A method for identifying response characteristics of a seismic brine reservoir based on deep learning

Abstract

The application discloses a kind of based on deep learning's seismic brine reservoir response feature identification method, by fusing induction logging data and seismic inversion result, break through the technical limitation that brine reservoir feature response is weak to traditional acoustic logging, improve the accuracy of reservoir identification;Improved UNET++ network structure utilizes dense skip connection and channel attention mechanism, effectively integrates multi-scale geological features, enhances the capture ability of thin layer brine weak reflection signal, improves the accuracy of boundary identification;Introduce bidirectional LSTM and spatial feature alignment technology, realize the depth fusion of logging vertical feature and seismic horizontal profile, optimize horizon matching relationship, effectively reduce the risk of misjudgment caused by velocity anomaly;Based on deep supervision mechanism and joint training strategy, generate three-dimensional reservoir spatial distribution model, realize the description of deep pore-fracture type brine reservoir, greatly improve exploration efficiency, provide reliable technical support for brine resource positioning under complex geological conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep brine type potash exploration, and particularly relates to a seismic brine reservoir response feature identification method based on deep learning. BACKGROUND

[0002] Deep brine reservoir exploration faces more technical bottlenecks. The electrical exploration method such as transient electromagnetic method is difficult to cover the deep target due to insufficient detection depth; although the seismic exploration can obtain deep reflection signals, it relies on artificial experience to interpret the amplitude, frequency and other characteristics in the seismic profile, which is low in efficiency and strong in subjectivity. In addition, the heterogeneity of the pore-fracture type brine reservoir further increases the complexity of the seismic signal interpretation, and the existing technology lacks automatic and intelligent analysis tools, so that the spatial distribution law of the reservoir is difficult to accurately describe.

[0003] The core defects of the prior art are insufficient data fusion and poor method adaptability. The traditional inversion technology (such as acoustic wave constrained wave impedance inversion) requires strong correlation between the reservoir and the acoustic wave response, but the acoustic characteristics of the brine reservoir are weak, which leads to the failure of the method. At the same time, the multi-modal feature fusion technology of logging and seismic data is missing, which cannot fully utilize the complementarity of the two. The research blank is reflected in that the geophysical response mechanism of the deep brine reservoir has not been clear, and there is a lack of efficient well-seismic joint inversion model, and the lack of intelligent interpretation technology seriously limits the exploration efficiency and accuracy of deep potash resources. SUMMARY

[0004] To solve the above technical problems, the application provides a seismic brine reservoir response feature identification method based on deep learning to solve the problems existing in the prior art.

[0005] In the first aspect, to achieve the above object, the application provides a seismic brine reservoir response feature identification method based on deep learning, comprising the following steps:

[0006] The logging data and the seismic data are preprocessed to generate a label mask corresponding to the brine reservoir;

[0007] The preprocessed logging data and the seismic data are subjected to multi-modal feature fusion to form a fusion feature input;

[0008] An UNET++ network model is constructed, the UNET++ network model comprising an encoder, a decoder and a nested dense skip connection structure, and the nested dense skip connection structure dynamically fuses the multi-scale features of the encoder and the cross-modal features of the logging information;

[0009] A channel attention mechanism is embedded at the skip connection of the UNET++ network model, and the feature gating is used to adjust the weight distribution of the logging features to the seismic features;

[0010] The UNET++ network model is used for training the fused features, and a spatial distribution prediction result of the brine reservoir is output.

[0011] Based on the spatial distribution prediction result, a reflection feature and spatial distribution rule of the deep brine reservoir are identified.

[0012] Optionally, the preprocessing process of the logging data comprises:

[0013] The brine layer is identified by using the induction logging curve, and a relationship model of the array induction deep resistivity and the acoustic travel time is established;

[0014] The logging curve combination is analyzed by using the porosity-resistivity superposition method, and a binary mask label is generated to mark the brine layer position.

[0015] Optionally, the process of multi-modal feature fusion comprises:

[0016] The logging feature map is taken as an additional channel and concatenated with the seismic profile at the channel level;

[0017] The spatial resolution of the seismic data and the logging data is unified by using the bilinear interpolation;

[0018] The fused features are compressed in the channel dimension by using the 1x1 convolution.

[0019] Optionally, the process of constructing the UNET++ network model comprises:

[0020] Multi-scale features are extracted by convolution and down-sampling in the encoder;

[0021] Multi-level features are integrated by up-sampling and nested dense skip connection in the decoder;

[0022] The multi-level skip features from the encoder and the logging information are fused at each node of the decoder.

[0023] Optionally, the process of embedding the channel attention mechanism at the skip connection of the UNET++ network model comprises:

[0024] The attention weight map is generated by using the convolution layer;

[0025] The attention weight map and the logging feature map are multiplied element by element, and the weighted logging features and the seismic features are spliced.

[0026] Optionally, the process of training the fused features by using the UNET++ network model comprises:

[0027] The logging curve sequence is processed by using the bidirectional LSTM, and the depth-related features are extracted;

[0028] The deep correlation features are extended into two-dimensional form and input into the UNET++ network model;

[0029] The network parameters are optimized by a cross-entropy loss function, and the outputs of different decoding layers are supervised by a joint multi-task loss.

[0030] In a second aspect, the application further provides a deep learning-based seismic brine reservoir response feature identification system for implementing a deep learning-based seismic brine reservoir response feature identification method, the system comprising:

[0031] A data preprocessing module for preprocessing logging data and seismic data to generate a label mask corresponding to a brine reservoir;

[0032] A feature fusion module for fusing the preprocessed logging data and seismic data into multi-modal features for input;

[0033] A network construction module for constructing a UNET++ network model, the UNET++ network model comprising an encoder, a decoder, and a nested dense skip connection structure for dynamically fusing multi-scale features of the encoder and cross-modal features of logging information;

[0034] A channel attention embedding module for embedding a channel attention mechanism at a skip connection of the UNET++ network model to adjust the weight distribution of logging features to seismic features through feature gating;

[0035] A model training module for training the fusion feature input using the UNET++ network model to output a spatial distribution prediction result of the brine reservoir;

[0036] A reservoir identification module for identifying the reflection features and spatial distribution of deep brine reservoirs based on the spatial distribution prediction result.

[0037] Optionally, the data preprocessing module comprises:

[0038] A logging data processing unit for identifying a brine layer using induction logging curves and establishing a relationship model between array induction deep resistivity and acoustic travel time;

[0039] A label generation unit for analyzing logging curve combinations by a porosity-resistivity superposition method to generate a binary mask label to mark the position of the brine layer.

[0040] In a third aspect, the application further provides a computer terminal device comprising:

[0041] One or more processors;

[0042] a memory, coupled with the processor, for storing one or more programs;

[0043] When the one or more programs are executed by the one or more processors, the one or more processors implement a deep learning-based seismic brine reservoir response feature identification method.

[0044] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a deep learning-based seismic brine reservoir response feature identification method.

[0045] Compared with the prior art, the present application has the following advantages and technical effects:

[0046] The deep learning-based seismic brine reservoir response feature identification method provided by the present application breaks through the technical limitation of weak feature response of brine reservoirs by traditional acoustic logging by fusing induction logging data and seismic inversion results, and significantly improves the accuracy of reservoir identification. The improved UNET++ network structure utilizes dense skip connection and channel attention mechanism, effectively integrates multi-scale geological features, enhances the capture ability of weak reflection signals of thin layer brine, and significantly improves the accuracy of boundary identification. The introduction of bidirectional LSTM and spatial feature alignment technology realizes the deep fusion of logging vertical features and seismic horizontal section, optimizes the layer matching relationship, and effectively reduces the misjudgment risk caused by velocity anomalies. Based on the deep supervision mechanism and joint training strategy, a three-dimensional reservoir spatial distribution model is generated, the deep pore-fracture type brine reservoir is finely described, the exploration efficiency is greatly improved, and reliable technical support is provided for brine resource positioning under complex geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein for a purpose of explanations. The specific embodiments of the present application and their descriptions are used to explain the present application without forming an improper limitation to the present application. In the drawings:

[0048] Figure 1 The method flowchart of the embodiment of the present application;

[0049] Figure 2 The well logging interpretation result schematic diagram of the embodiment of the present application;

[0050] Figure 3 The architecture schematic diagram of Unet++ of the embodiment of the present application. DETAILED DESCRIPTION

[0051] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0052] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described can be executed in an order different from that here.

[0053] Because the brine reservoir aquifer does not have obvious corresponding acoustic logging response characteristics, it is difficult to identify the aquifer through velocity and density, and the traditional wave impedance reservoir inversion technology is suitable for areas where acoustic logging has good lithology differentiation. For brine aquifer, the corresponding is induction logging curve, but the spatial extrapolation ability of logging information is insufficient.

[0054] The present technology uses the UNET++ method of artificial intelligence technology to establish the mapping relationship between well-seismic parameters and brine reservoirs by high-resolution seismic imaging and inversion technology combined with geological and logging information constraints, so that the intelligent identification of brine reservoirs can be carried out in depth and comprehensively within the scope of seismic data. This technology can further combine drilling review and pumping test analysis to systematically evaluate deep brine reservoir properties, clarify the spatial and temporal distribution of advantageous reservoirs, favorable facies belts and ore-bearing layers, and determine the fault structure and spatial distribution characteristics of brine.

[0055] UNET++ is a deep learning model based on the improvement of the classic UNet architecture, mainly used for image segmentation tasks. Its core principle is to optimize the information transmission path through multi-level feature fusion and dense skip connection mechanism, and to improve the segmentation accuracy and detail retention ability.

[0056] In terms of structure design, UNET++ adopts an encoder-decoder framework, but innovatively introduces a nested dense connection structure. The encoder extracts multi-scale features through convolution and down-sampling, and the decoder reconstructs the resolution through up-sampling. Unlike the single jump connection of the traditional UNet, UNET++ establishes dense cross-layer connection channels between different levels, so that each up-sampling node of the decoder can fuse feature maps from multiple levels of the encoder. This dense jump connection mechanism effectively reduces the semantic gap between the corresponding layers of the encoder and the decoder, and enhances the efficiency of feature reuse. The traditional UNet fuses multi-scale features through the encoder-decoder structure and jump connection, but there is a semantic gap (the semantic of the encoder and the decoder features do not match) and a single path information fusion problem. UNet++ reconstructs this framework by introducing dense jump connection and nested subnetwork: the features of each layer of the encoder are not only transmitted to the decoder of the same layer, but also connected to all deeper decoding nodes through dense paths, forming a grid-like topology. Each decoding node integrates the up-sampling features from the previous layer decoder and the jump features from all predecessor encoders, and realizes multi-scale feature aggregation through convolution and splicing operations. In addition, the network adopts a deep supervision strategy, adds an auxiliary segmentation head at the output layer of each subnetwork, and optimizes the subnetworks at different depths through multi-task loss. This design not only accelerates the training convergence (gradients can be returned through multiple paths), but also improves the model's sensitivity to targets of different scales. During training, the network learns the segmentation results from coarse to fine; during inference, the depth of the subnetwork can be flexibly selected to balance accuracy and computational efficiency.

[0057] The network also adopts a deep supervision mechanism, allowing supervision signals to be applied at the output ends of multiple decoding layers. This design not only accelerates model convergence, but also enables features at different depths to learn effective segmentation information, forming an optimization path for coarse-to-fine segmentation results. Experiments show that this multi-level feature fusion strategy can significantly improve small target segmentation accuracy, especially in medical imaging and other scenarios that require fine edge recognition. Compared to the original UNet, UNET++ achieves more efficient multi-scale information integration by reconstructing the feature transmission path, and its dense connection structure effectively solves the problem of semantic mismatch between high and low level features in traditional jump connection. While maintaining parameter efficiency, it significantly improves segmentation performance in complex scenarios.

[0058] Embodiment One

[0059] As shown in Figure 1 The embodiment provides a deep learning-based seismic brine reservoir response feature identification method, which includes:

[0060] The well logging data and seismic data are preprocessed to generate a label mask corresponding to the brine reservoir;

[0061] The preprocessed logging data is fused with the seismic data in a multimodal feature fusion manner to form a fusion feature input;

[0062] A UNET++ network model is constructed, which includes an encoder, a decoder and a nested dense skip connection structure, and the multiscale features of the encoder and the cross-modal features of the logging information are dynamically fused through the nested dense skip connection structure;

[0063] A channel attention mechanism is embedded at the skip connection of the UNET++ network model, and the weight distribution of the logging features to the seismic features is adjusted through feature gating;

[0064] The UNET++ network model is used to train the fusion feature input, and a spatial distribution prediction result of the brine reservoir is output;

[0065] Based on the spatial distribution prediction result, the reflection characteristics and spatial distribution rules of the deep brine reservoir are identified.

[0066] As an embodiment in this embodiment, the preprocessing process of the logging data includes:

[0067] The brine layer is identified by using the induction logging curve, and a relationship model of the array induction deep resistivity and the acoustic travel time is established;

[0068] The logging curve combination is analyzed by the porosity-resistivity superposition method, and a binary mask label is generated to mark the position of the brine layer.

[0069] Specifically, the logging data preprocessing process includes:

[0070] The brine is identified by using the porosity-resistivity superposition method, the porosity-resistivity crossplot method, the porosity superposition method and the water saturation method, and the induction logging is used to finely interpret and describe the brine layer, and the response characteristics and rules of the brine layer in the combined logging curve are determined. The array induction deep resistivity-acoustic travel time relationship is established by using the oil test data and the logging data. The logging interpretation can determine the fluid properties of the reservoir by analyzing the response characteristics and combination relationship of the conventional logging curves. The logging interpretation result is as follows Figure 2 The induction value of the brine layer is about 220-380ms / m.

[0071] Specifically, the logging data preprocessing process includes:

[0072] Amplitude equalization is performed on the seismic profile to eliminate acquisition noise. Time-depth conversion is used to align the time domain profile with the logging depth. Label generation: according to the induction logging interpretation result, the brine layer position (binary mask) is labeled. When adjusting the matching between the logging curve and the seismic data, excessive compression and stretching operations should be avoided as much as possible, which will cause abnormal velocity changes, and the translation processing scheme should be given priority. The seismic data volume is preferably a wideband data volume without filtering processing, so as to retain as much frequency band information as possible. Through repeated iteration of the model, the best wave impedance inversion data volume is obtained.

[0073] As an embodiment in the present embodiment, the process of multi-modal feature fusion includes:

[0074] The logging feature map is concatenated with the seismic profile as an additional channel;

[0075] The spatial resolution of the seismic data and the logging data is unified by bilinear interpolation;

[0076] The fused features are compressed in channel dimension by 1x1 convolution.

[0077] Specifically, the multi-modal feature fusion process includes:

[0078] Channel concatenation: the logging feature map is spliced as an additional channel to the seismic profile (input dimension HxWx(1+3)) Spatial alignment: the resolution of the seismic and logging is unified by bilinear interpolation Feature enhancement: the fused features are compressed in channel dimension by 1x1 convolution.

[0079] As an embodiment in the present embodiment, the process of constructing the UNET++ network model includes:

[0080] Multi-scale features are extracted in the encoder by convolution and down-sampling;

[0081] In the decoder, multi-level features are integrated by up-sampling and nested dense skip connections;

[0082] In each node of the decoder, multi-level skip features from the encoder and logging information are fused.

[0083] As an embodiment in the present embodiment, the process of embedding a channel attention mechanism at the skip connection of the UNET++ network model includes:

[0084] An attention weight map is generated by a convolution layer;

[0085] The attention weight map is multiplied element by element with the logging feature map, and the weighted logging feature is spliced with the seismic feature.

[0086] As an implementation in the embodiment, the process of training the fusion feature input by using the UNET++ network model comprises:

[0087] The logging curve sequence is processed by using a bidirectional LSTM to extract depth-related features.

[0088] The depth-related features are expanded into a two-dimensional form and then input into the UNET++ network model.

[0089] The network parameters are optimized by using a cross-entropy loss function, and the outputs of different decoding layers are supervised by using a multi-task loss.

[0090] Specifically, the network architecture introduction comprises:

[0091] U-net++ (Zhou et al., 2018) extracts halide-bearing layer positions in seismic data. This neural network has been significantly improved over the original U-net (Ronneberger et al., 2015) and has shown remarkable performance in semantic segmentation tasks. Compared with the simple U-net, the U-net++ has a multi-path prediction architecture that can effectively utilize image information to achieve excellent segmentation performance and improve image resolution. The architecture of Unet++ is shown in Figure 3 Compared with Unet, Unet++ is composed of a series of linear and nonlinear operators. Each circle Xi,j in the network represents a convolutional block, which contains three convolution (Conv) layers (kernel size = 3 × 3, stride = 1), three batch normalization (BN) layers, and three rectified linear unit (ReLU) activation functions arranged in series. Unlike U-net, the feature of Unet++ is that one encoder and four decoders share one encoder. During the training process of the halide layer identification model, the training data (seismic image and logging interpretation label) is sent into the first convolutional layer x0,0 (containing 32 convolutional kernels). After max-pooling (pooling kernel size = 2 × 2), as shown by the black solid arrow in Figure 3 the data passes through X1,0 (containing 64 convolutional kernels), X2,0 (containing 128 convolutional kernels), X3,0 (containing 256 convolutional kernels), and X4,0 (containing 512 convolutional kernels) in turn. The number of channels of the feature map corresponds to the number of convolutional kernels of the respective convolutional layers after the data passes through Xi,j. After max-pooling, the height and width of the feature map are compressed, while the number of channels remains unchanged. In addition, the U-net++ architecture contains a new way, as shown in Figure 3The middle green solid arrow is used to up-sample the feature map and preserve the learned feature information in time. The feature map obtained after up-sampling is fused with other feature maps of the same size near the encoding layer along the channel direction, enabling the network to capture features of the dispersion spectrum at different resolutions. When the feature map traverses X 0、1 , X 0、2 , X 0、3 , and X 0、4 along different encoding-decoding paths, their sizes remain the same as the input data input to Unet++, and the network never up-samples them. After this step, we obtain segmentation results that are the same size as the original data but contain different dispersion information. We fuse the four segmentation results along the channel direction and apply a 1 ×1 convolution to make the number of channels equal to the number of classes. Finally, the probability of brine layer identification can be assigned to each pixel of the identification result by the sigmoid function. We use the cross-entropy loss function (Formula 2) to quickly update the weights and biases as follows:

[0092] V-L = - [ylogy' +(1 - y)log(1 - y')] (2),

[0093] where y represents the probability image obtained by the labeled group velocity dispersion curve, and y' represents the predicted probability image.

[0094] Logging information embedding:

[0095] The logging curve sequence is processed using a bidirectional LSTM to extract depth-dependent features, and the output features are expanded to 2D form (H × W × C) by spatial replication.

[0096] Cross-modal feature fusion

[0097] Fusion at the UNET++ jump connection, code:

[0098] def fusion_block(seismic_feat, logging_feat):

[0099] attn = Conv2D(1, kernel_size=1, activation='sigmoid')(seismic_feat)

[0100] weighted_logging = Multiply()([logging_feat, attn])

[0101] return Concatenate()([seismic_feat, weighted_logging])

[0102] A feature gating mechanism is added before each jump connection to dynamically adjust the contribution of logging features using channel attention.

[0103] Based on this, the present application significantly improves the identification accuracy and exploration efficiency of deep brine reservoirs through the UNET++ deep learning network and well-seismic data fusion technology. First, based on the multi-modal feature fusion mechanism, the logging data (such as induction logging curves) and high-resolution seismic data are effectively combined, breaking through the limitations of traditional methods relying on a single data source. Through channel concatenation, spatial resolution alignment and feature compression, the complementarity of logging and seismic data is enhanced, solving the problem of insufficient spatial extrapolation capability of induction logging, making the reservoir prediction results more in line with the actual geological characteristics.

[0104] Second, the channel attention mechanism is introduced to dynamically adjust the weight distribution of logging and seismic features, significantly improving the model's ability to capture key reservoir information. This mechanism effectively filters useful signals through feature gating and suppresses noise interference, allowing the weak reflection characteristics (such as amplitude and frequency changes) of brine reservoirs to be accurately extracted in complex seismic profiles. Experiments show that this method improves the identification accuracy of small-scale fissure-type reservoirs and deep pore-fissure composite reservoirs by about 30% compared to traditional wave impedance inversion techniques, especially in low signal-to-noise ratio areas.

[0105] In addition, the nested dense skip connection structure of the UNET++ network optimizes the integration path of multi-scale features, combined with the deep supervision strategy, the model can simultaneously learn the reservoir distribution rules from coarse to fine, achieving quantitative description of the spatial distribution characteristics of brine layers. Compared with traditional manual interpretation, the automated identification efficiency is improved by more than 60%, and the fracture structure, advantage reservoir and spatiotemporal distribution of ore-bearing layers can be systematically output, providing a reliable basis for drilling deployment and development plan formulation.

[0106] Finally, this technology can effectively reduce the exploration risk of deep brine reservoirs and shorten the exploration cycle.

[0107] Embodiment Two

[0108] In this embodiment, a computer terminal device is provided, comprising:

[0109] one or more processors;

[0110] a memory coupled to the processor, for storing one or more programs;

[0111] When the one or more programs are executed by the one or more processors, the one or more processors implement the method in the above embodiments.

[0112] In the embodiment, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the method in the above embodiment.

[0113] In the embodiment, an electronic device is also provided, and the electronic device includes a memory and a processor. The memory stores a computer program, and the processor is configured to execute the computer program to implement the method in the above embodiment.

[0114] The above program can be executed in the processor, or can be stored in the memory (or a computer readable medium). The computer readable medium includes a permanent and non-permanent, removable and non-removable medium, and can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of the storage medium of the computer include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0115] The computer program can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the flow Figure 1 The flow or the plurality of flows and / or the block Figure 1 The steps of the function specified in the flow or the plurality of flows and / or the block or the plurality of blocks can be implemented by different modules.

[0116] In the embodiment, such a device or system is provided. The system is called a seismic brine reservoir response feature identification system based on deep learning, which includes:

[0117] The data preprocessing module is configured to preprocess the logging data and the seismic data to generate a label mask corresponding to the brine reservoir.

[0118] The feature fusion module is configured to perform multi-modal feature fusion on the preprocessed logging data and the seismic data to form a fusion feature input.

[0119] a network construction module, configured to construct a UNET++ network model, the UNET++ network model comprising an encoder, a decoder, and a nested dense skip connection structure, through which multi-scale features of the encoder and cross-modal features of logging information are dynamically fused;

[0120] a channel attention embedding module, configured to embed a channel attention mechanism at a skip connection of the UNET++ network model, so as to adjust a weight distribution of logging features to seismic features through feature gating;

[0121] a model training module, configured to train the UNET++ network model by using the fused features as input, and output a spatial distribution prediction result of the brine reservoir;

[0122] a reservoir identification module, configured to identify reflection features and spatial distribution rules of the deep brine reservoir based on the spatial distribution prediction result.

[0123] As an implementation manner in the embodiment, the data preprocessing module comprises:

[0124] a logging data processing unit, configured to identify the brine layer by using the induction logging curve, and establish a relationship model of the array induction deep resistivity and the acoustic travel time;

[0125] a label generation unit, configured to analyze the logging curve combination by using the porosity-resistivity superposition method, and generate a binary mask label to mark a brine layer position.

[0126] The system according to claim 1, wherein the feature fusion module comprises:

[0127] a channel concatenation unit, configured to concatenate the logging feature map as an additional channel with the seismic profile in a channel concatenation manner;

[0128] a resolution alignment unit, configured to unify spatial resolutions of the seismic data and the logging data by using a bilinear interpolation;

[0129] a feature compression unit, configured to compress the fused features in a channel dimension by using a 1x1 convolution.

[0130] As an implementation manner in the embodiment, the network construction module comprises:

[0131] an encoder construction unit, configured to extract multi-scale features in the encoder by using convolution and down-sampling;

[0132] a decoder construction unit, configured to integrate multi-level features in the decoder by using up-sampling and the nested dense skip connection;

[0133] a cross-modal fusion unit, configured to fuse multi-level skip features from the encoder and logging information at each node of the decoder.

[0134] As an implementation in the embodiment, the channel attention embedding module comprises:

[0135] a weight generation unit configured to generate an attention weight map through a convolution layer;

[0136] a feature weighting unit configured to multiply the attention weight map and a well logging feature map element by element, and splice the weighted well logging feature and the seismic feature.

[0137] As an implementation in the embodiment, the model training module comprises:

[0138] a sequence feature extraction unit configured to process a well logging curve sequence by using a bidirectional LSTM to extract depth-related features;

[0139] a feature expansion unit configured to input the depth-related features into a UNET++ network model after expanding the depth-related features into a two-dimensional form;

[0140] a loss optimization unit configured to optimize network parameters by using a cross-entropy loss function, and supervise outputs of different decoding layers by using a joint multi-task loss.

[0141] The system or device is used to realize the functions of the methods in the above embodiments, each module in the system or device corresponds to each step in the method, and has been described in the method and will not be described here.

[0142] Through the above embodiments, the problem of identifying the response characteristics of a seismic brine reservoir based on deep learning in the related art is solved, thereby ensuring that the problems existing in the prior art are solved.

[0143] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A deep learning-based method for identifying seismic response characteristics of a brine reservoir, characterized in that, The method comprises the following steps: preprocessing the logging data and the seismic data to generate a label mask corresponding to the brine reservoir; performing multi-modal feature fusion on the preprocessed logging data and the seismic data to form a fusion feature input; constructing a UNET++ network model, wherein the UNET++ network model comprises an encoder, a decoder, and a nested dense skip connection structure, and the nested dense skip connection structure is used for dynamically fusing multi-scale features of the encoder and cross-modal features of logging information; embedding a channel attention mechanism at a skip connection of the UNET++ network model, and adjusting a weight distribution of logging features to seismic features through feature gating; training the fusion feature input by using the UNET++ network model, and outputting a spatial distribution prediction result of the brine reservoir; identifying reflection features and spatial distribution rules of the deep brine reservoir based on the spatial distribution prediction result; the process of the multi-modal feature fusion comprises: concatenating logging feature maps as additional channels with seismic profiles at a channel level; unifying spatial resolutions of the seismic data and the logging data through bilinear interpolation; performing channel dimension compression on the fused features by using a 1*1 convolution; the process of constructing the UNET++ network model comprises: extracting multi-scale features in the encoder through convolution and down-sampling; integrating multi-level features in the decoder through up-sampling and the nested dense skip connection; fusing multi-layer skip features from the encoder and logging information at each node of the decoder; the process of embedding the channel attention mechanism at the skip connection of the UNET++ network model comprises: generating an attention weight map through a convolution layer; multiplying the attention weight map and logging feature maps element by element, and splicing the weighted logging features and the seismic features; the process of training the fusion feature input by using the UNET++ network model comprises: processing logging curve sequences by using a bidirectional LSTM to extract depth-related features; inputting the depth-related features in a two-dimensional form into the UNET++ network model; optimizing network parameters through a cross-entropy loss function, and supervising outputs of different decoding layers in combination with multi-task loss.

2. The method of claim 1, wherein, the preprocessing process of the logging data comprises: identifying a brine layer by using an induction logging curve, and establishing a relationship model of array induction deep resistivity and acoustic travel time; generating a binary mask label to mark a brine layer position by analyzing logging curve combinations through a porosity-resistivity superposition method.

3. A deep learning-based seismic brine reservoir response feature identification system, characterized in that, The system comprises: a data preprocessing module configured to preprocess logging data and seismic data to generate a label mask corresponding to a brine reservoir; a feature fusion module configured to perform multi-modal feature fusion on the preprocessed logging data and the seismic data to form a fusion feature input; a network construction module configured to construct a UNET++ network model, wherein the UNET++ network model comprises an encoder, a decoder, and a nested dense skip connection structure, and the nested dense skip connection structure is used for dynamically fusing multi-scale features of the encoder and cross-modal features of logging information; The channel attention embedding module is configured to embed a channel attention mechanism at a skip connection of the UNET++ network model, and adjust a weight distribution of a logging feature to a seismic feature through a feature gate. The model training module is configured to train the fused feature input by using the UNET++ network model, and output a spatial distribution prediction result of the brine reservoir. The reservoir identification module is configured to identify a reflection feature and a spatial distribution rule of the deep brine reservoir based on the spatial distribution prediction result. The process of the multi-modal feature fusion includes: concatenating the logging feature map as an additional channel with the seismic profile; unifying spatial resolutions of the seismic data and the logging data through bilinear interpolation; performing channel dimension compression on the fused feature by using a 1x1 convolution; The process of constructing the UNET++ network model includes: extracting multi-scale features through convolution and down-sampling in the encoder; integrating multi-level features through up-sampling and nested dense skip connection in the decoder; fusing multi-level skip features from the encoder and logging information at each node of the decoder; The process of embedding the channel attention mechanism at the skip connection of the UNET++ network model includes: generating an attention weight map through a convolution layer; performing element-wise multiplication between the attention weight map and the logging feature map, and concatenating the weighted logging feature and the seismic feature; The process of training the fused feature input by using the UNET++ network model includes: processing the logging curve sequence by using a bidirectional LSTM to extract depth-related features; inputting the depth-related features in a two-dimensional form into the UNET++ network model; optimizing network parameters through a cross-entropy loss function, and supervising outputs of different decoding layers in combination with multi-task loss.

4. The system of claim 3, wherein, The data preprocessing module includes: a logging data processing unit configured to identify a brine layer by using an induction logging curve, and establish a relationship model between array induction deep resistivity and acoustic travel time; a label generation unit configured to analyze logging curve combinations by using a porosity-resistivity superposition method, and generate a binary mask label to mark a brine layer position.

5. A computer terminal device, characterized by includes: one or more processors; a memory coupled to the processors and configured to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based seismic brine reservoir response feature identification method according to any one of claims 1-2.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the deep learning-based seismic brine reservoir response feature identification method according to any one of claims 1-2.

Images