Pre-stack and post-stack multi-attribute combined intelligent river channel identification method

Abstract

The invention discloses a pre-stack and post-stack multi-attribute combined intelligent river channel identification method, and belongs to the technical field of geophysical exploration. The method comprises the following steps: S1, analyzing pre-stack riverway response characteristics and determining sensitive attributes; s2, constructing a pre-stack and post-stack multi-attribute three-dimensional river channel sample data set; s3, constructing a pre-stack and post-stack multi-attribute combined intelligent river channel identification network model; s4, designing a loss function for pre-stack and post-stack multi-attribute combined river channel intelligent identification; s5, optimizing network parameters in the pre-stack and post-stack multi-attribute combined intelligent river channel identification network model in the step S3 based on the sample data set in the step S2 and the loss function in the step S4 to obtain an optimized intelligent river channel identification network model; and S6, carrying out three-dimensional river channel intelligent identification in an actual work area by using the river channel intelligent identification network model obtained in the S5. The method can more comprehensively capture the features of the underground river channel under complex geological conditions, improves the recognition precision of the three-dimensional river channel, and provides technical support for exploration and deployment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical exploration, more particularly to a pre-stack and post-stack multi-attribute joint river channel intelligent identification method. BACKGROUND

[0002] Fluvial facies is an important sedimentary system in continental environment, and the sand body formed is a good oil and gas reservoir. In the Mesozoic and Cenozoic oil and gas basins in eastern China, fluvial facies occupies an important proportion. For fluvial facies reservoir prediction and trap analysis, it is crucial to quickly and effectively identify the river channel. The identification of river channel sand body usually relies on P-wave stacked seismic data, and seismic attribute analysis methods such as root mean square amplitude, similarity and coherence are used to highlight the seismic response characteristics of the river channel interior or boundary by transforming the data in a specified time window. The advantage of the above method is that the physical meaning is clear. However, with the increase of data information, repeated adjustment of parameters and selection of attribute combination methods are needed in the process of attribute optimization and fusion analysis, which involves many subjective judgment factors. The reliability of river channel sand body prediction cannot be guaranteed, and the traditional analysis process still faces problems such as complicated workflow, low efficiency, difficulty in multi-dimensional data analysis, and cannot meet the needs of current efficient exploration and development.

[0003] In recent years, artificial intelligence technologies based on data-driven deep learning have developed rapidly and have achieved remarkable application results in natural image recognition, game competition and other fields. They have also received widespread attention in the fields of oil and gas exploration and development research such as waveform classification, rock hill identification, oil and gas detection, and river channel identification. In the application of river channel identification, researchers have explored the use of data-driven models to establish a quantitative relationship between the spatial occurrence of river channels and geophysical data. However, current research results are mostly based on post-stack P-wave seismic data, and in specific areas where the P-wave impedance of river sandstone and surrounding rock is similar, the reflection characteristics are not obvious, and the spatial structure of the river channel cannot be accurately described. In contrast, pre-stack seismic data provide more information, especially in reflecting the velocity difference of S-wave impedance, which is more effective. This difference is particularly evident when the reservoir contains a gas reservoir. Therefore, further method research is needed to fully exploit the multi-dimensional data information of pre-stack seismic data and improve the reliability and accuracy of the three-dimensional river channel intelligent identification network model.

[0004] The Chinese patent document with the publication number CN116609830A and the publication date of April 11, 2023 discloses a river channel sand body identification method based on sand body AVO response characteristics, which comprises the following steps: determining the prestack response characteristics of the river channel sand body by well-seismic calibration, and then determining the AVO response characteristics and AVO types of the river channel sand body through AVO forward modeling; performing prestack CRP gather processing under the quality control of sand body AVO forward modeling characteristics, so that the actual prestack gather AVO characteristics of the river channel sand body match the AVO response characteristics of the sand body forward modeling; under the guidance of the AVO response characteristics of the river channel sand body, performing dominant angle stacking on the CRP gather to obtain a seismic data volume; based on the seismic data volume, extracting the amplitude values of the fusion attribute volume by using high-density isochronous slice extraction based on global automatic volume tracking, and depicting the planar distribution of the sand body according to the amplitude values of the fusion attribute volume. The method of the invention can simultaneously consider different velocity river channel sand bodies and improve the accuracy of river channel depiction.

[0005] However, the above technical solution uses the isochronous slice of automatic tracking to identify the river channel, the three-dimensional prestack seismic attribute characteristics are not fully explored, and its essence is still based on poststack seismic data of dominant angle stacking. The gradient and intercept information of the prestack gather amplitude change and the specific geological and geophysical parameter analysis involved are not fully extracted and utilized, and specific geological and geophysical parameter analysis is not involved, so that some subtle geological information cannot be fully extracted and utilized, resulting in reduced accuracy of river channel identification. SUMMARY

[0006] In order to overcome the defects and deficiencies in the above-mentioned prior art, the present application provides a more efficient and reliable prestack-poststack multi-attribute joint river channel intelligent identification method, which can fully explore the multi-dimensional data information of prestack seismic data and provide technical support for exploration deployment.

[0007] In order to solve the problems existing in the above-mentioned prior art, the present application is realized by the following technical scheme:

[0008] A prestack-poststack multi-attribute joint river channel intelligent identification method, comprising the following steps:

[0009] S1, prestack river channel response characteristic analysis and sensitive attribute determination: based on the actual work area river channel geology and seismic data characteristics, a river channel geological model is designed, prestack seismic forward response characteristic analysis is carried out, and the river channel identification related sensitive seismic attributes under different lithology elastic parameters are determined;

[0010] S2, prestack-poststack multi-attribute three-dimensional river channel sample data set construction: using the river channel identification related sensitive seismic attributes obtained in S1, combining with the well logging data information, carrying out three-dimensional river channel fine interpretation in the work area, and simultaneously enhancing the prestack-poststack multi-attribute data and three-dimensional river channel interpretation results to obtain a sample data set;

[0011] S3, construct a pre-stack post-stack multi-attribute joint channel intelligent recognition network model;

[0012] S4, design a loss function for pre-stack post-stack multi-attribute joint channel intelligent recognition;

[0013] S5, based on the sample data set of S2 and the loss function of S4, optimize the network parameters in the pre-stack post-stack multi-attribute joint channel intelligent recognition network model in S3, and obtain an optimized channel intelligent recognition network model;

[0014] S6, using the channel intelligent recognition network model obtained in S5, carrying out three-dimensional channel intelligent recognition in actual work area.

[0015] Said S1, specifically comprising:

[0016] S11, according to the logging data and the lithology interpretation result, the geological and geophysical parameters are counted;

[0017] S12, based on the statistical geological and geophysical parameters, a channel geological model is designed, and seismic forward is carried out to obtain pre-stack gather data, combined with post-stack seismic data, analyze the characteristics of various channel pre-stack and post-stack seismic attributes, and obtain sensitive pre-stack and post-stack seismic attributes affecting channel recognition.

[0018] The geological and geophysical parameters include channel P-wave, S-wave, density elastic parameter, channel sand body shape parameter, seismic main frequency, phase, wavelet.

[0019] Said S2, specifically comprising:

[0020] S21, using sensitive pre-stack and post-stack seismic attributes, respectively from profile and along slice two directions, carrying out channel fine interpretation, combining with logging data information correction, obtaining three-dimensional channel label data;

[0021] S22, drift along the top of the layer section upwards and downwards, determine the top and bottom time of the time window; extract the original seismic data within the time window to obtain pre-stack and post-stack seismic attribute sample feature data D and channel sample label L;

[0022] S23, the pre-stack and post-stack seismic attribute sample feature data D and the channel sample label L inside the time window extraction result are enhanced;

[0023] S24, the enhanced and expanded multiple channel pre-stack and post-stack seismic attribute data are used as sample feature data, the enhanced and expanded channel interpretation result is used as sample label data, and it is randomly divided into training set, verification set and test set, as the sample data set of channel intelligent recognition model.

[0024] The top and bottom time of the time window is expressed as the following formula:

[0025] T top = T top_layer - alpha T wav (1)

[0026] T bot = T top_layer + (alpha T wav + delta T max ) (2)

[0027] Wherein, T top is the top formation time after the time window extraction, T bot is the bottom formation time after the time window extraction, alpha (alpha >= 1) is the time window scale factor, T top_layer is the top formation time of the target interval, delta T wav is the wavelet time length, delta T max is the maximum value of the time difference between the bottom and top of the target interval.

[0028] The enhancement processing includes in-layer rotation of the seismic data, simulation of the relief change, conversion of the original data into different size channels and corresponding seismic data by applying sinusoidal wave displacement in the vertical time direction, adding random noise to the seismic data or performing horizontal direction scale scaling on the seismic data.

[0029] In the S3, a multi-input channel convolutional neural network architecture is adopted to construct a river channel intelligent recognition network model for multi-attribute joint analysis, and the input of the river channel intelligent recognition network model is multi-channel three-dimensional data composed of multiple prestack and poststack seismic attributes, and the output result is three-dimensional river channel label data.

[0030] In the S4, the loss function is:

[0031] L total = alpha L labeled + (1-alpha) L unlabeled (3)

[0032] Wherein, L labeled is the loss of labeled data, L labeled = -sum (ylog (y') + (1-y) log (1-y')), y is the real label, y' is the model prediction; L unlabeled is the loss of unlabeled data, L unlabeled= -∑p(y~|x; θ)log(y^), wherein p(y~|x; θ) represents a predicted label distribution of unlabeled data given input x and model parameters θ; α is a weight parameter for balancing the contribution of labeled and unlabeled data to the total loss, and the setting of the weight takes into account the integrity of data labels and the confidence of model prediction, aiming to balance the contribution of labeled and unlabeled data to model training.

[0033] The S5 specifically comprises:

[0034] Based on the constructed sample data set of river channel intelligent identification, a loss function is used as an evaluation standard, an optimization calculation method is used, a river channel intelligent identification network weight is solved through repeated network forward propagation and backward propagation processes, a verification set is used to evaluate model performance after each training cycle, network hyperparameters are adjusted according to performance on the verification set, so that an optimal network model in the stage is determined, an independent test set is used to evaluate the generalization ability of the model after the network model is determined, if the application condition is not met, the above process is repeated to determine the final river channel intelligent identification network model.

[0035] The S6 specifically comprises:

[0036] S61, prestack and poststack seismic attribute layer flattening processing: using the top and bottom horizon information of the target layer section, layer flattening processing is performed on the prestack and poststack seismic attribute data;

[0037] S62, river channel intelligent identification of layer flattening processing results: applying the trained river channel intelligent identification network model to process the flattened data to identify the spatial distribution of the river channel;

[0038] S63, reverse layer flattening processing of river channel intelligent identification results: after identification, reverse layer flattening processing is performed, and the reverse layer flattening process is the inverse process of layer flattening, which restores the identification results to the original data space;

[0039] S64, result verification and correction: comparing the identification results after reverse layer flattening with actual geology and logging data to verify the accuracy, and checking and correcting according to exploration requirements.

[0040] Compared with the prior art, the beneficial technical effects brought by the present application are:

[0041] 1、The present application introduces prestack and poststack multi-attribute data, fully utilizes the complementarity between various prestack and poststack seismic attributes, fully excavates the three-dimensional prestack seismic attribute characteristics including sensitive elastic parameters, geological model design and seismic forward modeling, shape parameter and geophysical parameter statistics, comprehensive analysis of prestack and poststack seismic attribute characteristics and optimization of sensitive attributes, and these characteristics are more comprehensively and deeply processed, so that the accuracy and reliability of the river channel recognition are significantly improved; a prestack and poststack multi-attribute river channel sample data construction method is designed, and the sample data scale is effectively expanded through sample enhancement; on this basis, a river channel intelligent recognition network structure capable of effectively fusing various prestack and poststack seismic attributes is proposed; the river channel intelligent recognition loss function is improved for the problem of incomplete labeling of river channels; in practical application, the prestack and poststack seismic attribute results after layer flattening are used for river channel recognition, and the interference of structure on intelligent recognition results is reduced. Through the above improvements, the present application can more comprehensively capture the characteristics of the underground river channel under complex geological conditions, and is helpful to further improve the three-dimensional river channel recognition accuracy.

[0042] 2、The present application proposes a river channel intelligent recognition method fusing prestack multi-attribute data, which can more effectively identify the spatial position of the river channel when the fluid occurrence regularity of the river channel reservoir is complex compared with the traditional poststack attribute analysis.

[0043] 3、The present application proposes a prestack and poststack multi-attribute river channel sample data construction method, which optimizes the most sensitive attributes used for river channel recognition through elastic parameter sensitivity analysis, shape parameter statistical analysis and geophysical parameter statistical analysis, and constructs samples, and effectively expands the scale of the sample data set through sample enhancement technology, thereby providing rich training data for the deep learning model and improving the generalization ability of the model.

[0044] 4、The present application designs a river channel intelligent recognition network structure capable of effectively fusing various prestack and poststack seismic attributes, and especially for the common problem of incomplete labeling of river channel data, a new type of loss function is designed, so that the model can still maintain robustness when facing incomplete label data. The layer flattening technology is used to process the prestack and poststack seismic attribute data, which effectively reduces the interference of geological structure complexity on the intelligent recognition result, improves the reliability of the recognition work, especially in areas with significant structural changes.

[0045] 5、The present application adopts various data enhancement methods including rotation, vertical fluctuation twisting, adding noise and scale transformation for the enhancement processing of the time window extraction result, which can significantly improve the robustness, generalization ability and recognition accuracy of the river channel intelligent recognition model, improve the diversity of data from different angles, and make the model better cope with complex and variable geological conditions in practical application. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 Flow chart of the method of the present application;

[0047] Figure 2 Pre-stack seismic forward response characteristic analysis chart in embodiment 2 of the present application;

[0048] Figure 3 Three-dimensional seismic channel automatic interpretation network structure chart in embodiment 2 of the present application;

[0049] Figure 4 Actual data post-stack P-wave and pre-stack sensitive attribute plane chart in embodiment 3 of the present application;

[0050] Figure 5 River channel sample data enhancement processing schematic diagram in embodiment 3 of the present application; DETAILED DESCRIPTION

[0051] The technical solutions of the present application will be described clearly and completely below in combination with specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] Embodiment 1

[0053] As the most basic embodiment of the present application, the present embodiment discloses a pre-stack and post-stack multi-attribute joint river channel intelligent identification method, which comprises the following steps:

[0054] S1, pre-stack channel response characteristic analysis and sensitive attribute determination: based on the channel geology and seismic data characteristics of the actual work area, a channel geological model is designed, pre-stack seismic forward response characteristic analysis is carried out, and the channel identification related sensitive seismic attributes under different lithology elastic parameters are determined;

[0055] S2, construction of pre-stack and post-stack multi-attribute three-dimensional channel sample data set: using the channel identification related sensitive seismic attributes obtained in S1, combining with the well logging data information, three-dimensional channel fine interpretation in the work area is carried out, and the pre-stack and post-stack multi-attribute data and three-dimensional channel interpretation results are simultaneously enhanced, and the sample data set is obtained;

[0056] S3, construction of pre-stack and post-stack multi-attribute joint river channel intelligent identification network model;

[0057] S4, design of loss function for pre-stack and post-stack multi-attribute joint river channel intelligent identification;

[0058] S5, optimizing the network parameters in the prestack poststack multi-attribute joint river channel intelligent identification network model in S3 based on the sample data set in S2 and the loss function in S4, to obtain an optimized river channel intelligent identification network model;

[0059] S6, using the river channel intelligent identification network model obtained in S5 to carry out three-dimensional river channel intelligent identification in an actual work area.

[0060] In this embodiment, prestack and poststack multi-attribute data are introduced, and the complementarity between various prestack and poststack seismic attributes is fully utilized. A prestack and poststack multi-attribute river channel sample data construction method is designed, and the sample data scale is effectively expanded through sample enhancement. On this basis, a river channel intelligent identification network structure capable of effectively fusing various prestack and poststack seismic attributes is proposed. In view of the problem of incomplete labeling of river channels, the river channel intelligent identification loss function is improved. In actual application, the prestack and poststack seismic attribute results after layer flattening are used for river channel identification, which reduces the interference of structures on the intelligent identification results. Through the above improvements, this embodiment can more comprehensively capture the characteristics of underground river channels under complex geological conditions, which helps to further improve the three-dimensional river channel identification precision.

[0061] Embodiment 2

[0062] As a preferred embodiment of the present application, this embodiment discloses a prestack and poststack multi-attribute joint river channel intelligent identification method, which comprises the following steps:

[0063] S1, prestack river channel response feature analysis and sensitive attribute determination: based on the river channel geology and seismic data characteristics of an actual work area, a river channel geological model is designed, prestack seismic forward response feature analysis is carried out, and river channel identification related sensitive seismic attributes under different lithology elastic parameter conditions are determined;

[0064] According to the logging data and lithology interpretation results, the P-wave, S-wave and density elastic parameters of the river channel are counted, the shape parameters such as the thickness and width of the river channel sand body are counted, and the geophysical parameters such as the seismic main frequency, phase and wavelet are counted. Based on the above geological and geophysical parameter statistical results, a river channel geological model is designed, and seismic forward is carried out by using Zoeppritz equation set to obtain prestack gather data. Then, the characteristics of various river channel prestack and poststack seismic attributes are analyzed, and the sensitive prestack and poststack seismic attributes related to river channel identification are optimized;

[0065] For example, Figure 2 (d)- Figure 2(e) As shown in Fig. 1 (e), using the Zoeppritz equation set, the pre-stack gather is forward calculated with 35Hz Ricker wavelet, and after processing, the conventional post-stack P-wave profile, i.e. PP wave forward profile, is obtained. According to the slope and intercept attributes of the pre-stack gather, the difference attribute of the slope and intercept attributes is obtained, which is approximately the S-wave attribute, i.e. PS wave forward profile. Figure 2 (d) and Figure 2 (e) The response characteristics of different sandstones are different, and the latter has better reflection on medium-speed sandstone, which can effectively assist the identification of the channel.

[0066] As Figure 2 (a)- Figure 2 As shown in Fig. 1 (c), the designed forward geological model is a gas channel sandstone model, which includes 3 sets of strata. The top surrounding rock has a P-wave velocity of 4350m / s, a S-wave velocity of 2300m / s, and a density of 2.48g / cm3. The middle set of strata is composed of a series of non-uniform sand bodies, which have certain differences in P-wave velocity, S-wave velocity and density. The P-wave velocity decreases from 5000m / s to 4200m / s from left to right, the S-wave velocity decreases from 2850m / s to 2450m / s, and the density decreases from 2.49g / cm3 to 2.45g / cm3. The bottom stratum has a P-wave velocity of 4350m / s, a S-wave velocity of 2300m / s, and a density of 2.48g / cm3.

[0067] S2, pre-stack and post-stack multi-attribute three-dimensional channel sample data set construction: using the channel identification related sensitive seismic attributes obtained from S1, combined with well logging data information, the three-dimensional channel fine interpretation in the work area is carried out, and the pre-stack and post-stack multi-attribute data and the three-dimensional channel interpretation results are simultaneously enhanced, to obtain the sample data set;

[0068] Specifically, it includes: (1) using the optimized pre-stack and post-stack seismic attributes, the channel fine interpretation is carried out from the profile and the layer slice in two directions respectively, to obtain the three-dimensional channel label data;

[0069] (2) Time window extraction of channel sample data in the target layer. The time window bottom is obtained by drifting upwards along the top stratum and drifting downwards along the top stratum. In order to ensure that the sample data can accurately and comprehensively represent the target geological characteristics, and provide sufficient, the specific time window size requirement can be expressed as the following formula;

[0070] T top =T top_layer -αT wav (1)

[0071] T bot =T top_layer +(αT wav +ΔTmax ) (2)

[0072] wherein, T top is the top formation time after time window extraction, T bot is the bottom formation time after time window extraction, a (a≥1) is the time window scaling factor, T top_layer is the top formation time of the target interval, T wav is the wavelet time length, and ΔT max is the maximum value of the time difference between the bottom and top of the target interval;

[0073] According to the above time window extraction of the original seismic data, the prestack and poststack seismic attribute sample feature data D and the channel sample label L are obtained, at this time the two data are flattened into a cuboid data, the horizontal direction size is the same as the original data size, and the vertical direction size is 2aT wav +ΔT max , which can effectively reduce the calculation amount required for subsequent sample enhancement processing and training;

[0074] (3) Enhancement expansion of the time window extraction result of the channel sample data: the prestack and poststack seismic attribute sample feature data D and the channel sample label L inside the time window extraction result are enhanced;

[0075] (4) The enhanced and expanded multiple channel prestack and poststack seismic attribute data are taken as sample feature data, the enhanced and expanded channel interpretation result is taken as sample label data, and they are randomly split into a training set, a validation set and a test set as a sample data set for channel intelligent recognition;

[0076] S3, a channel intelligent recognition network model combined with prestack and poststack multi-attributes is constructed;

[0077] As shown in the network structure diagram of Figure 3 , the input of the channel intelligent recognition network is multi-channel three-dimensional data composed of multiple prestack and poststack seismic attributes, and the output result is three-dimensional channel label data. The multiple prestack and poststack seismic attributes include but are not limited to poststack P-wave seismic data, intercept attribute of Amplitude Versus Offset (AVO) of prestack seismic trace set amplitude and offset, AVO slope attribute and its combination attribute, etc. The network uses a convolutional neural network architecture with multiple input channels, each seismic attribute as an independent input channel, to realize effective feature extraction and information fusion.

[0078] As shown in the network structure diagram of Figure 3The network structure diagram shown in FIG. 1, the weight sharing mechanism is applied to the convolutional layer, the purpose is to improve the parameter efficiency and enhance the model learning ability of the correlation between different seismic attributes. The network structure contains three-dimensional convolutional layer, which is specially used to extract the three-dimensional spatial features of seismic data, which is crucial for accurate analysis of river channel geological structure. The encoder-decoder architecture is adopted, in which the encoder gradually reduces the spatial dimension of the data to extract features, and the decoder gradually restores the spatial dimension to reconstruct the detailed information of the river channel structure. In the decoder, the transpose convolution or up-sampling layer is used to fine recover details, to ensure that the three-dimensional river channel label data volume output finally is consistent with the size of the input data. In order to improve the model's ability to identify key spatial features of the river channel, the attention mechanism is integrated to enhance the model's ability to perceive key features. In addition, the network model uses the skip connection technology to directly transmit high-resolution features from the encoder to the decoder, so as to better preserve the spatial information. Multi-scale feature fusion strategy is also adopted to ensure that the network can use the features extracted by each layer at different scales to fully capture the river channel features;

[0079] S4, design a loss function for intelligent river channel identification of prestack and poststack multi-attribute joint;

[0080] In view of the possible incomplete or partially labeled situation of river channel interpretation results, a probability model is introduced to process the unlabeled data part. Specifically, for the labeled sample part, the conventional supervised semantic segmentation loss function is adopted, including but not limited to cross-entropy loss, to ensure the consistency between the prediction and the true label. While for the unlabeled data part, the loss function uses the confidence of the model prediction to estimate a soft label to estimate the loss of the unlabeled sample.

[0081] L labeled is the loss of labeled data, and the calculation method is:

[0082] L labeled =-∑(ylog(y')+(1-y)log(1-y'))

[0083] where y is the true label y' is the model prediction.

[0084] L unlabeled is the loss of unlabeled data, and the calculation method is:

[0085] L unlabeled =-∑p(y~∣x;θ)log(y^)

[0086] Where p(y~|x;θ) represents the distribution of predicted labels for unlabeled data given input x and model parameters θ. α is a weighting parameter used to balance the contributions of labeled and unlabeled data to the total loss. The weighting takes into account the completeness of data labels and the confidence of model predictions, aiming to balance the contributions of labeled and unlabeled data to model training.

[0087] The overall loss function is a weighted sum of the labeled data loss and the unlabeled data loss, in the following form:

[0088] L total =αL labeled +(1-α)L unlabeled (3)

[0089] By designing the loss function described above, the model can be trained using unlabeled data even when the labels are incomplete, thereby improving the performance of intelligent river identification without sacrificing model accuracy.

[0090] S5. Based on the constructed sample data training set, validation set, and test set, using the loss function as the evaluation criterion, and employing optimization calculation methods, the weights of the intelligent river identification network are solved through repeated forward and backward propagation processes. After each training cycle, the validation set is used to evaluate the model performance. Model validation includes, but is not limited to, metrics such as loss function value, precision, and recall. Based on the performance on the validation set, the network hyperparameters are adjusted. Hyperparameters include, but are not limited to, learning rate, weight decay coefficient, or network depth, thereby determining the optimal network model for the current stage. After determining the network model, an independent test set is used to evaluate the model's generalization ability. If the application conditions are not met, the above process is repeated to obtain the desired final intelligent river identification network model.

[0091] S6. Using the intelligent river identification network model obtained in S5, three-dimensional intelligent river identification is carried out in the actual work area.

[0092] In this embodiment, a prestack and poststack multi-attribute channel sample data construction method is proposed. Through elastic parameter sensitivity analysis, morphological parameter statistical analysis, and geophysical parameter statistical analysis, the most sensitive attributes for channel recognition are selected and samples are constructed. Through sample enhancement technology, the size of the sample data set is effectively expanded, providing rich training data for the deep learning model, thereby improving the generalization ability of the model. A channel intelligent recognition network structure capable of effectively fusing multiple prestack and poststack seismic attributes is designed. A new type of loss function is designed to address the common problem of incomplete labeling of channel data, enabling the model to remain robust when facing incomplete label data. The layer flattening technology is used to process prestack and poststack seismic attribute data, effectively reducing the interference of geological structure complexity on intelligent recognition results, and improving the reliability of the recognition work, especially in areas with significant structural changes.

[0093] Embodiment 3

[0094] As the best embodiment of the present application, a prestack and poststack multi-attribute joint channel intelligent recognition method includes the following steps:

[0095] S1, prestack channel response feature analysis and sensitive attribute determination: based on the channel geology and seismic data characteristics of the actual work area, a channel geological model is designed, prestack seismic forward response feature analysis is carried out, and the relevant sensitive seismic attributes for channel recognition under different lithology elastic parameter conditions are determined.

[0096] As shown in Figure 4 , the prestack and poststack attributes of the actual data are analyzed, and the sensitive prestack and poststack seismic attributes related to channel recognition are selected. Figure 4 (a) is a poststack P-wave data root mean square attribute plane, Figure 4 (b) is a prestack AVO gather slope and intercept difference (pseudo-S-wave) root mean square attribute data plane, Figure 4 (c) is a poststack P-wave data multi-channel coherence plane, Figure 4 (d) is a prestack AVO gather slope and intercept difference (pseudo-S-wave) instantaneous phase data plane. Through sensitive attribute selection, the spatial morphological characteristics of the channel are further highlighted, and the attribute extraction results are integrated into the subsequent data training and prediction process, which is beneficial to improving the channel recognition accuracy;

[0097] S2, prestack and poststack multi-attribute three-dimensional channel sample data set construction: using the channel recognition related sensitive seismic attributes obtained in S1, combined with well logging data information, three-dimensional channel fine interpretation is carried out within the work area, and prestack and poststack multi-attribute data and three-dimensional channel interpretation results are simultaneously enhanced to obtain a sample data set.

[0098] The specific process includes:

[0099] (1) Using the preferred pre-stack post-stack seismic attributes, respectively from the profile and along the slice two directions to carry out the channel fine interpretation, obtain the three-dimensional channel label data;

[0100] (2) The time window extraction of the channel sample data in the target layer. Along the top layer of the section, the time window bottom is obtained by drifting upward along the top layer of the section. In order to ensure that the sample data can accurately and comprehensively represent the target geological features and provide sufficient during the channel interpretation and label enhancement process. The specific time window size requirement can be expressed as the following formula:

[0101] T top =T top_layer -αT wav (1)

[0102] T bot =T top_layer +(αT wav +ΔT max ) (2)

[0103] Wherein, T top is the top layer time after time window extraction, T bot is the bottom layer time after time window extraction, α (α≥1) is the time window ratio factor, T top_layer is the top layer time of the target layer, T wav is the wavelet length, and ΔT max is the maximum value of the time difference between the bottom and top layer positions of the target layer.

[0104] According to the above time window extraction of the original seismic data, the pre-stack post-stack seismic attribute sample characteristic data D and the channel sample label L are obtained. At this time, the two data are flattened into cuboid data, the horizontal direction size is the same as the original data size, and the vertical direction size is 2αT wav +ΔT max , which can effectively reduce the calculation amount required for subsequent sample enhancement processing and training.

[0105] (3) Enhancement expansion of the time window extraction result of the channel sample data. The pre-stack post-stack seismic attribute sample characteristic data D and the channel sample label L inside the time window extraction result are enhanced, and the enhancement processing mode includes but is not limited to:

[0106] a. Rotation of three-dimensional seismic data R

[0107] Rotating the seismic data along the layer, the rotation operation is represented as R θ (D, L), wherein θ represents the rotation angle.

[0108] b. Vertical fluctuation twist F

[0109] The simulation of the relief of the stratum is realized by applying a series of sine wave displacements in the vertical time direction. The relief operation is represented as wherein is the phase, and A is the amplitude.

[0110] c. Adding noise N

[0111] Random noise is added to the seismic data, and the standard deviation range is set to a certain proportion of the amplitude of the data. The noise addition operation is represented as N μ,σ (D, L), wherein μ and σ are the mean and standard deviation of the noise, respectively.

[0112] d. Scale transformation S

[0113] The seismic data is scaled in the horizontal direction, and the original data is converted into different size channels and corresponding seismic data. The scale transformation operation is represented as S s (D, L), wherein s is the scaling ratio.

[0114] As Figure 5 shown, the sample feature data of the prestack and poststack seismic attributes and the channel sample label data are simultaneously enhanced, Figure 5 (a) is the difference between the pre-enhancement prestack AVO gather slope and intercept (pseudo-shear wave) root mean square attribute data plane graph, Figure 5 (b) is the channel sample label data plane graph before enhancement processing, Figure 5 (c) is the prestack AVO gather slope and intercept difference (pseudo-shear wave) root mean square attribute data plane graph after rotating 288 degrees in the azimuth angle and expanding 1.3 times in size, Figure 5 (d) is the channel sample label data plane graph after rotating 288 degrees in the azimuth angle and expanding 1.3 times in size, Figure 5 (e) is the superimposed profile graph of the seismic attribute data and the channel interpretation label before enhancement processing, Figure 5 (f) is the superimposed profile graph of the seismic attribute data and the channel interpretation label after vertical relief. Through the adjustment of the sample enhancement processing parameters, the sample data size can be effectively expanded, and more abundant geological and geophysical situations can be simulated.

[0115] (4) The enhanced and expanded various channel prestack and poststack seismic attribute data are taken as sample feature data, the enhanced and expanded channel interpretation results are taken as sample label data, and they are randomly split into a training set, a validation set and a test set, which are used as sample data sets for channel intelligent recognition;

[0116] S3, constructing a channel intelligent recognition network model combined with prestack and poststack multi-attributes;

[0117] The input of the river channel intelligent recognition network is multi-channel three-dimensional data composed of multiple prestack and poststack seismic attributes, and the output result is a three-dimensional river channel label data volume. The input of multiple prestack and poststack seismic attributes includes but is not limited to poststack P-wave seismic data, intercept attribute of amplitude versus offset (AVO) of prestack seismic trace set, AVO slope attribute and combination attribute, etc. The network uses a multi-input channel convolutional neural network architecture, with each seismic attribute as an independent input channel, to realize effective feature extraction and information fusion.

[0118] In the network, a weight sharing mechanism is applied in the convolutional layer, aiming to improve parameter efficiency and enhance the model's ability to learn the correlation features between different seismic attributes. The network structure contains a three-dimensional convolutional layer, which is specifically used to extract the three-dimensional spatial features of seismic data, which is crucial for accurately analyzing the river channel geological structure. An encoder-decoder architecture is adopted, in which the encoder gradually reduces the spatial dimension of the data to extract features, and the decoder gradually restores the spatial dimension to reconstruct the detailed information of the river channel structure. In the decoder, the transpose convolution or up-sampling layer is used to finely restore the details to ensure that the three-dimensional river channel label data volume output is consistent in size with the input data. In order to improve the model's ability to identify key spatial features of the river channel, an attention mechanism is integrated to enhance the model's perception of key features. In addition, the network model uses the skip connection technology to directly pass high-resolution features from the encoder to the decoder, to better preserve spatial information. A multi-scale feature fusion strategy is also adopted to ensure that the network can utilize features of different scales extracted by each layer to comprehensively capture river channel features;

[0119] S4, design a prestack and poststack multi-attribute joint river channel intelligent recognition loss function;

[0120] In view of the possible incomplete or partially labeled situation of the river channel interpretation result, a probability model is introduced to process the unlabeled data part. Specifically, for the labeled sample part, a conventional supervised semantic segmentation loss function is used, including but not limited to cross-entropy loss, to ensure the consistency between the prediction and the true label. For the unlabeled data part, the loss function uses the confidence of the model prediction to estimate a soft label to estimate the loss of the unlabeled sample.

[0121] L labeled is the loss of labeled data, and the calculation method is:

[0122] L labeled = -∑(ylog(y') + (1-y)log(1-y'))

[0123] where y is the true label y' is the model prediction.

[0124] Lunlabeled is the loss of unlabelled data, which is calculated as:

[0125] L unlabeled = -∑p(y~|x; θ)log(y^)

[0126] where p(y~|x; θ) represents the predicted label distribution of unlabelled data given input x and model parameters θ. α is a weight parameter to balance the contribution of labelled and unlabelled data to the total loss, which is set considering the completeness of data labels and the confidence of model prediction, aiming to balance the contribution of labelled and unlabelled data to model training.

[0127] The total loss function is the weighted sum of labelled data loss and unlabelled data loss, which is in the form of:

[0128] L total = αL labeled +(1-α)L unlabeled (3)

[0129] Through the design of the above loss function, the model can utilize unlabelled data for training under the condition of incomplete labels, thereby improving the performance of river intelligent identification without sacrificing model accuracy;

[0130] S5, optimize the prestack and poststack multi-attribute joint river intelligent identification network parameters to obtain a river intelligent identification network model with strong generalization ability and high prediction accuracy. Based on the constructed sample data training set, verification set and test set, the loss function is used as the evaluation standard, and the optimal calculation method is used to solve the river intelligent identification network weight through repeated network forward propagation and backward propagation process. After each training cycle, the verification set is used to evaluate the model performance. Model verification includes but is not limited to loss function value, precision, recall rate and other indicators. According to the performance on the verification set, the network hyperparameters are adjusted, including but not limited to learning rate, weight decay coefficient or network depth, etc., so as to determine the optimal network model in this stage. After determining the network model, the generalization ability of the model is evaluated using an independent test set. If the application conditions are not met, the above process is repeated to obtain the expected final river intelligent identification network model.

[0131] S6, using the river intelligent identification network model obtained in S5, three-dimensional river intelligent identification is carried out in actual work area. In actual application, in order to reduce the influence of stratigraphic structure change, first, layer flattening processing is performed, then the river intelligent identification network model is applied, and finally, reverse layer flattening is performed to restore the identification result to the original data space. The specific process is as follows:

[0132] (1) Pre-stack and post-stack seismic attribute layer flattening processing. Using the top and bottom horizon information of the target layer, layer flattening processing is performed on pre-stack and post-stack seismic attribute data to eliminate the influence of structure such as stratigraphic tilt and fold, and simplify the task of channel identification.

[0133] (2) Intelligent channel identification of layer flattening processing results. The trained intelligent channel identification network model is applied to process the flattened data to identify the spatial distribution of the channel.

[0134] (3) Reverse layer flattening processing of intelligent channel identification results. After identification, reverse layer flattening processing is performed. The reverse layer flattening process is the inverse process of layer flattening, which restores the identification results to the original data space.

[0135] (4) Result verification and correction. The identification results after reverse layer flattening are compared with the actual geology and logging data to verify their accuracy, and are checked and corrected according to the exploration requirements.

[0136] In this embodiment, for the enhancement processing of the time window extraction results, multiple data enhancement methods are used, including rotation, vertical fluctuation twisting, adding noise and scale transformation, which can significantly improve the robustness, generalization ability and identification accuracy of the intelligent channel identification model, and improve the diversity of data from different angles, so that the model can better cope with complex and variable geological conditions in actual application.

Claims

1. A pre-stack post-stack multi-attribute joint channel intelligent identification method, characterized in that, The method comprises the following steps: S1, pre-stack channel response characteristic analysis and sensitive attribute determination: based on the actual work area channel geology and seismic data characteristics, a channel geological model is designed, pre-stack seismic forward response characteristic analysis is carried out, and the channel recognition related sensitive seismic attributes under different lithology elastic parameters are determined; S2, construction of pre-stack and post-stack multi-attribute three-dimensional channel sample data set: using the channel recognition related sensitive seismic attributes obtained in S1, combining with the well logging data information, three-dimensional channel fine interpretation is carried out in the work area, and the pre-stack and post-stack multi-attribute data and the three-dimensional channel interpretation results are simultaneously enhanced, and the sample data set is obtained; S3, construction of pre-stack and post-stack multi-attribute joint channel intelligent recognition network model; S4, design of loss function for pre-stack and post-stack multi-attribute joint channel intelligent recognition; S5, based on the sample data set of S2 and the loss function of S4, the network parameters in the pre-stack and post-stack multi-attribute joint channel intelligent recognition network model in S3 are optimized, and the optimized channel intelligent recognition network model is obtained; S6, using the channel intelligent recognition network model obtained in S5, three-dimensional channel intelligent recognition is carried out in the actual work area.

2. The method according to claim 1, characterized in that: The S1 specifically comprises: S11, according to the well logging data and lithology interpretation results, the geological and geophysical parameters are counted; S12, based on the counted geological and geophysical parameters, a channel geological model is designed, and seismic forward is carried out to obtain pre-stack gather data, combined with post-stack seismic data, the characteristics of various channel pre-stack and post-stack seismic attributes are analyzed, and the sensitive pre-stack and post-stack seismic attributes affecting channel recognition are obtained.

3. The method according to claim 1, characterized in that: The geological and geophysical parameters include channel P-wave, S-wave, density elastic parameters, channel sand body shape parameters, seismic main frequency, phase, wavelet.

4. The method according to claim 1, characterized in that: The S2 specifically comprises: S21, using the sensitive pre-stack and post-stack seismic attributes, fine channel interpretation is carried out from profile and layer slice directions respectively, combined with well logging data information correction, three-dimensional channel label data is obtained; S22, the top and bottom times of the time window are determined by drifting along the top stratum of the layer section upward and downward; the original seismic data in the time window range is extracted to obtain pre-stack and post-stack seismic attribute sample feature data D and channel sample label L; S23, the pre-stack and post-stack seismic attribute sample feature data D and the channel sample label L in the time window extraction result are enhanced; S24, the enhanced and expanded multiple channel pre-stack and post-stack seismic attribute data are used as sample feature data, the enhanced and expanded channel interpretation results are used as sample label data, and they are randomly divided into training set, verification set and test set as sample data set of channel intelligent recognition model.

5. The method according to claim 4, characterized in that: The top and bottom times of the time window are expressed as the following formula: T top = T top_layer - αT wav (1) T bot = T top_layer + (aT wav + ΔT max ) (2) wherein T top is the top formation time after time window extraction, T bot is the bottom formation time after time window extraction, α (α≥1) is a time window scaling factor, T top_layer is the top formation time of the target interval, T wav is the wavelet time length, ΔT max is the maximum value of the time difference between the bottom and top of the target interval.

6. The method according to claim 4, characterized in that: The enhancement processing includes rotating the seismic data in the layer, simulating the fluctuation change of the stratum, applying sinusoidal wave displacement in the vertical time direction, adding random noise in the seismic data or scaling the seismic data in the horizontal direction, and converting the original data into different size channels and corresponding seismic data.

7. The method according to claim 1, wherein the method is characterized in that: In the S3, a multi-input channel convolutional neural network architecture is adopted to construct a river channel intelligent recognition network model for multi-attribute joint analysis, the input of the river channel intelligent recognition network model is multi-channel three-dimensional data composed of multiple prestack and poststack seismic attributes, and the output result is a three-dimensional river channel label data body.

8. The method according to claim 1, characterized in that: In the S4, the loss function is: L total = aL labeled + (1 - a)L unlabeled (3) where L labeled is the loss of labeled data, L labeled = -∑(ylog(y') + (1 - y)log(1 - y')), y is the real label, y' is the model prediction; L unlabeled is the loss of unlabeled data, L unlabeled = -∑p(y~|x; θ)log(y^), where p(y~|x; θ) represents the predicted label distribution of unlabeled data given input x and model parameters θ; α is a weight parameter to balance the contribution of labeled and unlabeled data to the total loss, the setting of the weight takes into account the integrity of the data label and the confidence of the model prediction, aiming to balance the contribution of labeled and unlabeled data to the model training.

9. The method according to claim 1, characterized in that: The S5 specifically includes: Based on the constructed sample data set of river channel intelligent recognition, the loss function is used as an evaluation standard, an optimal calculation method is used to solve the river channel intelligent recognition network weight through repeated network forward propagation and backward propagation processes, a verification set is used to evaluate the model performance after each training cycle, the network hyperparameters are adjusted according to the performance on the verification set, so as to determine the optimal network model in the stage, an independent test set is used to evaluate the generalization ability of the model after the network model is determined, if the application condition is not met, the above process is repeated to determine the final river channel intelligent recognition network model.

10. The method according to claim 1, characterized in that: The S6 specifically includes: S61, prestack and poststack seismic attribute layer flattening processing: layer flattening processing is performed on prestack and poststack seismic attribute data by using the top and bottom horizon information of the target layer section; S62, river channel intelligent recognition of layer flattening processing result: the trained river channel intelligent recognition network model is applied to process the flattened data to identify the spatial distribution of the river channel; S63, reverse layer flattening processing of river channel intelligent recognition result: after the identification is completed, reverse layer flattening processing is performed, the reverse layer flattening process is the inverse process of layer flattening, and the identification result is restored to the original data space; S64, result verification and correction: the identification result after the reverse layer flattening is compared with the actual geology and logging data to verify the accuracy, and the verification and correction are performed according to the exploration requirements.

Images