obtain a current seismic feature map, and taking the current seismic feature map as an input parameter for next feature extraction; performing time series feature extraction of different scales on the sample seismic feature map to obtain a time series feature map; performing time series dependent modeling and time domain noise filtering on the time series feature map to obtain a preliminary seismic inversion sequence; performing up-sampling on the preliminary seismic inversion sequence, and mapping the up-sampled data from a feature domain to a target domain through a regression layer to obtain a sample seismic inversion result of the target training sample; performing parameter tuning based on a comprehensive loss value corresponding to the sample seismic inversion result.

6. An electronic device, comprising: comprise: a processor; and a memory storing programs, wherein the programs comprise instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-4.

7. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to cause the computer to perform the method of any one of claims 1-4. The computer instructions are used to cause the computer to perform the method of any one of claims 1-4.

Images