Semi-supervised intelligent seismic inversion method and device based on frequency domain feature stabilization module

Abstract

The application provides a semi-supervised intelligent seismic inversion method and device based on a frequency domain feature stability module. In the training process of a target seismic inversion model, the difference between a large number of training seismic trace sets and corresponding seismic trace prediction values and the difference between a small number of seismic attribute actual values and seismic attribute prediction values are used to train an initial seismic inversion model. While ensuring the large-scale seismic data driving advantage, the local high-precision calibration characteristics of the well logging data are effectively utilized, the overfitting risk of small sample well logging data in the traditional method is relieved, and the training accuracy of the target seismic inversion model is improved. The frequency domain feature stability module can extract frequency domain features of target seismic trace data in the frequency domain space. The frequency domain features can reflect the characteristics of the target seismic trace data in different frequency intervals, better retain the stratum information contained in the target seismic trace data, and improve the accuracy of subsequent seismic inversion based on the frequency domain features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic inversion, in particular to a semi-supervised seismic intelligent inversion method and device based on a frequency domain feature stable module. BACKGROUND

[0002] Seismic inversion is one of the core technologies in geophysical exploration (especially oil and gas, mineral resources exploration), and its essence is to inversely deduce the physical properties (such as velocity, density, porosity, etc.) and geometric shape of the underground geological body from the observed seismic data. It is a bridge connecting seismic data and geological model, which can convert indirect information of seismic data into more intuitive geological information, and provides key basis for resource exploration, reservoir prediction, etc.

[0003] Due to the complex nonlinear characteristics and incomplete observation data of the seismic inversion problem, the traditional method often depends on the accuracy of the initial model and the complex regularization constraint. In recent years, deep learning provides a new way to break through this technical bottleneck by constructing a data-driven nonlinear mapping relationship. Compared with the conventional seismic inversion method which needs artificial intervention to establish an initial model, the seismic intelligent inversion based on deep learning can learn geological rules directly from massive data and realize self-compensation for the missing of observation data through an end-to-end training method. SUMMARY

[0004] Therefore, the embodiments of the present application provide a semi-supervised seismic intelligent inversion method and device based on a frequency domain feature stable module to improve the accuracy of seismic inversion.

[0005] According to an aspect of the present application, a semi-supervised seismic intelligent inversion method based on a frequency domain feature stable module is provided, which comprises:

[0006] Obtaining target seismic trace data;

[0007] Inputting the target seismic trace data into a target seismic inversion model to obtain a target seismic attribute output by the target seismic inversion model, wherein the target seismic attribute comprises target impedance field data; the target seismic inversion model is obtained by pre-training an initial seismic inversion model, and the initial seismic inversion model comprises an inversion sub-model and a forward sub-model, the inversion sub-model comprises a frequency domain feature stable module, and the target seismic inversion model is pre-trained by the following steps:

[0008] Obtaining a training seismic trace set, wherein the training seismic trace set comprises a plurality of training seismic trace data, and at least one of the training seismic trace data has a corresponding real attribute label;

[0009] inputting the training seismic trace set into an initial seismic inversion model, the inversion sub-model outputting a seismic attribute prediction result corresponding to the training seismic trace set based on the training seismic trace set and the frequency domain feature extracted by the frequency domain feature stabilization module;

[0010] inputting the seismic attribute prediction result into the forward sub-model, the forward sub-model outputting a corresponding seismic trace prediction value based on the seismic attribute prediction result;

[0011] training the initial seismic inversion model based on the difference between the seismic trace prediction value and the training seismic trace set and the difference between the seismic attribute prediction result and the true attribute label until a preset training stop condition is reached, to obtain a target seismic inversion model, the target seismic inversion model containing the frequency domain feature stabilization module.

[0012] According to another aspect of the present application, a semi-supervised seismic intelligent inversion device based on a frequency domain feature stabilization module is provided, the device comprising:

[0013] an acquisition module configured to acquire target seismic trace data;

[0014] an output module configured to input the target seismic trace data into a target seismic inversion model and acquire a target seismic attribute output by the target seismic inversion model, wherein the target seismic attribute contains target impedance field data;

[0015] a training module configured to obtain a target seismic inversion model by pre-training an initial seismic inversion model, the initial seismic inversion model containing an inversion sub-model and a forward sub-model, the inversion sub-model containing a frequency domain feature stabilization module;

[0016] acquiring a training seismic trace set containing a plurality of training seismic trace data, at least one of the training seismic trace data having a corresponding true attribute label;

[0017] inputting the training seismic trace set into an initial seismic inversion model, the inversion sub-model outputting a seismic attribute prediction result corresponding to the training seismic trace set based on the training seismic trace set and the frequency domain feature extracted by the frequency domain feature stabilization module;

[0018] inputting the seismic attribute prediction result into the forward sub-model, the forward sub-model outputting a corresponding seismic trace prediction value based on the seismic attribute prediction result;

[0019] The initial seismic inversion model is trained based on differences between the predicted values of the seismic traces and the training seismic trace set and differences between the seismic attribute prediction results and the real attribute labels until a preset training stop condition is reached, so as to obtain a target seismic inversion model, and the target seismic inversion model comprises the frequency domain feature stabilization module.

[0020] According to another aspect of the present application, an electronic device is provided, comprising a processor and a memory storing programs,

[0021] The programs comprise instructions which, when executed by the processor, cause the processor to perform the semi-supervised seismic intelligent inversion method based on the frequency domain feature stabilization module according to any one of the preceding aspects.

[0022] According to another aspect of the present application, a non-transitory computer readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to perform the semi-supervised seismic intelligent inversion method based on the frequency domain feature stabilization module according to any one of the preceding aspects.

[0023] In the one or more technical solutions provided in the embodiments of the present application, after target seismic trace data is acquired, the target seismic trace data is input into a pre-trained target seismic inversion model to acquire target seismic attributes output by the target seismic inversion model. Specifically, target impedance field data corresponding to the target seismic trace data can be output. The seismic inversion model used in the training process can comprise an inversion sub-model and a forward sub-model, and the inversion sub-model comprises a frequency domain feature stabilization module. In the model training process, the frequency domain feature stabilization module can extract frequency domain features of a training seismic trace set in a frequency domain space, the inversion sub-model can output a preliminary prediction result of seismic attributes based on the frequency domain features of the training seismic trace set and the training seismic trace set, the forward sub-model can deduce a predicted value of the training seismic trace set corresponding to the preliminary prediction result of the seismic attributes, and the initial seismic inversion model is trained based on differences between the predicted value of the training seismic trace set and the training seismic trace set and differences between the seismic attribute prediction results and real attribute labels corresponding to part of the training seismic trace set until the differences converge, so as to obtain a target seismic inversion model.

[0024] In the model training process, the difference between a large amount of training seismic trace sets and corresponding seismic trace prediction values and the difference between a small amount of seismic attribute actual values and seismic attribute prediction values are used to train an initial seismic inversion model, so that the advantages of large-scale seismic data driving are ensured, the local high-precision calibration characteristics of the logging data are effectively utilized, the overfitting risk of small sample logging data in the traditional method is alleviated, the non-physical generation problem caused by the lack of physical constraints in the pure data-driven model is avoided, and the training accuracy of the target seismic inversion model is improved. Meanwhile, the inversion sub-model comprises a frequency domain feature stabilization module, which can extract frequency domain features of the target seismic trace data in the frequency domain space, and the frequency domain features can reflect the characteristics of the target seismic trace data in different frequency intervals. Therefore, the frequency domain features of the target seismic trace data are extracted, so that the stratum information contained in the target seismic trace data can be better preserved, and the accuracy of subsequent seismic inversion based on the frequency domain features is improved. BRIEF DESCRIPTION OF DRAWINGS

[0025] In the following description of exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the present application are disclosed, in which:

[0026] Figure 1 A flowchart of a semi-supervised seismic intelligent inversion method based on a frequency domain feature stabilization module provided by the embodiment of the present application is shown in the figure.

[0027] Figure 2 A flowchart of training a target seismic inversion model in the semi-supervised seismic intelligent inversion method based on the frequency domain feature stabilization module provided by the embodiment of the present application is shown in the figure.

[0028] Figure 3 A structural diagram of the frequency domain feature stabilization module provided by the embodiment of the present application is shown in the figure.

[0029] Figure 4 A structural diagram of the inversion sub-model provided by the embodiment of the present application is shown in the figure.

[0030] Figure 5 Another flowchart of training a target seismic inversion model in the semi-supervised seismic intelligent inversion method based on the frequency domain feature stabilization module provided by the embodiment of the present application is shown in the figure.

[0031] Figure 6 An elastic impedance diagram generated in the process of generating seismic data used in the numerical experiment test in the embodiment of the present application is shown in the figure.

[0032] Figure 7 A diagram of the seismic data used in the numerical experiment test in the embodiment of the present application is shown in the figure.

[0033] Figure 8 Fig. 7 is a schematic diagram of independent inversion results of the frequency domain feature stabilization module;

[0034] Figure 9 Fig. 8 is a schematic diagram of inversion results of different models;

[0035] Figure 10 Fig. 9 is a schematic diagram of residual errors of inversion results of different models;

[0036] Figure 11 Fig. 10 is a well test result of different models;

[0037] Figure 12 Fig. 11 is a blind well test result of different models;

[0038] Figure 13 Fig. 12 is a schematic diagram of seismic data used in the effectiveness of the actual seismic data processing test method;

[0039] Figure 14 Fig. 13 is a schematic diagram of wave impedance corresponding to the seismic data used in the effectiveness of the actual seismic data processing test method;

[0040] Figure 15 Fig. 14 is a schematic diagram of inversion results obtained by using a traditional inversion method;

[0041] Figure 16 Fig. 15 is a schematic diagram of inversion results obtained by using the semi-supervised seismic intelligent inversion method based on the frequency domain feature stabilization module provided by the embodiment of the present application;

[0042] Figure 17 Fig. 16 is a structural schematic diagram of the semi-supervised seismic intelligent inversion device based on the frequency domain feature stabilization module provided by the embodiment of the present application;

[0043] Figure 18 Fig. 17 shows a structural block diagram of an exemplary electronic device that can be used to implement embodiments of the present application. DETAILED DESCRIPTION

[0044] Embodiments of the present application will be described in more detail by referring to the attached drawings. Although certain 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 should not be used to limit the scope of protection of the present application.

[0045] It should be understood that each step described 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 the steps shown. The scope of the present application is not limited in this respect.

[0046] The term "comprising" and variations thereof as used herein are open-ended, and mean "including but not limited to". The term "based on" means "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 reference to "first", "second" and the like in the present disclosure indicates "different" devices, modules or units and does not mean that these are "first" and "second" in order of their implementation or importance.

[0047] It should be noted that the modification of "one", "multiple" mentioned in the present application is illustrative and not restrictive, and those skilled in the art should understand that unless otherwise explicitly indicated in the context, it should be understood as "one or more".

[0048] 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.

[0049] In order to improve the accuracy of seismic inversion, the present application provides a semi-supervised intelligent seismic inversion method based on a frequency domain feature stable module, a device, an electronic equipment and a storage medium. The semi-supervised intelligent seismic inversion method based on the frequency domain feature stable module can be applied to any electronic equipment with intelligent seismic inversion function. The electronic equipment can be a server, a computer, an industrial computer or a mobile terminal, etc. The scheme of the present application is described below with reference to the accompanying drawings:

[0050] Figure 1 A flowchart of the semi-supervised intelligent seismic inversion method based on the frequency domain feature stable module provided by the embodiments of the present application can include the following steps:

[0051] S101, acquiring target seismic trace data;

[0052] S102, input the target seismic trace data into a target seismic inversion model to obtain target seismic attributes output by the target seismic inversion model, wherein the target seismic attributes comprise target impedance field data; the target seismic inversion model is obtained by pre-training an initial seismic inversion model, the initial seismic inversion model comprises an inversion sub-model and a forward sub-model, the inversion sub-model comprises a frequency domain feature stabilization module, and the target seismic inversion model is pre-trained by the following steps:

[0053] S201, obtain a training seismic trace set, wherein the training seismic trace set comprises a plurality of training seismic trace data, and at least one piece of the training seismic trace data has a corresponding real attribute label;

[0054] S202, input the training seismic trace set into an initial seismic inversion model, wherein the inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic trace set based on the training seismic trace set and frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module;

[0055] S203, input the seismic attribute prediction result into the forward sub-model, wherein the forward sub-model outputs a corresponding seismic trace prediction value based on the seismic attribute prediction result;

[0056] S204, train the initial seismic inversion model based on differences between the seismic trace prediction value and the training seismic trace set and differences between the seismic attribute prediction result and the real attribute label until a preset training stop condition is reached to obtain a target seismic inversion model, wherein the target seismic inversion model comprises the frequency domain feature stabilization module.

[0057] In the embodiment of the application, after obtaining target seismic trace data, the target seismic trace data is input into a pre-trained target seismic inversion model to obtain target seismic attributes output by the target seismic inversion model, and specifically, target impedance field data corresponding to the target seismic trace data can be output. The seismic inversion model used in the training process can comprise an inversion sub-model and a forward sub-model, and the inversion sub-model comprises a frequency domain feature stabilization module. In the model training process, the frequency domain feature stabilization module can extract frequency domain features of a training seismic trace set in a frequency domain space, the inversion sub-model can output a preliminary prediction result of seismic attributes based on the frequency domain features of the training seismic trace set and the training seismic trace set, the forward sub-model can deduce a seismic trace set prediction value corresponding to the preliminary prediction result of the seismic attributes based on the preliminary prediction result of the seismic attributes, and the initial seismic inversion model is trained based on differences between the seismic trace set prediction value and the training seismic trace set data and differences between the seismic attribute prediction result and real seismic attribute labels corresponding to part of the training seismic trace set data until the differences converge to obtain a target seismic inversion model.

[0058] By applying the embodiments of the present application, in the model training process, the difference between a large amount of training seismic trace sets and corresponding seismic trace prediction values and the difference between a small amount of seismic attribute actual values and seismic attribute prediction values are used to train the initial seismic inversion model, while ensuring the advantages of large-scale seismic data driving, effectively utilizing the local high-precision calibration characteristics of the logging data, and alleviating the overfitting risk of small sample logging data in the traditional method, while avoiding the non-physical generation problem caused by the lack of physical constraints in the pure data-driven model, and improving the training accuracy of the target seismic inversion model. At the same time, since the inversion sub-model includes a frequency domain feature stabilization module, the frequency domain feature stabilization module can extract frequency domain features of the target seismic trace data in the frequency domain space, and the frequency domain features can reflect the characteristics of the target seismic trace data in different frequency intervals. Since the underlying information reflected by the seismic trace data is different at different frequencies, extracting the frequency domain features of the target seismic trace data can better preserve the stratum information contained in the target seismic trace data, and improve the accuracy of subsequent seismic inversion based on the frequency domain features.

[0059] In seismic exploration, seismic trace is the basic data unit for recording the reflection or refraction of seismic signals in underground medium, and is the core carrier of seismic data acquisition, processing and interpretation. In actual application scenarios, artificial seismic sources are usually used to emit seismic waves into the underground, the seismic waves propagate in the underground medium, and reflection occurs when they encounter interfaces with different lithology or physical properties. The reflected waves are received by geophones arranged on the ground and receivers arranged in the well, and are converted into electrical signals. The electrical signals are recorded by related instruments and arranged in time sequence to form a time-amplitude curve, which is the seismic trace data. In S101, the seismic trace data detected by each acquisition point can be obtained as target seismic trace data according to a preset time period, which can be set according to actual application scenarios, such as one week, one month, etc.

[0060] A pre-trained target seismic inversion model can be deployed in an electronic device. After obtaining the target seismic trace data, the target seismic trace data can be input into the target seismic inversion model. The target seismic inversion model can output target seismic attributes for the target seismic trace data. The type of the target seismic attributes can be selected according to actual application scenarios. For example, the target seismic attributes can include one or more reservoir parameters, such as elastic impedance (EI) parameter field, density, shear wave velocity, and compressional wave velocity, etc. The specific type of the target seismic inversion data should be the same as the label type used in the loss calculation during the training process of the target seismic inversion model. The present application will be exemplarily described below taking the elastic impedance parameter as an example. The elastic impedance field is a physical field that specifically describes the spatial distribution characteristics of underground rock in the elastic mechanics attribute.

[0061] The target seismic inversion model can be pre-trained through S201-S204, which are exemplarily described as follows:

[0062] The training seismic trace set can be obtained from historical seismic trace data. Exemplarily, after each measured seismic trace data is obtained, it can be stored in a database. When training the seismic inversion model is needed, a plurality of historical seismic trace data can be obtained from the database to form the training seismic trace set. In actual application, wells are drilled and measured at some geological points to obtain the real impedance field data of these geological points. Therefore, the seismic trace data at these well points in the database has corresponding real impedance field data. As a possible implementation manner, the seismic trace data stored in the database can include seismic trace data acquisition point information, and the impedance field data measured at each well point can also include corresponding acquisition point information. Then, the seismic trace data and the impedance field data can be associated based on the seismic trace data acquisition point information and the impedance field data acquisition point information, and stored in the database correspondingly. When the training seismic trace set is obtained, seismic trace data including impedance field labels and seismic trace data not including impedance field labels can be obtained from the database to form the training seismic trace set.

[0063] The initial seismic inversion model can include an inversion sub-model and a forward sub-model. The inversion sub-model is used to predict the impedance field based on the input training seismic trace set, and the forward sub-model is used to deduce the seismic trace data based on the predicted impedance field output by the inversion sub-model to obtain the seismic trace prediction value. The inversion sub-model and the forward sub-model are exemplarily described as follows:

[0064] The inversion sub-model can include a frequency domain feature stabilization module, which can extract the frequency domain features of each training seismic trace data in the training seismic trace set. The training seismic trace data included in the training seismic trace set is usually time domain data, that is, it can reflect the change of seismic wave with time. In related technologies, the seismic trace data features are usually extracted by globally convolving the time domain seismic trace data, but this method has many limitations. It needs a convolution kernel that matches the input sequence length, which will cause excessive padding, dimension compression and increase in computational complexity. In order to overcome the above problems that may be caused by convolution of time domain seismic trace data, in the present application, the frequency domain analysis of seismic trace data can be performed through the frequency domain feature stabilization module. Frequency domain analysis is a process of decomposing a signal into different frequency components. The information of the stratum reflected by the seismic signal is different at different frequencies. Low frequency corresponds to large-scale structure, and high frequency corresponds to details. Therefore, by extracting frequency domain features, more comprehensive information of the seismic trace data can be retained. In inversion, directly using time domain data is easily affected by local outliers, while after frequency domain conversion, the algorithm can process features of different scales respectively, thereby reducing local interference.

[0065] In a possible embodiment, the frequency domain feature stabilization module can include an embedding sub-module, a frequency domain conversion sub-module, a channel learning sub-module, a frequency learning sub-module, and a time domain conversion sub-module; the frequency domain feature stabilization module extracts the frequency domain features of the training seismic trace set through the following steps:

[0066] S21, the embedding sub-module embeds the training seismic trace set to obtain a second seismic trace embedding vector;

[0067] S22, the frequency domain conversion sub-module converts the second seismic trace embedding vector in the frequency domain to obtain a frequency domain vector corresponding to the second seismic trace embedding vector;

[0068] S23, the channel learning sub-module linearly transforms the frequency domain vector in the channel dimension to obtain a channel vector;

[0069] S24, the frequency learning sub-module linearly transforms the channel vector in the frequency dimension to obtain a frequency vector;

[0070] S25, the time domain conversion sub-module converts the frequency vector in the time domain to obtain the frequency domain features of the training seismic trace set.

[0071] The embedding process can be performed in the time domain. As a possible implementation, the embedding sub-module can include a trainable parameter matrix, based on which the input training seismic trace data can be transformed in the channel dimension to obtain the corresponding embedding result. This process can be represented by the following formula:

[0072] (1)

[0073] wherein, is the input data, represents a real number field, N is the number of seismic traces, C is the number of channels, and T is the number of sampling points, is a learnable parameter matrix, and C'x C is the dimension of the learnable parameter matrix, is the embedding result, i.e., the second seismic trace embedding vector. Then, the second seismic trace embedding vector and the original training seismic trace data can be concatenated in the channel dimension to obtain the concatenated seismic trace data X concat , which can be represented by the following formula:

[0074] (2)

[0075] By concatenating the original training seismic trace data and the second seismic trace embedding vector, the original data and the embedding information can be effectively fused, and the representation ability of the feature space can be improved.

[0076] The frequency domain conversion submodule can perform time domain-frequency domain conversion on the spliced seismic trace data, i.e., convert the time domain spliced seismic trace data into frequency domain spliced seismic trace data, which can contain real and imaginary parts. For example, the frequency domain conversion submodule can perform a short-time Fourier transform (DFT) on the spliced seismic trace data to obtain its frequency domain representation.

[0077] The channel learning submodule and the frequency learning submodule can perform channel dimension and frequency dimension feature extraction on the spliced seismic trace data. The channel learning submodule can include a trainable channel parameter matrix for performing linear transformation on the spliced seismic trace data in the channel dimension, and an activation layer that can activate the features obtained after linear transformation by a preset activation function, such as Tanh. The channel learning submodule can output channel features (channel vectors described above) to the frequency learning submodule.

[0078] The frequency learning submodule can include a trainable frequency parameter matrix for performing linear transformation on the channel features in the frequency dimension, and an activation layer that can activate the features obtained after linear transformation by a preset activation function.

[0079] The trainable parameter matrices in the channel learning submodule and the frequency learning submodule can exist in the form of a fully connected layer. The global convolution operation in the time domain can be represented as:

[0080] (3)

[0081] wherein, is a global convolution kernel, is an input signal, is an output signal, and N is the number of channels, which can be set according to the actual application scenario.

[0082] Discrete Fourier transform (DFT) is performed on the input signal x and the convolution kernel W to obtain their frequency domain representations X and W f :

[0083] (4)

[0084] (5)

[0085] wherein, F represents a discrete Fourier transform, is a complex vector space, and the frequency domain representation of the output signal y can be:

[0086] (6)

[0087] wherein, represents element-wise multiplication. Since the input signal in time domain and the frequency domain representation obtained after Fourier transform of the convolution kernel are usually in complex number form, the above equation can be split into real and imaginary parts as follows:

[0088] (7)

[0089] wherein, is the real part of the output signal Y, is the imaginary part of the output signal Y, is the real part of the convolution kernel, is the imaginary part of the convolution kernel, is the real part of the input signal, is the imaginary part of the input signal.

[0090] Then the operation in frequency domain can be represented as:

[0091] (8)

[0092] (9)

[0093] The concatenated seismic trace data after time-to-frequency domain conversion is linearly transformed according to equation (8) and (9), and there is a redundant calculation of real and imaginary parts. The essence of the fully connected layer is linear transformation, therefore, the trainable parameter matrix in the channel learning sub-module and the frequency learning sub-module can exist in the form of a fully connected layer, which not only simplifies the operation process, but also improves the calculation efficiency and the scalability of the model.

[0094] The channel learning sub-module and the frequency learning sub-module described above can constitute a frequency domain feature extraction module, and the frequency domain feature stabilization module can include a plurality of such frequency domain feature extraction modules, and the plurality of frequency domain feature extraction modules can optimize the frequency domain features through multiple rounds of iteration.

[0095] In a possible embodiment, in order to strengthen the model's capture efficiency of global significant features, a hierarchical normalization strategy can be adopted, such as implementing channel-by-channel normalization and inverse normalization processing before and after time-to-frequency domain conversion. Specifically, a normalization layer can be arranged before the frequency domain feature extraction module to perform channel-by-channel normalization on the concatenated seismic trace data in the frequency domain, so that the feature value range of each channel is unified. The normalization layer can use any feasible normalization method to perform channel-by-channel normalization on the concatenated seismic trace data, which is not limited in the present application. The inverse normalization layer can be arranged after the frequency domain feature extraction module, that is, the candidate frequency domain features extracted by the frequency domain feature extraction module can be input to the inverse normalization layer, and the inverse normalization layer can inverse normalize the candidate frequency domain features according to the parameters of the normalization layer to restore the features to the original scale.

[0096] As Figure 3As shown, Figure 3 A structural schematic diagram of the frequency domain feature stabilization module provided by the embodiment of the present application is shown. Specifically, after the training seismic trace data is input into the frequency domain feature stabilization module, the embedding layer first embeds the training seismic trace data, and the embedded data and the training seismic trace data are spliced in the channel dimension in the form of a complex number. The spliced seismic trace data contains a real part real and an imaginary part imag. Normalization-frequency domain feature extraction-reverse normalization can be performed on the spliced seismic trace data.

[0097] The frequency domain feature extraction is realized through a channel learning submodule and a frequency learning submodule. The channel learning submodule includes a linear transformation layer Linear layer of channel (channel learning layer), a group normalization layer GroupNormalization, and an activation layer. The linear transformation layer can be a full connection layer, which is used to perform linear transformation on the features in the channel dimension of the normalized spliced seismic trace data, so as to adjust the feature combination between channels. The group normalization layer is used to normalize the features after channel grouping, stabilize the training process, and reduce internal covariate shift. The activation layer can activate the features obtained after group normalization through a preset activation function.

[0098] The frequency learning submodule includes a linear transformation layer (frequency learning layer), a group normalization layer GroupNormalization, and an activation layer. The frequency learning layer can perform linear transformation on the channel features output by the channel learning submodule in the frequency dimension, so as to adjust the feature distribution in the frequency domain. The group normalization layer is used to perform group normalization after frequency grouping. The activation layer can activate the features obtained after group normalization through a preset activation function.

[0099] The channel learning submodule and the frequency learning submodule can constitute a repeated module, so as to optimize the features in the channel and frequency dimensions through multiple iterations, and obtain candidate frequency domain features.

[0100] The candidate frequency domain features can then be subjected to reverse normalization. The reverse normalization layer can record the mean and variance of each channel data before normalization in each batch, and use the corresponding statistical data to perform reverse normalization in the subsequent reverse normalization layer, with each channel in each batch being recorded separately. The data after reverse normalization can be split into a real part real and an imaginary part imag, that is, the data after reverse normalization is converted into a complex number form. The data after reverse normalization in the complex domain is then inversely transformed (time-frequency conversion), to obtain the frequency domain features corresponding to the training seismic trace data in the time domain form.

[0101] In a possible embodiment, the inversion submodel can further include a global feature extraction module, which can extract features of different scales from the input training seismic trace data.

[0102] The global feature extraction module comprises an embedding layer, a multi-head attention layer and a convolution layer, and the convolution layer comprises convolution blocks of different scales, which specifically refer to convolution blocks with different receptive fields. The receptive field of a convolution block describes the size of the region perceived by the neurons in the convolution block in the input data, and the size of the receptive field is related to the size of the convolution kernel, the convolution step, the padding method and the number of network layers. The method further comprises:

[0103] S31, inputting the training seismic trace set to the global feature extraction module, and embedding each training seismic trace data by the embedding layer to obtain a first seismic trace embedding vector;

[0104] S32, inputting the first seismic trace embedding vector to the multi-head attention layer, and performing attention calculation on the first seismic trace embedding vector by the multi-head attention layer to obtain seismic trace attention features;

[0105] S33, inputting the seismic trace attention features and the first seismic trace embedding vector after splicing to the convolution layer, and extracting different scale features of the spliced data by the convolution layer through the convolution blocks of different scales; wherein the spliced data is obtained by splicing the seismic trace attention features and the first seismic trace embedding vector;

[0106] S34, fusing the different scale features of the spliced data to obtain the global features of the training seismic trace set;

[0107] The inversion sub-model outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the training seismic trace set and the frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module, comprising:

[0108] The inversion sub-model outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the global features of the training seismic trace set and the frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module.

[0109] As a possible implementation, after inputting the training seismic trace set to the inversion sub-model, the embedding layer in the global feature extraction module can embed the training seismic trace data to obtain its vector representation, which will be referred to as the first seismic trace embedding vector in the following for convenience of description.

[0110] The multi-head attention layer can perform multi-head attention calculation on each first seismic trace embedding vector. The multi-head attention layer can include multiple heads, and each head can independently perform attention calculation on the training seismic trace vector. For example, the multi-head attention layer can include multiple heads, and each head can include a query matrix Q, a key matrix K, and a value matrix V. Linear transformation of the QKV matrix in each head can obtain a query vector, a key vector, and a value vector in each head. Each head in the multi-head attention layer can transform the input training seismic trace data according to the query vector, the key vector, and the value vector to obtain features extracted by each head. The features extracted by each head are spliced to obtain the attention features output by the multi-head attention layer.

[0111] In a possible embodiment, the attention features can be connected in residual connection with the first seismic trace embedding vector to obtain spliced data, so as to alleviate the gradient disappearance problem of the deep network and make information more smoothly transmitted within the module.

[0112] The convolution layer can include multiple convolution blocks with different receptive fields. As a possible implementation, the attention layer can include four groups of parallel one-dimensional dilated convolution layers, and the dilated factors of the convolution layers can be 1, 3, 6, and 9, respectively, to construct a ladder-shaped receptive field and focus on the analysis of features of reservoir microscopic lithology changes, medium-scale sedimentary units, and regional tectonic background, respectively.

[0113] The multi-scale features extracted by the convolution blocks with different scales can be spliced according to the channels to fuse information of different scales. The features obtained after channel splicing can be set to be convolved by the convolution layer to integrate global features.

[0114] In a possible embodiment, the inversion sub-model can further include a time series modeling module, which can be constructed based on a Bi-LSTM (Bidirectional Long Short-Term Memory Network). LSTM is a special recurrent neural network (RNN) designed to solve the gradient vanishing and gradient exploding problems existing in traditional RNN, so as to effectively process long sequence data. Its core components include: memory cell (Cell): used to store long-term information, which can decide which information to keep in memory and which information to forget. Input gate (Input Gate): controls which information in the current input data will be added to the memory cell. Forget gate (Forget Gate): decides which old information in the memory cell will be discarded. Output gate (Output Gate): decides the final output according to the state of the memory cell and the current input. Bi-LSTM can include a forward LSTM and a backward LSTM, wherein the forward network is used to process data in the forward order of the input sequence, and the backward network is used to process data in the reverse order of the input sequence. The hidden states of the two directions of LSTM are connected at each time step and then passed to the next layer (such as a fully connected layer) for further processing.

[0115] The time series modeling module can include a plurality of Bi-LSTM modules connected in series, and the training seismic trace set is subjected to bidirectional recursive operation through the plurality of Bi-LSTM modules to obtain the time series features corresponding to the training seismic trace set.

[0116] The inversion sub-model outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the global features of the training seismic trace set and the frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module, and includes:

[0117] The inversion sub-model outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the global features of the training seismic trace set, the time series features, and the frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module.

[0118] Through the above technical solution, bidirectional recursive operation is adopted to simulate the wave field propagation path from shallow to deep and from deep to shallow, and the stratum deposition sequence rule contained in the seismic signal is analyzed layer by layer. The information can be selectively retained and updated, so that the network can learn the long-distance dependence.

[0119] In a possible embodiment, the inversion sub-module can further include an up-sampling module, and the method further includes:

[0120] The global features and the time series features are spliced to obtain global time series features;

[0121] The global time sequence feature is input to the upsampling module, and the upsampling module upsamples the global time sequence feature through a plurality of deconvolution submodules to obtain an expanded feature;

[0122] The inversion submodel outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the global feature of the training seismic trace set, the time sequence feature, and the frequency domain feature extracted by the frequency domain feature stabilization module.

[0123] The inversion submodel outputs the seismic attribute prediction result corresponding to the training seismic trace set based on the expanded feature and the frequency domain feature extracted by the frequency domain feature stabilization module.

[0124] The upsampling module can upsample the time sequence feature of the training seismic trace data, which can include the feature extracted by the global feature extraction module and the feature extracted by the time sequence modeling module. As a possible implementation, a two-level differentiated step transposed convolution architecture can be used in the upsampling module, and the vertical resolution of the feature map is gradually improved through the de-projection mechanism and dynamic kernel parameter optimization. Specifically, the upsampling module can include two deconvolution modules, and the steps of the two deconvolution modules are different. For example, the step of one of the deconvolution modules is 2, which can perform a two-fold upsampling operation to increase the resolution of the feature map.

[0125] The deconvolution module can include a deconvolution layer, a normalization layer, and an activation function. The deconvolution layer can include a deconvolution kernel, and the parameter of the deconvolution kernel is a trainable parameter. The normalization layer is used for grouping and normalizing the features obtained by deconvolution, and the activation function is used to activate the normalized features to adaptively recover the high-frequency detail features of the seismic signal and bridge the vertical resolution difference between the seismic data and the logging data. The activation function can be a hyperbolic tangent activation function.

[0126] The inversion submodel can also include a regression module. The regression module is a terminal mapping unit that fuses the frequency domain features output by the frequency domain stabilization module and the high-frequency features output by the upsampling through a multi-layer perceptron (MLP) architecture and realizes weighted superposition of full-band feature fields through residual connection. The MLP internally analyzes the petrophysical correlation rules between seismic attributes and elastic impedance through hidden layer nonlinear transformation, and finally maps the fused features to the target parameter space to generate high-resolution elastic impedance field prediction values that conform to the wave equation constraint.

[0127] As shown in Figure 4 , the upsampling module can include two deconvolution modules, and the steps of the two deconvolution modules are different. For example, the step of one of the deconvolution modules is 2, which can perform a two-fold upsampling operation to increase the resolution of the feature map. Figure 4A structural schematic diagram of the inversion sub-model provided by the embodiment of the present application is shown in FIG. 1. The input of the inversion sub-model is original seismic data Seismic, which contains wave information of underground geological structure. The inversion sub-model extracts features of seismic trace data from different angles through three feature extraction modules, which can specifically include:

[0128] A global multi-scale feature extraction module, which includes an embedding layer for converting the original seismic data into a feature vector that can be processed by the model. A multi-head attention layer: allows the model to focus on key areas in the seismic data, such as strong reflection events and faults, and capture long-range dependencies, such as the correlation of seismic signals at different locations. Residual connection: alleviates the gradient vanishing problem of deep networks and stabilizes training. Multiple ConvBlock (convolutional layers): extract multi-scale local features through convolution operations (small convolution kernels capture fine-grained details, and large convolution kernels capture macro trends). Concatenation + convolution block: concatenate the results of multi-scale convolution and then use the convolution block to fuse them to obtain multi-scale features from a global perspective.

[0129] A sequence module (i.e., the above-mentioned time series modeling module) includes multiple LSTM (Long Short-Term Memory networks): LSTM is good at processing sequence data, and seismic data is naturally a "time series", such as the amplitude change of seismic trace over time or the spatial sequence in the direction of the survey line. Through LSTM, the model can learn the "sequence dependency" of seismic data, such as the waveform correlation of adjacent seismic traces, the energy change law of the same trace at different times, etc.

[0130] Frequency domain numerical stabilisation module: Embedding layer + time-frequency conversion layer: convert seismic data to frequency domain (e.g. Fourier transform) and split into real / imaginary parts, etc. complex form, while doing feature encoding. Channel learning module is used to learn the features of the channel dimension (if the input is multi-channel seismic data, the channel corresponds to different seismic traces; if it is single-channel, the channel corresponds to different feature dimensions). Frequency learning module is used to learn the features of the frequency dimension, i.e. it can learn seismic waves of different frequency components, such as low-frequency seismic waves reflecting deep structures and high-frequency seismic waves reflecting shallow details. The repeated Learner structure allows the model to capture the frequency characteristics of the seismic wave more finely in the frequency domain, while numerical stability ensures that the frequency domain operation will not distort the results due to numerical problems (such as overflow, oscillation). The upsampling module includes two deconvolution blocks to enlarge the resolution of the feature map, as seismic data usually requires high-resolution output, such as small-scale reservoir prediction, and enlarging the resolution of the feature map is beneficial to subsequent output of high-resolution prediction results. The regression module includes LSTM and linear layers. Specifically, the regression module first concatenates the frequency domain features and the high-resolution features output by the upsampling module, and inputs the concatenated features into the LSTM. The LSTM can continue to learn the sequence rules of the features after upsampling. The linear layer combines the low-frequency information (such as frequency domain features) and the features output by the LSTM to output seismic attributes through linear transformation. Seismic attributes are indicators that quantitatively describe underground geological features (such as amplitude envelope reflecting oil and gas bearing reservoirs, and coherence reflecting faults / fractures), providing key evidence for geological interpretation and reservoir prediction.

[0131] The model extracts the features of seismic data through global multi-scale + sequence dependence + frequency domain fine-grained three dimensions, and then outputs high-resolution and accurate seismic attributes through upsampling and regression. This multi-module cooperation can simultaneously consider the "global-local" "time-space" "spatial-frequency" features of seismic data, improving the accuracy and stability of attribute prediction.

[0132] The seismic attribute prediction parameters output by the inversion sub-model can be input into the forward sub-model, which can be constructed based on the CNN structure, for predicting the corresponding seismic trace parameters based on seismic attribute prediction. The convolution kernel parameters in the forward sub-model are trainable values, which simulate the seismic wave field propagation process through parameterized convolution kernels, achieving efficient gradient calculation while ensuring physical rationality.

[0133] The loss function used in the training process can combine the seismic trace loss and the seismic attribute loss. As described above, the training seismic trace data contained in the training seismic trace set has data with seismic attribute labels. For this part of the training seismic trace data, the difference between the corresponding seismic attribute prediction value and the seismic attribute label, and the sum of the difference between the training seismic trace data and the seismic trace prediction value, can be used as the corresponding loss function. Specifically, the total loss function (L(Θ)) can be calculated by the following formula:

[0134] (10)

[0135] wherein, is the actual EI label, the function represents the forward process, i.e. the processing of the forward sub-model, the function represents the inversion process, represents the seismic data, i is the number of seismic data, is the number of available well records, is the total number of seismic traces, in seismic exploration , α and β are preset coefficients.

[0136] Based on the loss, the parameters in the seismic inversion model can be adjusted by gradient descent method until the total loss converges, the preset training stop condition is reached, and the target seismic inversion model is obtained. Specifically, the trained inversion sub-model can be used as the target seismic inversion model. The total loss convergence refers to that the total loss is less than a preset loss threshold or the difference between the total losses obtained by two times of training is less than a preset difference threshold.

[0137] As shown in FIG. Figure 5 , the target seismic inversion model training process in the embodiment of the present application is a schematic diagram, specifically, the low-frequency training seismic trace data (i.e. low-frequency data, seismic data in the figure) is input into the inversion model in the inversion sub-model, Figure 5 The inversion sub-model performs seismic attribute prediction based on the training seismic trace data to obtain the seismic attribute prediction value corresponding to the training seismic trace data. The seismic attribute prediction value is input into the forward model in the forward sub-model, Figure 5 The forward sub-model performs seismic trace data prediction based on the seismic attribute prediction value to obtain the corresponding seismic trace prediction value (predicted seismic in the figure). The attribute loss losswell is calculated based on the difference between the seismic attribute prediction value and the well logging data. Based on the attribute loss, the parameters of the inversion sub-model can be updated. The seismic loss lossseismic is calculated based on the seismic trace prediction value and the training seismic trace data. Based on the seismic loss, the parameters of the inversion sub-model are updated. Figure 5

[0138] ​By the above technical solution, while ensuring the advantages of large-scale seismic data driving, the local high-precision calibration characteristics of the logging data are effectively utilized. This collaborative optimization mechanism significantly alleviates the overfitting risk of small sample logging data in the traditional method, and avoids the non-physical generation problem caused by the lack of physical constraints in the pure data-driven model.

[0139] After the model training is completed, the trained inversion sub-model can be stored in the electronic device as a target seismic inversion model to realize the prediction of target seismic trace data.

[0140] Figures 6-16 The numerical experiment test results and the actual seismic data test results provided by the embodiments of the present application are shown, wherein the numerical experiment data set is constructed based on the Marmousi 2 classical elastic model in the numerical experiment test, and the performance difference between the semi-supervised seismic intelligent inversion method based on the frequency domain feature stable module provided by the embodiments of the present application and the benchmark model is compared. The inversion effects of the base sequence model, the simplified seismic inversion model without the frequency domain stable module and the complete seismic inversion model are respectively tested by the control variable method, and the impedance recovery accuracy and the lateral continuity in the wide-angle seismic profile of different models are mainly compared, the improvement effect of the frequency domain feature stable module on the inversion stability and the noise resistance is verified, and the method theory feasibility is confirmed.

[0141] The actual seismic data test method effectiveness selects the actual seismic data (a total of 667 traces) of M area in eastern China to carry out application test, and carries out elastic impedance reconstruction by using the semi-supervised inversion framework based on the frequency domain feature stable module. Through the lateral consistency analysis of the inversion results, the well logging interpretation results and the geological structure model, the high resolution characterization ability of the method to the complex stratum structure and the applicability to the actual exploration scene are verified, and the effectiveness of the inversion method is tested. The following two kinds of tests are exemplarily described:

[0142] Based on the Marmousi 2 classical model, the depth domain elastic parameters (ρ (density), Vp (longitudinal wave velocity), Vs (transverse wave velocity)) are converted into a high-fidelity time domain model (sampling interval 1 ms), and an effective time window (678-2,814 ms) is intercepted to avoid distortion area; based on the Whitcombe theory, the three-incidence-angle (θ=0°, θ=15° and θ=30°) elastic impedance model (EI) is calculated as shown in Figure 6 , combined with the linear convolution model to synthesize multi-angle seismic response, after introducing 15 dB Gaussian white noise to simulate actual interference, 6 times down-sampling is implemented to generate low-resolution seismic data, as shown in Figure 7As shown, a sixth-order Butterworth filter is used to extract low-frequency background constraints, and pseudo-well samples are constructed along the target layer. The final dataset integrates downsampled seismic data, pseudo-well elastic parameters, and regional low-frequency information. Through a multi-scale collaborative mechanism, frequency band gaps are filled and the ability to identify thin interlayers is enhanced, providing a foundation for reservoir inversion that combines macroscopic geological constraints with microscopic feature representation.

[0143] The independent inversion results of the frequency characteristic stabilization module are as follows: Figure 8 As shown, Figure 8 The leftmost diagram shows the results of the frequency domain feature stabilization module (the prediction results of the frequency module). It can be seen that it can accurately capture the overall structural features of the data, but there is a systematic bias in the values. The right diagram shows the absolute error (residual) between the inversion results and the true EI. The right diagram shows that while the prediction results maintain excellent lateral continuity, there is still an overall bias. Experiments show that the frequency domain feature stabilization module is good at extracting global features and constraining the time-domain inversion process, effectively improving lateral continuity and the ability to guide the inversion direction. However, due to the existence of systematic errors, it is not suitable to independently undertake the entire inversion task.

[0144] Inversion results of different models, such as Figure 9 As shown, the residuals of the inversion results are as follows: Figure 10 As shown. Figure 9 as well as Figure 10 From left to right, the figures are: a normal sequence model, a sequence model containing a frequency domain feature stabilization module (the frequency module in the figure), the target earthquake inversion model provided by this invention without a frequency domain feature stabilization module, the inversion result of the target earthquake inversion model provided by this invention, and the residual of the inversion result.

[0145] Experimental results show that the sequence model incorporating the frequency domain numerical stabilization module (frequency domain feature stabilization module) significantly improves the lateral continuity of the inversion results compared to the traditional time series model. Especially in regions of tectonic abrupt change, the frequency domain feature stabilization module effectively suppresses the lateral non-uniformity noise caused by the temporal limitations of conventional sequence modeling networks through a global frequency feature reweighting mechanism, thereby improving the matching degree between the stratigraphic interface extension features and the regional tectonic background.

[0146] Figure 11 The results of the well bypass test are as follows. Figure 12 The results of blind well testing include, specifically, the well-side tunnel test results and blind well test results of the ordinary sequence model, the sequence model with a frequency domain module, the target seismic inversion model without a frequency domain feature stabilization module (the suggested model without a frequency model in the figure), and the target seismic inversion model provided in the embodiments of the present invention (the suggested model in the figure).

[0147] Test results show that in the 1250-1500 depth range, the introduction of the frequency stabilization module significantly improves the sequence model's ability to capture low-frequency phase features; for the 1800-2100 depth range, the module also effectively suppresses amplitude attenuation and phase shift in the proposed model. Particularly in the 1500-1750 depth range, the complete proposed model exhibits superior waveform continuity compared to the basic sequence model, validating its architecture's advantage in representing complex geological signals. Although the overall prediction accuracy decreases at distant wells, the frequency module still achieves optimal waveform matching across angles through multi-scale feature fusion. The two figures jointly verify the key role of the frequency stabilization module in enhancing generalization ability.

[0148] The numerical model test specifically involves applying the semi-supervised intelligent seismic inversion method based on frequency domain characteristic stabilization modules provided in this invention to post-stack seismic data from an exploration block in the Shengli Oilfield in eastern China, such as... Figure 13 As shown, this dataset contains 667 seismic traces with a time sampling interval of 1 ms. Based on the well-by-well logging curves and lithofacies interpretation results within the work area, a low-frequency wave impedance model is constructed using a local Kriging interpolation algorithm, as follows: Figure 14 As shown, this provides initial geological constraints for the inversion process.

[0149] Compared with traditional inversion methods ( Figure 15 The method provided by this invention ( Figure 16 The proposed method demonstrates significant advantages in both vertical resolution and lateral continuity: the impedance boundaries of thin layers are more clearly delineated, and the transition zone between formation pinch-out points and fault contact zones exhibits smoother seismic response characteristics. The spatial distribution of faults within the black-circled area shows higher confidence in the inversion results, with improved agreement with fault interpretation results from adjacent well locations. Well logging curve comparisons verify that the proposed method outperforms traditional methods in velocity-impedance relationships within thin interbedded sandstone and mudstone sections. Its improved lateral continuity effectively reduces the ambiguity of structural boundaries such as formation pinch-outs and faults, validating the method's applicability in actual work areas.

[0150] Based on the same inventive concept, this invention also provides a semi-supervised intelligent seismic inversion device based on a frequency domain characteristic stabilization module, such as... Figure 17 As shown, the device 1700 may include:

[0151] Module 1701 is used to acquire target seismic trace data;

[0152] The output module 1702 is used to input the target seismic trace data into the target seismic inversion model and obtain the target seismic attributes output by the target seismic inversion model, wherein the target seismic attributes include target impedance field data;

[0153] The training module 1703 is configured to train a target seismic inversion model in advance according to the following steps, wherein the target seismic inversion model is obtained by pre-training an initial seismic inversion model, the initial seismic inversion model comprises an inversion sub-model and a forward sub-model, and the inversion sub-model comprises a frequency domain feature stabilization module;

[0154] Obtain a training seismic trace set, wherein the training seismic trace set comprises a plurality of training seismic trace data, and at least one piece of the training seismic trace data has a corresponding true attribute label;

[0155] Input the training seismic trace set into the initial seismic inversion model, wherein the inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic trace set based on the training seismic trace set and frequency domain features of the training seismic trace set extracted by the frequency domain feature stabilization module;

[0156] Input the seismic attribute prediction result into the forward sub-model, wherein the forward sub-model outputs a corresponding seismic trace prediction value based on the seismic attribute prediction result;

[0157] Train the initial seismic inversion model based on differences between the seismic trace prediction value and the training seismic trace set and differences between the seismic attribute prediction result and the true attribute label until a preset training stop condition is reached, to obtain a target seismic inversion model, wherein the target seismic inversion model comprises the frequency domain feature stabilization module.

[0158] In a possible embodiment, the inversion sub-model further comprises a global feature extraction module, the global feature extraction module comprises an embedding layer, a multi-head attention layer and a convolution layer, and the convolution layer comprises convolution blocks of different scales;

[0159] The training module is configured to input the training seismic trace set into the global feature extraction module, and the embedding layer embeds each piece of the training seismic trace data to obtain a first seismic trace embedding vector;

[0160] The first seismic trace embedding vector is input into the multi-head attention layer, the multi-head attention layer performs attention calculation on the first seismic trace embedding vector to obtain seismic trace attention features;

[0161] The seismic trace attention features and the first seismic trace embedding vector are spliced and input into the convolution layer, the convolution layer extracts features of different scales of spliced data through the convolution blocks of different scales; wherein the spliced data is obtained by splicing the seismic trace attention features and the first seismic trace embedding vector;

[0162] Fusing features of different scales of the spliced data to obtain global features of the training seismic gather;

[0163] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic gather based on the global features of the training seismic gather and the frequency domain features of the training seismic gather extracted by the frequency domain feature stabilization module, and the inversion sub-model includes the following steps:

[0164] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic gather based on the global features of the training seismic gather and the frequency domain features of the training seismic gather extracted by the frequency domain feature stabilization module;

[0165] The frequency domain feature stabilization module includes an embedding sub-module, a frequency domain conversion sub-module, a channel learning sub-module, a frequency learning sub-module, and a time domain conversion sub-module, and the frequency domain feature stabilization module extracts the frequency domain features of the training seismic gather through the following steps:

[0166] The embedding sub-module embeds the training seismic gather to obtain a second seismic trace embedding vector;

[0167] The frequency domain conversion sub-module performs frequency domain conversion on the second seismic trace embedding vector to obtain a frequency domain vector corresponding to the second seismic trace embedding vector;

[0168] The channel learning sub-module performs linear transformation on the frequency domain vector in the channel dimension to obtain a channel vector;

[0169] The frequency learning sub-module performs linear transformation on the channel vector in the frequency dimension to obtain a frequency vector;

[0170] The time domain conversion sub-module performs time domain conversion on the frequency vector to obtain the frequency domain features of the training seismic gather;

[0171] The inversion sub-model includes a plurality of Bi-LSTM modules, and the plurality of Bi-LSTM modules are connected in series, and the training module is configured to perform bidirectional recursive operation on the training seismic gather through the plurality of Bi-LSTM modules to obtain time sequence features corresponding to the training seismic gather;

[0172] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic gather based on the global features of the training seismic gather and the frequency domain features of the training seismic gather extracted by the frequency domain feature stabilization module, and the inversion sub-model includes the following steps:

[0173] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic trace set based on the global feature of the training seismic trace set, the time sequence feature, and the frequency domain feature extracted by the frequency domain feature stabilization module.

[0174] The inversion sub-model further comprises an upsampling module, the upsampling module comprising a plurality of deconvolution sub-modules, and the training module is configured to splice the global feature and the time sequence feature to obtain a global time sequence feature.

[0175] The global time sequence feature is input into the upsampling module, and the upsampling module upsamples the global time sequence feature through a plurality of deconvolution sub-modules to obtain an expanded feature.

[0176] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic trace set based on the global feature of the training seismic trace set, the time sequence feature, and the frequency domain feature extracted by the frequency domain feature stabilization module.

[0177] The inversion sub-model outputs a seismic attribute prediction result corresponding to the training seismic trace set based on the global feature of the training seismic trace set, the time sequence feature, and the frequency domain feature extracted by the frequency domain feature stabilization module.

[0178] The initial seismic inversion model is trained based on the difference between the seismic trace prediction value and the training seismic trace set, and the difference between the seismic attribute prediction result and the real attribute label.

[0179] The total loss is constructed based on the difference between the seismic trace prediction value and the training seismic trace set, and the difference between the seismic attribute prediction result and the real attribute label.

[0180] The initial seismic inversion model is trained based on the total loss.

[0181] The exemplary embodiments of the present application also 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 is used to make the electronic device execute the method according to the embodiments of the present application when executed by the at least one processor.

[0182] The exemplary embodiments of the present application also provide a non-transitory computer readable storage medium storing a computer program, wherein the computer program is used to make the computer execute the method according to the embodiments of the present application when executed by the processor of the computer.

[0183] The exemplary embodiments of this application further provide a computer program product comprising a computer program which, when executed by a processor of a computer, is adapted to cause the computer to carry out the method according to the embodiments of this application.

[0184] Reference Figure 18 The structure block diagram of an electronic device 1800 which can be a server or a client of the present application, which is an example of a hardware device that can be applied to aspects of the present application, will now be described. The electronic device is intended to represent a wide variety of digital electronic computer devices such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. The electronic device can also represent a wide variety of mobile devices such as personal digital assistants, 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.

[0185] As Figure 18 shown, the electronic device 1800 includes a computing unit 1801 which can perform various appropriate actions and processes in accordance with a computer program stored in a read only memory (ROM) 1802 or a computer program loaded from a storage unit 1808 into a random access memory (RAM) 1803. In the RAM 1803, various programs and data required for the operation of the electronic device 1800 can also be stored. The computing unit 1801, the ROM 1802, and the RAM 1803 are connected to each other through a bus 1804. An input / output (I / O) interface 1805 is also connected to the bus 1804.

[0186] The various components in the electronic device 1800 are connected to the I / O interface 1805, including an input unit 1806, an output unit 1807, a storage unit 1808, and a communication unit 1809. The input unit 1806 can be any type of device capable of inputting information to the electronic device 1800, 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 1807 can be any type of device capable of presenting 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 1808 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 1809 allows the electronic device 1800 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 WiMax device, a cellular communication device, and / or the like.

[0187] The computing unit 1801 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1801 performs various methods and processes described above. For example, in some embodiments, any of the above-described semi-supervised intelligent seismic inversion methods based on frequency domain feature stabilization module can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1808. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 1800 via the ROM 1802 and / or the communication unit 1809. In some embodiments, the computing unit 1801 can be configured to perform any of the above-described semi-supervised intelligent seismic inversion methods based on frequency domain feature stabilization module by any other appropriate means, such as by means of firmware.

[0188] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, causes the machine to perform the functions / operations specified in the flow diagrams and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, partially on a machine and partially on a remote machine or entirely on a remote machine or server.

[0189] 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 an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an 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.

[0190] 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 to provide 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 that can be used to provide machine instructions and / or data to a programmable processor.

[0191] 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 CRT (cathode ray tube) or LCD (liquid crystal display) 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.

[0192] 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.

[0193] 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.

Claims

1. A semi-supervised intelligent seismic inversion method based on a frequency domain characteristic stabilization module, characterized in that, The method includes: Acquire target seismic trace data; The target seismic trace data is input into the target seismic inversion model to obtain the target seismic attributes output by the target seismic inversion model, wherein the target seismic attributes include target impedance field data; the target seismic inversion model is obtained by pre-training an initial seismic inversion model, which includes an inversion sub-model and a forward sub-model, and the inversion sub-model includes a frequency domain feature stabilization module; the target seismic inversion model is pre-trained through the following steps: Obtain a training seismic trace set, which contains multiple training seismic traces, and at least one of the training seismic traces has a corresponding real attribute label. The training seismic gathers are input into the initial seismic inversion model. The inversion sub-model outputs the seismic attribute prediction results corresponding to the training seismic gathers based on the training seismic gathers and the frequency domain features of the training seismic gathers extracted by the frequency domain feature stabilization module. The earthquake attribute prediction results are input into the forward sub-model, and the forward sub-model outputs the corresponding earthquake trace prediction value based on the earthquake attribute prediction results; Based on the differences between the predicted seismic trace values ​​and the training seismic trace set, and the differences between the predicted seismic attributes and the true attribute labels, the initial seismic inversion model is trained until a preset training stop condition is met, resulting in a target seismic inversion model, which includes the frequency domain feature stabilization module.

2. The method according to claim 1, characterized in that, The inversion sub-model further includes a global feature extraction module, which comprises an embedding layer, a multi-head attention layer, and a convolutional layer. The convolutional layer contains convolutional blocks of different scales. The method further includes: The training seismic trace set is input into the global feature extraction module, and the embedding layer embeds each of the training seismic trace data to obtain the first seismic trace embedding vector; The first seismic trace embedding vector is input into the multi-head attention layer, and the multi-head attention layer performs attention calculation on the first seismic trace embedding vector to obtain the seismic trace attention features. The seismic trace attention features and the first seismic trace embedding vector are concatenated and input into the convolutional layer. The convolutional layer extracts features of different scales from the concatenated data through convolutional blocks of different scales. The concatenated data is obtained by concatenating the seismic trace attention features and the first seismic trace embedding vector. The features at different scales of the stitched data are fused to obtain the global features of the training seismic gather; The inversion sub-model, based on the training seismic gathers and the frequency domain features extracted from the training seismic gathers by the frequency domain feature stabilization module, outputs the seismic attribute prediction results corresponding to the training seismic gathers, including: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the global features of the training earthquake gather and the frequency domain features extracted by the frequency domain feature stabilization module.

3. The method according to claim 2, characterized in that, The frequency domain feature stabilization module includes: an embedding submodule, a frequency domain transformation submodule, a channel learning submodule, a frequency learning submodule, and a time domain transformation submodule; the frequency domain feature stabilization module extracts the frequency domain features of the training seismic gather through the following steps: The embedding submodule embeds the training seismic traces to obtain a second seismic trace embedding vector; The frequency domain conversion submodule performs frequency domain conversion on the second seismic trace embedding vector to obtain the frequency domain vector corresponding to the second seismic trace embedding vector. The channel learning submodule performs a linear transformation of the frequency domain vector along the channel dimension to obtain the channel vector. The frequency learning submodule performs a linear transformation on the channel vector along the frequency dimension to obtain a frequency vector. The frequency vector is transformed in the time domain by the time domain transformation submodule to obtain the frequency domain features of the training seismic gather.

4. The method according to claim 3, characterized in that, The inversion sub-model includes multiple Bi-LSTM modules, and the multiple Bi-LSTM modules are connected in series. The method also includes: The training seismic gathers are subjected to bidirectional recursive operations by multiple Bi-LSTM modules to obtain the temporal features corresponding to the training seismic gathers. The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gathers based on the global features of the training earthquake gathers and the frequency domain features extracted by the frequency domain feature stabilization module, including: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the global features of the training earthquake gather, the temporal features, and the frequency domain features extracted by the frequency domain feature stabilization module.

5. The method according to claim 4, characterized in that, The inversion sub-model further includes: an upsampling module, which comprises multiple deconvolution sub-modules; the method further includes: The global features and the temporal features are concatenated to obtain the global temporal features; The global temporal features are input into the upsampling module, which upsamples the global temporal features through multiple deconvolution sub-modules to obtain extended features. The inversion sub-model outputs earthquake attribute prediction results corresponding to the training earthquake gathers based on the global features of the training earthquake gathers, the temporal features, and the frequency domain features extracted by the frequency domain feature stabilization module. These results include: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the extended features and the frequency domain features extracted by the frequency domain feature stabilization module.

6. The method according to claim 1, characterized in that, The process of training the initial seismic inversion model based on the differences between the predicted seismic trace values ​​and the training seismic trace set, and the differences between the predicted seismic attributes and the true attribute labels, includes: The total loss is constructed based on the difference between the predicted seismic trace values ​​and the training seismic trace set, and the difference between the predicted seismic attributes and the true attribute labels. The initial seismic inversion model is trained based on the total loss.

7. A semi-supervised intelligent seismic inversion device based on a frequency domain characteristic stabilization module, characterized in that, The device includes: The acquisition module is used to acquire target seismic trace data; The output module is used to input the target seismic trace data into the target seismic inversion model and obtain the target seismic attributes output by the target seismic inversion model, wherein the target seismic attributes include target impedance field data; The training module is used to pre-train a target earthquake inversion model according to the following steps, wherein the target earthquake inversion model is obtained by pre-training an initial earthquake inversion model, the initial earthquake inversion model includes an inversion sub-model and a forward sub-model, and the inversion sub-model includes a frequency domain feature stabilization module. Obtain a training seismic trace set, which contains multiple training seismic traces, and at least one of the training seismic traces has a corresponding real attribute label. The training seismic gathers are input into the initial seismic inversion model. The inversion sub-model outputs the seismic attribute prediction results corresponding to the training seismic gathers based on the training seismic gathers and the frequency domain features of the training seismic gathers extracted by the frequency domain feature stabilization module. The earthquake attribute prediction results are input into the forward sub-model, and the forward sub-model outputs the corresponding earthquake trace prediction value based on the earthquake attribute prediction results; Based on the differences between the predicted seismic trace values ​​and the training seismic trace set, and the differences between the predicted seismic attributes and the true attribute labels, the initial seismic inversion model is trained until a preset training stop condition is met, resulting in a target seismic inversion model, which includes the frequency domain feature stabilization module.

8. The apparatus according to claim 7, characterized in that, The inversion sub-model also includes a global feature extraction module, which includes an embedding layer, a multi-head attention layer, and a convolutional layer, wherein the convolutional layer contains convolutional blocks of different scales. The training module is used to input the training seismic trace set into the global feature extraction module, and the embedding layer embeds each of the training seismic trace data to obtain a first seismic trace embedding vector. The first seismic trace embedding vector is input into the multi-head attention layer, and the multi-head attention layer performs attention calculation on the first seismic trace embedding vector to obtain the seismic trace attention features. The seismic trace attention features and the first seismic trace embedding vector are concatenated and input into the convolutional layer. The convolutional layer extracts features of different scales from the concatenated data through convolutional blocks of different scales. The concatenated data is obtained by concatenating the seismic trace attention features and the first seismic trace embedding vector. The features at different scales of the stitched data are fused to obtain the global features of the training seismic gather; The inversion sub-model, based on the training seismic gathers and the frequency domain features extracted from the training seismic gathers by the frequency domain feature stabilization module, outputs the seismic attribute prediction results corresponding to the training seismic gathers, including: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the global features of the training earthquake gather and the frequency domain features extracted by the frequency domain feature stabilization module. The frequency domain feature stabilization module includes: an embedding submodule, a frequency domain transformation submodule, a channel learning submodule, a frequency learning submodule, and a time domain transformation submodule; the frequency domain feature stabilization module extracts the frequency domain features of the training seismic gather through the following steps: The embedding submodule embeds the training seismic traces to obtain a second seismic trace embedding vector; The frequency domain conversion submodule performs frequency domain conversion on the second seismic trace embedding vector to obtain the frequency domain vector corresponding to the second seismic trace embedding vector. The channel learning submodule performs a linear transformation of the frequency domain vector along the channel dimension to obtain the channel vector. The frequency learning submodule performs a linear transformation on the channel vector along the frequency dimension to obtain a frequency vector. The frequency vector is transformed in the time domain by the time domain transformation submodule to obtain the frequency domain features of the training seismic gather; The inversion sub-model includes multiple Bi-LSTM modules, which are connected in series. The training module is used to perform bidirectional recursive operations on the training seismic gathers through the multiple Bi-LSTM modules to obtain the temporal features corresponding to the training seismic gathers. The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gathers based on the global features of the training earthquake gathers and the frequency domain features extracted by the frequency domain feature stabilization module, including: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the global features of the training earthquake gather, the temporal features, and the frequency domain features extracted by the frequency domain feature stabilization module. The inversion sub-model further includes: an upsampling module, which contains multiple deconvolution sub-modules; and a training module, which is used to concatenate the global features and the temporal features to obtain global temporal features. The global temporal features are input into the upsampling module, which upsamples the global temporal features through multiple deconvolution sub-modules to obtain extended features. The inversion sub-model outputs earthquake attribute prediction results corresponding to the training earthquake gathers based on the global features of the training earthquake gathers, the temporal features, and the frequency domain features extracted by the frequency domain feature stabilization module. These results include: The inversion sub-model outputs the earthquake attribute prediction results corresponding to the training earthquake gather based on the extended features and the frequency domain features extracted by the frequency domain feature stabilization module. The process of training the initial seismic inversion model based on the differences between the predicted seismic trace values ​​and the training seismic trace set, and the differences between the predicted seismic attributes and the true attribute labels, includes: The total loss is constructed based on the difference between the predicted seismic trace values ​​and the training seismic trace set, and the difference between the predicted seismic attributes and the true attribute labels. The initial seismic inversion model is trained based on the total loss.

9. An electronic device, characterized in that, include: processor; And the memory for storing programs, The program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-6.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, in, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-6.

Images