Seismic phase classification method, device, electronic equipment and medium

Abstract

The present application provides a seismic phase classification method, device, electronic device and medium. First, a two-dimensional scattering transform is used to extract multi-scale feature information with local deformation stability and translation invariance from two-dimensional seismic data. That is, when the seismic signal is affected by underground fault activity and produces local micro-deformations and time shifts, the extracted multi-scale features will not change significantly, which is conducive to the accurate classification of seismic data. Then, a densely connected convolutional network is constructed based on DenseBlock, and convolutional neural networks of different depths are used to further process and fuse feature information of different scales, that is, multi-resolution processing, and finally realize seismic phase classification of seismic data. Since DenseBlock adopts a fully interconnected structure, feature reuse can be achieved, so that feature information can be better retained during the layer-by-layer processing, further improving the accuracy of seismic phase classification results.

Description

Technical Field

[0001] The present application relates to the technical field of seismic phase analysis, and in particular to a seismic phase classification method, device, electronic equipment and medium. Background Art

[0002] Seismic facies identification is a key technology in seismic exploration and a fundamental and crucial step in oil and gas resource development. Accurate seismic facies identification can reveal the distribution of underground reservoirs, thereby guiding oil exploration. Traditionally, seismic facies interpretation is performed manually through collaboration between geophysicists, petrophysicists, and geologists. This method is time-consuming and heavily dependent on the experience of the individuals involved, resulting in low efficiency and high subjectivity.

[0003] With the development of machine learning, many automatic seismic phase interpretation methods based on machine learning have emerged and been successfully applied. In addition, many researchers have attempted to combine time-frequency analysis with machine learning to achieve phase interpretation of seismic data. However, these methods also have certain drawbacks:

[0004] (1) Although the seismic phase interpretation method based on machine learning can improve the efficiency of seismic phase interpretation to a certain extent, the seismic attributes of its input have a significant impact on the seismic phase interpretation results. Only when its input attributes are sufficiently sensitive to the seismic phase can a more accurate seismic phase interpretation result be obtained. The sensitive attributes corresponding to the seismic phases in data sets from different regions are not exactly the same. Therefore, it is necessary to manually select the optimal seismic attributes according to the conditions of different regions, which will increase the manual workload to a certain extent.

[0005] (2) The currently commonly used time-frequency transformation method does not have local deformation stability and translation invariance. That is, when the signal undergoes local slight deformation or translation, the characteristic information extracted using the current time-frequency transformation will undergo significant changes, which is not conducive to seismic phase classification of seismic data. Summary of the Invention

[0006] The present application provides a seismic phase classification method, device, electronic device and medium to solve or at least partially solve the above-mentioned problems.

[0007] A first aspect of an embodiment of the present application provides a seismic facies classification method, the method comprising:

[0008] Acquire two-dimensional seismic data of the target area, and pre-process the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point;

[0009] A two-dimensional wavelet scattering transform is performed on the two-dimensional seismic data using a pre-set scattering network to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform that have local deformation stability and translation invariance;

[0010] The multi-scale features of the seismic data corresponding to the above-mentioned seismic data sampling points are input into a preset densely connected convolutional network for multi-resolution processing and feature fusion to obtain the seismic phase classification results of the seismic data corresponding to the above-mentioned seismic data sampling points, and the above-mentioned seismic phase classification results are deblocked to obtain the seismic phase profile of the above-mentioned target area.

[0011] Optionally, the densely connected convolutional network includes a connected first-level DenseBlock group, a second-level DenseBlock group, a third-level DenseBlock group, a fully connected layer, and a classifier, wherein any DenseBlock group includes several DenseBlocks and convolutional layers of different depths, and two of the above DenseBlocks are connected through the above convolutional layer;

[0012] The multi-scale features are input into a preset densely connected convolutional network for multi-resolution processing and feature fusion to obtain the corresponding seismic phase classification results, including:

[0013] Inputting the scattering coefficients of different layers of the scattering transform into the first-level DenseBlock group of the densely connected convolutional network for the first feature extraction to obtain the first-level feature information;

[0014] The feature information extracted from the scattering coefficients of the first two layers of the scattering transformation using the first-level DenseBlock group is spliced ​​and input into the second-level DenseBlock group for a second feature fusion and extraction to obtain the second-level feature information;

[0015] The second-level feature information is concatenated with the feature information extracted from the scattering coefficient of the third layer of the scattering transform using the first-level DenseBlock group, and input into the third-level DenseBlock group for a third feature fusion and extraction to obtain the third-level feature information;

[0016] The third-level feature information is further fused and extracted through the fully connected layer, and the probability that the seismic data sampling point belongs to each type of seismic phase is calculated through the classifier, and the seismic phase category with the largest probability is selected as the seismic phase classification result of the seismic data corresponding to the seismic data sampling point.

[0017] Optionally, the above method further includes:

[0018] Acquire two-dimensional post-stack seismic data and randomly select a number of two-dimensional seismic data with seismic phase labels to form a training sample set; any training sample in the training sample set includes: two-dimensional seismic data corresponding to a sampling point and a seismic phase label, wherein the seismic phase label is used to characterize the seismic phase category to which the sampling point belongs;

[0019] A scattering network is constructed, and two-dimensional wavelet scattering transform is used to extract multi-scale feature information with local deformation stability and translation invariance of the two-dimensional seismic data corresponding to the sampling points in the above training sample set;

[0020] The initial densely connected convolutional network is trained using the multi-scale feature information of the two-dimensional seismic data corresponding to the sampling points in the above-mentioned training sample set and the seismic phase labels, thereby obtaining a densely connected convolutional network that can perform multi-resolution processing and fusion of the above-mentioned multi-scale features to predict the seismic phase category to which the above-mentioned sampling points belong.

[0021] Optionally, the generation of the training sample set includes:

[0022] Obtain sample 2D seismic data and preprocess them to obtain 2D post-stack seismic data Y, Y∈R N×M ; Where R represents a real number set, M represents the number of seismic traces of 2D seismic data, N represents the number of sampling points of each seismic data trace, R N×M represents a two-dimensional real number space of size N×M;

[0023] Select several two-dimensional seismic data from the above two-dimensional post-stack seismic data and obtain their corresponding seismic phase labels to form a training sample set Y train , Y train ∈R N×L , where L represents the number of seismic traces with known seismic phase labels, N also represents the number of sampling points for each seismic data, and R N×L represents a two-dimensional real number space of size N×L;

[0024] The two-dimensional seismic data corresponding to the above sampling points are: the two-dimensional post-stack seismic data are discretized into N×M sampling points, and for any of the above sampling points, the two-dimensional post-stack seismic data are divided using a square window with the sampling point as the center, and the two-dimensional seismic data of all sampling points within the square window are used as the two-dimensional seismic data corresponding to the sampling point.

[0025] Optionally, the step of constructing a scattering network includes:

[0026] A two-dimensional wavelet transform function family is constructed by rotating and scaling the bandpass filter Ψ(u):

[0027]

[0028] Among them, u is a two-dimensional variable, q∈Z is a scale parameter, and Z represents an integer set; Δ -θ ∈H is the rotation parameter, which means that the angle of rotation of variable u along the instantaneous clockwise direction is θ, and H represents the two-dimensional real number set R 2 A discrete rotation set of -θ The relationship between and the rotation angle θ can be expressed as:

[0029]

[0030] When the maximum scale is 2 Q When , the two-dimensional seismic data of a training sample input is a two-dimensional seismic data F, and the wavelet transform of the two-dimensional seismic data F can be expressed as:

[0031]

[0032] Among them, * represents convolution operation, W Q represents the wavelet operator, represents a low-pass filter, and {} represents a set of filters of different scales;

[0033] make Easy to know Then the modulus of the two-dimensional wavelet transform can be expressed as:

[0034]

[0035] Among them, M Q represents the wavelet modulus operator, For bandpass filters parameter The set of all values, represents a low-pass filter, and {} represents a set of filters of different scales;

[0036] Thus, the 0th-order scattering coefficient and the 1st-order wavelet modulus coefficient of the first layer of the scattering network can be obtained. The scattering coefficient of the scattering network is used to characterize the low-frequency characteristics of the two-dimensional seismic data F, which has local deformation stability and translation invariance. The wavelet modulus coefficient is used to characterize the high-frequency characteristics of the two-dimensional seismic data F; wherein the 0th-order scattering coefficient is expressed as:

[0037] The first-level wavelet modulus coefficients are expressed as:

[0038]

[0039] Where Ξ1 is the bandpass filter used by the first-order wavelet mode coefficient parameter The set of all values;

[0040] The 0th-level scattering coefficient is used as the output of the first layer of the scattering network, and the first-level wavelet mode coefficient is input into the second layer of the scattering network to calculate the first-level scattering coefficient and the second-level wavelet mode coefficient of the scattering network:

[0041]

[0042]

[0043] Wherein, Ξ2 is the bandpass filter used by the second-order wavelet mode coefficient parameter The set of all values;

[0044] Repeat the above wavelet convolution, nonlinear modulus and low-pass filtering operations, take the scattering coefficient of each layer of the scattering network as the output of the layer, and input the wavelet modulus coefficient of the layer into the next layer of the scattering network for calculation, so as to obtain the scattering coefficient and wavelet modulus coefficient of different scales in a step-by-step iterative manner;

[0045] The scattering coefficients output by the first several layers of the scattering network are combined to obtain the multi-scale features corresponding to the two-dimensional seismic data F.

[0046] Optionally, the training of the densely connected convolutional network includes:

[0047] Build a densely connected convolutional network and set the initial network parameters;

[0048] Inputting the multi-scale features of the two-dimensional seismic data corresponding to the sampling points in the training sample set extracted by the scattering network into the densely connected convolutional network, performing multi-resolution processing and feature fusion, and obtaining the predicted seismic phase category corresponding to the sampling points;

[0049] Calculating an objective function based on the predicted seismic phase category of the sampling point and the seismic phase label corresponding to the sampling point, and performing a reverse gradient update on the network parameters of the densely connected convolutional network based on the objective function to train the densely connected convolutional network and update the parameters;

[0050] The seismic phase label is represented as a binary vector obtained by one-hot encoding, and the objective function is:

[0051]

[0052] Among them, y kj Indicates the kth value of the one-hot encoding corresponding to the seismic phase label of the jth sampling point, It represents the kth value of the one-hot code corresponding to the predicted seismic phase category of the jth sampling point, and J represents the number of sampling points.

[0053] Optionally, the initial learning rate of the above densely connected convolutional network is set to lr initial =10 -3 , the learning decay constant is set to lr decay =10 -2 , the maximum number of iterations of the network is set to N max =50, then the Nth i The learning rate of the secondary network can be expressed as:

[0054]

[0055] in, Indicates the Nth i The learning rate of times.

[0056] A second aspect of an embodiment of the present application provides a seismic phase classification device, comprising:

[0057] A data acquisition module is used to acquire two-dimensional seismic data of a target area and pre-process the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point;

[0058] A multi-scale feature extraction module is configured to perform a two-dimensional wavelet scattering transform on the two-dimensional seismic data using a pre-set scattering network to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform that have local deformation stability and translation invariance;

[0059] The seismic phase category prediction module is used to input the multi-scale features of the seismic data corresponding to the above-mentioned seismic data sampling points into a preset densely connected convolutional network, perform multi-resolution processing and feature fusion, obtain the seismic phase classification results of the seismic data corresponding to the above-mentioned seismic data sampling points, and perform deblocking processing on the above-mentioned seismic phase classification results to obtain the seismic phase profile of the above-mentioned target area.

[0060] A third aspect of an embodiment of the present application provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps in the method described above in the first aspect of the present application are implemented.

[0061] The fourth aspect of an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the steps of the method described above in the first aspect of the present application.

[0062] Compared with the prior art, this application has the following advantages:

[0063] In an embodiment of the present application, first, a two-dimensional scattering transform is used to extract multi-scale feature information with local deformation stability and translation invariance from the two-dimensional seismic data of the seismic data sampling point step by step, that is, when the seismic signal produces local micro-deformation and time shift due to the influence of underground fault activity, the extracted multi-scale features will not change significantly, thereby ensuring that the seismic phase category to which the seismic signal belongs will not be misjudged due to adverse interference factors. Then, a densely connected convolutional neural network is constructed based on DenseBlock (densely connected block), and convolutional neural networks of different depths are used to further process and fuse feature information of different scales, that is, multi-resolution processing, and finally achieve seismic phase classification of seismic data of seismic data sampling points. Since the densely connected neural network adopts a fully interconnected structure, it is possible to reuse features, so that feature information can be better retained during the layer-by-layer processing, further improving the accuracy of the seismic phase classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a seismic facies classification method proposed in one embodiment of the present application;

[0065] Figure 2 This is a seismic data cross-section diagram of a seismic phase classification method proposed in an embodiment of the present application;

[0066] Figure 3 This is a schematic diagram of processing input data in a seismic phase classification method proposed in one embodiment of the present application;

[0067] Figure 4 This is a comparison diagram of the results of a seismic facies classification method proposed in one embodiment of the present application, wherein (a) shows the actual seismic facies section of the seismic data, and (b) shows the seismic facies classification result predicted using the method of this specification;

[0068] Figure 5 This is a structural diagram of a seismic phase classification device proposed in one embodiment of the present application.

[0069] Reference numerals: 1. Data acquisition module; 2. Multi-scale feature extraction module; 3. Seismic phase category prediction module. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0071] Seismic data contain a wealth of geological information. Seismic phase analysis is an important technique for extracting effective information based on the wave group characteristics of seismic reflections, thereby studying the distribution of sedimentary phases in three-dimensional space, the sedimentary environment of stratigraphic sequences, stratigraphic combination patterns, and regional geological background. In actual production, the accurate division of seismic phases is of great guiding significance for a deep understanding of the relationship between source, reservoir, and cap rock configuration in oil reservoirs and the prediction of high-quality reservoirs. Traditionally, the division of seismic phases mainly relies on manual experience to divide the reflection characteristics and phase axis morphology on the seismic profile to obtain a specific seismic phase structure. However, this method is very time-consuming, and the interpretation results are heavily dependent on the experience of the relevant personnel. It has the disadvantages of low efficiency and strong subjectivity.

[0072] With the development of machine learning, many automatic seismic phase interpretation methods based on machine learning have emerged and been successfully applied. The main idea is to use computer technology and certain mathematical methods to process and calculate seismic data to extract different attributes, and then combine each attribute according to the geological meaning corresponding to the spatial variation of each attribute to divide the seismic phase. In addition, many researchers have tried to combine time-frequency analysis with machine learning to achieve phase interpretation of seismic data. However, this type of method also has certain defects:

[0073] (1) Although the seismic phase interpretation method based on machine learning can improve the efficiency of seismic phase interpretation to a certain extent, the seismic attributes of its input have a great influence on the seismic phase interpretation results, and the seismic phase interpretation results are very dependent on the type of attributes selected and the number of attributes used. Only when its input attributes are sufficiently sensitive to the seismic phase can a more accurate seismic phase interpretation result be obtained. Moreover, the sensitive attributes corresponding to the seismic phases of data sets in different regions are not exactly the same. Therefore, it is necessary to manually select the optimal seismic attributes according to the conditions of different regions, which will increase the manual workload to a certain extent.

[0074] (2) The currently commonly used time-frequency transformation method does not have local deformation stability and translation invariance. That is, when the signal undergoes local slight deformation or translation, the characteristic information extracted using the current time-frequency transformation will undergo significant changes, which is not conducive to seismic phase classification of seismic data.

[0075] In view of this, this application first uses two-dimensional scattering transform to extract multi-scale feature information with local deformation stability and translation invariance from two-dimensional seismic data step by step, and then uses DenseBlock (densely connected block) as the basic unit and convolutional neural networks of different depths to further process and fuse the extracted feature information of each scale, thereby realizing phase classification of seismic data.

[0076] Please refer to Figure 1 , Figure 1This is a flow chart of a seismic phase classification method proposed in one embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0077] Step S101: Acquire two-dimensional seismic data of a target area, and pre-process the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point.

[0078] In this embodiment, seismic data refers to data that records earthquake activity and seismic wave propagation. It typically includes seismic waveforms, temporal and spatial information of seismic events, earthquake magnitudes, and source parameters. Seismic data is of great significance for the study of seismic activity, earthquake physics, and earthquake engineering. Processing seismic data can reveal subsurface conditions, thereby aiding the search for resources such as oil and gas.

[0079] For the target area to be surveyed, experiments can be conducted by simulating seismic waves, collecting relevant seismic data, and performing preprocessing, including denoising, static correction, dynamic correction, deconvolution, and time migration, to obtain two-dimensional seismic data for observation and generate a two-dimensional seismic data profile. Figure 2 As shown, the horizontal axis of the cross-sectional diagram represents the line number of the seismic trace, the vertical axis represents the depth, and CDP represents the common depth point (if different seismic traces all come from a common point on the underground interface, then the common point is the common depth point). Then, the two-dimensional seismic data is discretized into a number of seismic data sampling points to form a two-dimensional data grid, and the two-dimensional seismic data corresponding to a number of seismic data sampling points in the area around a certain seismic data sampling point are taken as the two-dimensional seismic data corresponding to the seismic data sampling point. For example, a certain seismic data sampling point can be used as the window center, and a square window of size I×I is used to divide the two-dimensional seismic data, and the two-dimensional seismic data corresponding to all seismic data sampling points in the square window are taken as the two-dimensional seismic data corresponding to the seismic data sampling point, that is, the size of the two-dimensional seismic data corresponding to the seismic data sampling point is I×I. Afterwards, the two-dimensional seismic data is processed by wavelet transform and densely connected convolutional network to predict the seismic phase category corresponding to the seismic data sampling point. After obtaining the seismic facies classifications for all seismic data sampling points, the seismic data sampling points are aligned one-to-one with the original 2D data grid positions and rearranged into a 2D seismic facies profile, thereby obtaining the 2D seismic facies profile of the target area. The specific seismic facies prediction process for each seismic data sampling point is detailed below.

[0080] Step S102: Using a pre-set scattering network, a two-dimensional wavelet scattering transform is performed on the two-dimensional seismic data to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform with local deformation stability and translation invariance.

[0081] As a powerful tool for analyzing nonstationary seismic signals, time-frequency transforms can extract characteristic information hidden in the time domain and sensitive to seismic phases. This can be used for seismic signal processing and feature extraction, leading to further phase identification. Deep learning, with its powerful feature learning and nonlinear problem-solving capabilities, can be used to process seismic signals or extract the complex nonlinear relationships between characteristic information and seismic phases from seismic signals using time-frequency transforms. Therefore, time-frequency transforms can be combined with deep learning to achieve seismic phase classification based on seismic data.

[0082] In this embodiment, the two-dimensional wavelet scattering transform is a calculation that outputs data simultaneously. During the scattering transform process, two coefficients are obtained for each layer: a low-frequency scattering coefficient and a high-frequency wavelet mode coefficient. The scattering coefficient of each layer has local deformation stability and translation invariance, and can represent the stable characteristic information of a certain frequency of the signal. Precisely because of the two stabilities of the scattering coefficient, when the signal is subjected to adverse factors that cause local deformation or slight translation, the extracted scattering coefficient will not undergo significant changes. This eliminates the impact of adverse factors on the final classification result of the seismic signal, thereby facilitating the phase classification of the seismic signal. The wavelet mode coefficient serves as input data for the next layer of the scattering transform and is used to continue calculating the scattering coefficient and wavelet mode coefficient of the next layer of the scattering network.

[0083] like Figure 3 As shown, after the two-dimensional seismic data corresponding to the seismic data sampling points is input into the scattering network, the first-layer operator WS1 first calculates the scattering coefficient S0F and wavelet mode coefficient U1F of the first layer. The scattering coefficient S0F serves as the output of the first layer and is input into the subsequent densely connected convolutional network. The wavelet mode coefficient U1F is also input into the second layer of the scattering network. Then, the second-layer operator WS2 calculates the scattering coefficient S1F and wavelet mode coefficient U2F of the second layer. The scattering coefficient S1F serves as the output of the second layer and is input into the subsequent densely connected convolutional network. The wavelet mode coefficient U2F is also input into the third layer of the scattering network for further calculation. This process continues, performing a two-dimensional wavelet scattering transform on the two-dimensional seismic data at the seismic data sampling points step by step. This iterative process obtains scattering coefficients and wavelet mode coefficients at different scales, thereby obtaining multi-scale features corresponding to the two-dimensional seismic data. Considering that the signal energy gradually decreases during the scattering transform process, the scattering coefficients of several layers before the scattering transform can be used as the output of the scattering network, depending on the actual application scenario. For example, the scattering coefficients of the first three layers can be used as the output of the entire scattering network. At this time, the scattering coefficients S0F, S1F and S2F of these three layers can be understood as containing characteristic information of different frequencies of the two-dimensional seismic data.

[0084] Step S103: Input the multi-scale features of the seismic data corresponding to the above-mentioned seismic data sampling points into a preset densely connected convolutional network, perform multi-resolution processing and feature fusion, obtain the seismic phase classification results of the seismic data corresponding to the above-mentioned seismic data sampling points, and perform deblocking processing on the above-mentioned seismic phase classification results to obtain the seismic phase profile of the above-mentioned target area.

[0085] In this embodiment, the densely connected convolutional network includes a connected first-level DenseBlock group, a second-level DenseBlock group, a third-level DenseBlock group, a fully connected layer and a classifier, wherein any DenseBlock group includes several DenseBlocks and convolutional layers of different depths, and two of the above DenseBlocks are connected through the above convolutional layer.

[0086] The multi-scale features are input into a preset densely connected convolutional network for multi-resolution processing and feature fusion to obtain corresponding seismic phase classification results, including: inputting the scattering coefficients of different layers of the scattering transform into the first-level DenseBlock group of the densely connected convolutional network for the first feature extraction to obtain first-level feature information; splicing the feature information extracted from the scattering coefficients of the first two layers of the scattering transform using the first-level DenseBlock group, and inputting it into the second-level DenseBlock group for the second feature fusion and extraction to obtain second-level feature information; splicing the second-level feature information with the feature information extracted from the scattering coefficients of the third layer of the scattering transform using the first-level DenseBlock group, and inputting it into the third-level DenseBlock group for the third feature fusion and extraction to obtain third-level feature information; further fusing and extracting the third-level feature information through the fully connected layer, and calculating the probability that the seismic data sampling point belongs to each type of seismic phase through the classifier, and selecting the seismic phase category with the largest probability as the seismic phase classification result of the seismic data corresponding to the seismic data sampling point.

[0087] Please refer to Figure 3 The convolutional network of this application is based on several DenseBlocks (Densely Connected Blocks). Each DenseBlock contains four convolutional layers, and the layers are densely connected. That is, the input of each layer comes from the feature information extracted by all the previous layers, and the output of each layer is directly connected to the input of all the subsequent layers, thereby achieving feature reuse (i.e., comprehensively re-exploring features of different "levels" - different representations), improving computational efficiency and processing effects. Figure 3In the example, the number of convolution kernels in each convolution layer of DenseBlock1 is 16, the number of convolution kernels in each convolution layer of DenseBlock2 is 32, the number of convolution kernels in each convolution layer of DenseBlock3 is 64, and the number of convolution kernels in each convolution layer of DenseBlock4 is 128. Except for the number of convolution kernels, the rest of the structures of each DenseBlock are the same.

[0088] like Figure 3 As shown in the figure, taking the scattering coefficients of the first three layers of the scattering network as an example, when processing the multi-scale features extracted by the scattering network, the scattering coefficient S0F output by the first layer of the scattering network is processed by two DenseBlocks, DenseBlock1 and DenseBlock2. These two DenseBlocks and the three convolutional layers connecting them together constitute the first-level DenseBlock group corresponding to the scattering coefficient S0F. The scattering coefficient S1F output by the second layer of the scattering network is processed by a DenseBlock, DenseBlock2. This DenseBlock and the convolutional layers connected to it together constitute the first-level DenseBlock group corresponding to the scattering coefficient S1F. The scattering coefficient S2F output by the third layer of the scattering network is processed by two DenseBlocks, DenseBlock2 and DenseBlock3. These two DenseBlocks and the two convolutional layers connecting them together constitute the first-level DenseBlock group corresponding to the scattering coefficient S2F. The main function of the convolutional layer is to reduce the dimensionality of the features and control the number of output features.

[0089] Thus, by inputting the scattering coefficients S0F, S1F, and S2F of the first three layers output by the scattering network into their respective corresponding first-level DenseBlock groups, a first feature extraction is performed separately to obtain the corresponding first-level feature information. Then, the first-level feature information of the extracted scattering coefficients S0F and S1F is spliced ​​and input into the second-level DenseBlock group (consisting of DenseBlock2, DenseBlock3, and a convolutional layer) for feature fusion and a second feature extraction to obtain second-level feature information. Thereafter, the fused second-level feature information is spliced ​​with the first-level feature information of the scattering coefficient S2F extracted above and input into the third-level DenseBlock group (consisting of DenseBlock3, DenseBlock4, and two convolutional layers) for feature fusion and a third feature extraction to obtain more complete third-level feature information. Finally, two fully connected layers (FC1 and FC2) are used for feature fusion and dimensionality reduction, and a Softmax classifier is used for classification prediction. The posterior probability that the two-dimensional seismic data at the seismic data sampling point belongs to a certain seismic phase is obtained. The seismic phase category with the highest probability is selected as the seismic phase classification result corresponding to the seismic data sampling point. At the same time, the scattering coefficient S0F is processed by a total of 6 DenseBlocks, the scattering coefficient S1F is processed by a total of 5 DenseBlocks, and the scattering coefficient S2F is processed by a total of 4 DenseBlocks. This allows for smoother fusion of multi-resolution features during the feature extraction process, preventing feature information from being lost due to a sudden drop in feature quantity, which would adversely affect the seismic phase classification results.

[0090] Therefore, the multi-resolution hybrid network model composed of a scattering network and a densely connected convolutional network can stably and accurately perform seismic facies classification, and the seismic facies classification results are deblocked to obtain the seismic facies profile of the target area, which can be used to assist in determining the sedimentary facies and lateral distribution range of the reservoir, laying the foundation for comprehensive reservoir prediction.

[0091] In another embodiment of the present application, the above method further includes:

[0092] Step S201: Acquire two-dimensional post-stack seismic data and randomly select a number of two-dimensional seismic data with seismic phase labels to form a training sample set; any training sample in the above training sample set includes: two-dimensional seismic data and seismic phase labels corresponding to a sampling point, and the above seismic phase label is used to characterize the seismic phase category to which the above sampling point belongs.

[0093] In this embodiment, firstly, sample 2D seismic data is obtained and preprocessed to obtain 2D post-stack seismic data Y, Y∈R N×M; Where R represents a real number set, M represents the number of seismic traces of 2D seismic data, N represents the number of sampling points of each seismic data trace, R N×M Represents a two-dimensional real number space of size N×M. Then, several two-dimensional seismic data are selected from the two-dimensional post-stack seismic data and their corresponding seismic phase labels are obtained to form a training sample set Y train , Y train ∈R N×L , where L represents the number of seismic traces with known seismic phase labels, N also represents the number of sampling points for each seismic data, and R N×L Represents a two-dimensional real number space of size N × L. The two-dimensional post-stack seismic data is then discretized into N × M sampling points. The two-dimensional seismic data corresponding to several sampling points in the area surrounding a certain sampling point are taken as the two-dimensional seismic data corresponding to the sampling point and input into the network for processing and prediction to determine the seismic facies category to which the sampling point belongs.

[0094] Specifically, for any sampling point, the corresponding two-dimensional seismic data is obtained in the following manner: with the sampling point as the center, the two-dimensional post-stack seismic data is divided using a square window, and the two-dimensional seismic data of all sampling points within the square window are used as the two-dimensional seismic data corresponding to the sampling point. For example, a square window of size I×I is used to divide the two-dimensional post-stack seismic data, so that the size of the two-dimensional seismic data corresponding to a certain sampling point is I×I. The divided data is then redistributed into a three-dimensional data form according to the number of sampling points × the two-dimensional seismic data corresponding to each sampling point earthquake, and the entire two-dimensional post-stack seismic data set becomes Y new ∈R NM×I×I , the training set becomes Y newtrain ∈R NL×I×I , where NM and NL refer to the number of sampling points in the entire dataset and the training set respectively, and R NM×I×I With R NL×I×I They refer to the three-dimensional real space with sizes of NM×I×I and NL×I×I respectively, and the three dimensions refer to the number of sampling points, the longitudinal length of the two-dimensional seismic data, and the lateral length of the two-dimensional seismic data.

[0095] Step S202: constructing a scattering network, and using a two-dimensional wavelet scattering transform to extract multi-scale feature information with local deformation stability and translation invariance of the two-dimensional seismic data corresponding to the sampling points in the training sample set.

[0096] In this embodiment, a two-dimensional wavelet transform function family is first constructed by rotating and scaling the bandpass filter Ψ(u):

[0097]

[0098] Among them, u is a two-dimensional variable, q∈Z is a scale parameter, and Z represents an integer set; Δ-θ ∈H is the rotation parameter, which means that the angle of rotation of variable u along the instantaneous clockwise direction is θ, and H represents the two-dimensional real number set R 2 A discrete rotation set of -θ The relationship between and the rotation angle θ can be expressed as:

[0099]

[0100] When the maximum scale is 2 Q When , the two-dimensional seismic data of a training sample input is a two-dimensional seismic data F, and the wavelet transform of the two-dimensional seismic data F can be expressed as:

[0101]

[0102] Among them, * represents convolution operation, W Q represents the wavelet operator, represents a low-pass filter, and {} represents a set of filters of different scales;

[0103] make Easy to know Then the modulus of the two-dimensional wavelet transform can be expressed as:

[0104]

[0105] Among them, M Q represents the wavelet modulus operator, For bandpass filters parameter The set of all values, Represents a low-pass filter, and {} represents a set of filters of different scales.

[0106] The wavelet modulus operation based on the above two-dimensional wavelet transform modulus can obtain the zero-order scattering coefficient with translation invariance and local deformation stability, which is expressed as: Because of the The signal is convolved with the low-pass filter, and the information obtained is only low-frequency information, so it will cause the loss of high-frequency information. In order to recover the lost high-frequency information, the high-frequency component in the two-dimensional wavelet transform modulus formula can be Continue iterating, and the high-frequency component is also called the first-level wavelet mode coefficient. The first-level wavelet mode coefficient can be expressed as:

[0107]

[0108] Where Ξ1 is the bandpass filter used by the first-order wavelet mode coefficient parameter That is, the scattering coefficient of the scattering network is used to characterize the low-frequency characteristics of the two-dimensional seismic data F, and the wavelet modulus coefficient is used to characterize the high-frequency characteristics of the two-dimensional seismic data F.

[0109] Afterwards, the 0th-level scattering coefficient is used as the output of the first layer of the scattering network, and the first-level wavelet mode coefficient is input into the second layer of the scattering network to calculate the first-level scattering coefficient and the second-level wavelet mode coefficient of the scattering network:

[0110]

[0111]

[0112] Wherein, Ξ2 is the bandpass filter used by the second-order wavelet mode coefficient parameter The set of all possible values.

[0113] By repeating the above wavelet convolution, nonlinear modulus, and low-pass filtering operations, the scattering coefficient of each layer of the scattering network is used as the output of that layer, and the wavelet modulus coefficient of that layer is input into the next layer of the scattering network for calculation. The scattering coefficients and wavelet modulus coefficients of different scales are obtained in a step-by-step iterative manner. Finally, the scattering coefficients output by the first several layers of the scattering network are combined to obtain the multi-scale features corresponding to the two-dimensional seismic data F.

[0114] Step S203: The initial densely connected convolutional network is trained using the multi-scale feature information of the two-dimensional seismic data corresponding to the sampling points in the extracted training sample set and the seismic phase labels to obtain a densely connected convolutional network capable of performing multi-resolution processing and fusion of the multi-scale features to predict the seismic phase category to which the sampling points belong.

[0115] In this embodiment, a densely connected convolutional network is first constructed based on DenseBlock. DenseBlock can realize feature reuse. Its output contains the low-level features of the input data, making the feature information obtained richer. In order to establish the mapping relationship between the scattering coefficient obtained by scattering transformation of two-dimensional seismic data and the seismic phase category, the initial network parameters are set. For example, the initial learning rate is set to lr initial =10 -3 , the learning decay constant is set to lr decay =10 -2 , the maximum number of iterations of the network is set to N max =50, then the Nth i The learning rate of the secondary network can be expressed as:

[0116]

[0117] in, Indicates the Nth i The learning rate of times.

[0118] The multi-scale features of the two-dimensional seismic data corresponding to the sampling points in the training sample set extracted by the scattering network are input into the densely connected convolutional network for multi-resolution processing and feature fusion. The classifier is used to calculate the posterior probability that the sampling point belongs to a certain seismic phase, and the seismic phase corresponding to the maximum value is taken as the seismic phase to which the current sampling point belongs, so as to obtain the predicted seismic phase category corresponding to the sampling point. Then, the objective function is calculated based on the predicted seismic phase category of the sampling point and the seismic phase label corresponding to the sampling point, and the network parameters of the densely connected convolutional network are reversely gradient updated based on the objective function to train the densely connected convolutional network and update the parameters. For the convenience of calculation, when calculating the objective function, the seismic phase label needs to be One-hot encoded first, that is, the classification value is mapped to an integer value. Then, each integer value is represented as a binary vector, and except for the integer index value 1, the other positions are all zero values. Therefore, the above seismic phase label is represented as: a binary vector obtained by One-hot encoding, and the above objective function is:

[0119]

[0120] Among them, y kj Indicates the kth value of the one-hot encoding corresponding to the seismic phase label of the jth sampling point, It represents the kth value of the one-hot code corresponding to the predicted seismic phase category of the jth sampling point, and J represents the number of sampling points.

[0121] The initial densely connected convolutional network is trained with the Adam optimizer using the above objective function to obtain the optimal parameter set of the network model suitable for seismic phase classification, thereby forming a multi-resolution network capable of seismic phase classification.

[0122] Please refer to Figure 4 , Figure 4 This is a comparison chart of the results of a seismic phase classification method proposed in one embodiment of the present application. The sample data comes from the F3 seismic dataset in the offshore area of ​​the Netherlands. Figure 4 middle, Figure 4 (a) shows the original seismic phase distribution profile. The data contains 869 seismic records, of which 40 have known seismic phase labels. Each seismic record contains 255 sampling points. The seismic data contains 6 types of seismic phases: Upper North, Middle North, Lower North, Chalk / Rijnland, Scruff, and Zechstein. Figure 4 (b) is the seismic phase profile predicted by the present invention. Figure 4 (a) with Figure 4 (b) It can be seen that the seismic phase classification results predicted by the multi-resolution convolutional network composed of the scattering network + densely connected convolutional network of the present invention are relatively accurate, and good results are achieved in describing the continuity within the same type of seismic phases and depicting the boundaries of different types of seismic phases, providing a favorable foundation for subsequent further improvements.

[0123] Based on the same inventive concept, an embodiment of the present application provides a seismic phase classification device. Figure 5 , Figure 5 : is a schematic structural diagram of a seismic phase classification device proposed in one embodiment of the present application, comprising:

[0124] Data acquisition module 1 is used to acquire two-dimensional seismic data of the target area and pre-process the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point;

[0125] Multi-scale feature extraction module 2 is used to perform a two-dimensional wavelet scattering transform on the two-dimensional seismic data using a pre-set scattering network to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform with local deformation stability and translation invariance;

[0126] The seismic phase category prediction module 3 is used to input the multi-scale features of the seismic data corresponding to the above-mentioned seismic data sampling points into a preset densely connected convolutional network, perform multi-resolution processing and feature fusion, obtain the seismic phase classification results of the seismic data corresponding to the above-mentioned seismic data sampling points, and perform deblocking processing on the above-mentioned seismic phase classification results to obtain the seismic phase profile of the above-mentioned target area.

[0127] Optionally, the densely connected convolutional network includes a connected first-level DenseBlock group, a second-level DenseBlock group, a third-level DenseBlock group, a fully connected layer, and a classifier, wherein any DenseBlock group includes several DenseBlocks of different depths and a convolutional layer, and two of the above DenseBlocks are connected through the above convolutional layer; the above seismic phase category prediction module 3 includes:

[0128] A first-level feature extraction submodule is used to input the scattering coefficients of different layers of the scattering transform into the first-level DenseBlock group of the densely connected convolutional network for the first feature extraction to obtain first-level feature information;

[0129] The second-level feature extraction submodule is used to splice the feature information extracted from the scattering coefficients of the first two layers of the scattering transformation using the above-mentioned first-level DenseBlock group, and input it into the above-mentioned second-level DenseBlock group for a second feature fusion and extraction to obtain the second-level feature information;

[0130] The third-level feature extraction submodule is used to splice the second-level feature information with the feature information extracted from the scattering coefficient of the third layer of the scattering transform using the first-level DenseBlock group, and input it into the third-level DenseBlock group for a third feature fusion and extraction to obtain the third-level feature information;

[0131] The feature fusion prediction module is used to further fuse and extract the above-mentioned third-level feature information through the above-mentioned fully connected layer, and calculate the probability that the above-mentioned seismic data sampling point belongs to each type of seismic phase through the above-mentioned classifier, and select the seismic phase category with the largest probability as the seismic phase classification result of the seismic data corresponding to the above-mentioned seismic data sampling point.

[0132] Optionally, the above device further includes:

[0133] A sample acquisition module is used to acquire two-dimensional post-stack seismic data and randomly select a number of two-dimensional seismic data with seismic phase labels to form a training sample set; any training sample in the training sample set includes: two-dimensional seismic data corresponding to a sampling point and a seismic phase label, wherein the seismic phase label is used to characterize the seismic phase category to which the sampling point belongs;

[0134] A scattering network construction module is used to construct a scattering network and extract multi-scale feature information with local deformation stability and translation invariance of the two-dimensional seismic data corresponding to the sampling points in the training sample set using a two-dimensional wavelet scattering transform;

[0135] The densely connected convolutional network training module is used to train the initial densely connected convolutional network using the multi-scale feature information of the two-dimensional seismic data corresponding to the sampling points in the above-mentioned training sample set and the seismic phase labels, so as to obtain a densely connected convolutional network that can perform multi-resolution processing and fusion of the above-mentioned multi-scale features to predict the seismic phase category to which the above-mentioned sampling points belong.

[0136] Optionally, the sample acquisition module includes:

[0137] The preprocessing submodule is used to obtain sample 2D seismic data and preprocess them to obtain 2D post-stack seismic data Y, Y∈R N×M ; Where R represents a real number set, M represents the number of seismic traces of 2D seismic data, N represents the number of sampling points of each seismic data trace, R N×M represents a two-dimensional real number space of size N×M;

[0138] The selection submodule is used to select several two-dimensional seismic data from the above two-dimensional post-stack seismic data and obtain their corresponding seismic phase labels to form a training sample set Y train , Y train ∈R N×L , where L represents the number of seismic traces with known seismic phase labels, N also represents the number of sampling points for each seismic data, and R N×L represents a two-dimensional real number space of size N×L;

[0139] The two-dimensional seismic data corresponding to the above sampling points are: the two-dimensional post-stack seismic data are discretized into N×M sampling points, and for any of the above sampling points, the two-dimensional post-stack seismic data are divided using a square window with the sampling point as the center, and the two-dimensional seismic data of all sampling points within the square window are used as the two-dimensional seismic data corresponding to the sampling point.

[0140] Optionally, the scattering network building module includes:

[0141] The wavelet transform construction submodule is used to construct a two-dimensional wavelet transform function family by rotating and scaling the bandpass filter Ψ(u):

[0142]

[0143] Among them, u is a two-dimensional variable, q∈Z is a scale parameter, and Z represents an integer set; Δ -θ ∈H is the rotation parameter, which means that the angle of rotation of variable u along the instantaneous clockwise direction is θ, and H represents the two-dimensional real number set R 2 A discrete rotation set of -θ The relationship between and the rotation angle θ can be expressed as:

[0144]

[0145] When the maximum scale is 2 Q When , the two-dimensional seismic data of a training sample input is a two-dimensional seismic data F, and the wavelet transform of the two-dimensional seismic data F can be expressed as:

[0146]

[0147] Among them, * represents convolution operation, W Q represents the wavelet operator, represents a low-pass filter, and {} represents a set of filters of different scales;

[0148] make Easy to know Then the modulus of the two-dimensional wavelet transform can be expressed as:

[0149]

[0150] Among them, M Q represents the wavelet modulus operator, For bandpass filters parameter The set of all values, represents a low-pass filter, and {} represents a set of filters of different scales;

[0151] The coefficient calculation submodule is used to calculate the 0th-order scattering coefficient and the 1st-order wavelet mode coefficient of the first layer of the scattering network. The scattering coefficient of the scattering network is used to characterize the low-frequency characteristics of the two-dimensional seismic data F, and the wavelet mode coefficient is used to characterize the high-frequency characteristics of the two-dimensional seismic data F. The 0th-order scattering coefficient is expressed as:

[0152] The first-level wavelet modulus coefficients are expressed as:

[0153]

[0154] Where Ξ1 is the bandpass filter used by the first-order wavelet mode coefficient parameter The set of all values;

[0155] The 0th-level scattering coefficient is used as the output of the first layer of the scattering network, and the first-level wavelet mode coefficient is input into the second layer of the scattering network to calculate the first-level scattering coefficient and the second-level wavelet mode coefficient of the scattering network:

[0156]

[0157]

[0158] Wherein, Ξ2 is the bandpass filter used by the second-order wavelet mode coefficient parameter The set of all values;

[0159] The output submodule is used to repeat the above wavelet convolution, nonlinear modulus and low-pass filtering operations, taking the scattering coefficient of each layer of the scattering network as the output of the layer, and inputting the wavelet modulus coefficient of the layer into the next layer of the scattering network for calculation, so as to iteratively obtain the scattering coefficients and wavelet modulus coefficients of different scales step by step;

[0160] The scattering coefficients output by the first several layers of the scattering network are combined to obtain the multi-scale features corresponding to the two-dimensional seismic data F.

[0161] Optionally, the densely connected convolutional network training module includes:

[0162] The initial construction submodule is used to build a densely connected convolutional network and set the initial network parameters;

[0163] A sample prediction submodule is configured to input the multi-scale features of the two-dimensional seismic data corresponding to the sampling points in the training sample set extracted by the scattering network into the densely connected convolutional network, perform multi-resolution processing and feature fusion, and obtain the predicted seismic phase category corresponding to the sampling points;

[0164] A training and updating submodule is configured to calculate an objective function based on the predicted seismic phase category of the sampling point and the seismic phase label corresponding to the sampling point, and perform a reverse gradient update on the network parameters of the densely connected convolutional network based on the objective function to train and update the parameters of the densely connected convolutional network;

[0165] The seismic phase label is represented as a binary vector obtained by one-hot encoding, and the objective function is:

[0166]

[0167] Among them, y kj Indicates the kth value of the one-hot encoding corresponding to the seismic phase label of the jth sampling point, It represents the kth value of the one-hot code corresponding to the predicted seismic phase category of the jth sampling point, and J represents the number of sampling points.

[0168] Optionally, the densely connected convolutional network training module further includes:

[0169] Set the submodule to set the initial learning rate to lr initial =10 -3 , the learning decay constant is set to lr decay =10 -2 , the maximum number of iterations of the network is set to N max =50, then the Nth i The learning rate of the secondary network can be expressed as:

[0170]

[0171] in, Indicates the Nth i The learning rate of times.

[0172] As for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0173] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0174] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0175] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0176] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0177] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0178] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0179] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

[0180] The above is a detailed introduction to the seismic phase classification method, device, electronic equipment and medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A seismic facies classification method, characterized in that: The method comprises: Acquire two-dimensional seismic data of the target area, and preprocess the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point; Performing a two-dimensional wavelet scattering transform on the two-dimensional seismic data using a pre-set scattering network to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform that have local deformation stability and translation invariance; Inputting the multi-scale features of the seismic data corresponding to the seismic data sampling points into a preset densely connected convolutional network, performing multi-resolution processing and feature fusion, obtaining seismic phase classification results of the seismic data corresponding to the seismic data sampling points, and performing deblocking processing on the seismic phase classification results to obtain a seismic phase profile of the target area; The densely connected convolutional network includes a connected first-level DenseBlock group, a second-level DenseBlock group, a third-level DenseBlock group, a fully connected layer and a classifier, wherein any DenseBlock group includes several DenseBlocks and convolutional layers of different depths, and two DenseBlocks are connected through the convolutional layer; The multi-scale features are input into a preset densely connected convolutional network to perform multi-resolution processing and feature fusion to obtain corresponding seismic phase classification results, including: Inputting the scattering coefficients of different layers of the scattering transform into the first-level DenseBlock group of the densely connected convolutional network for first feature extraction to obtain first-level feature information; The feature information extracted from the scattering coefficients of the first two layers of the scattering transformation using the first-level DenseBlock group is spliced ​​and input into the second-level DenseBlock group for a second feature fusion and extraction to obtain the second-level feature information; The second-level feature information is concatenated with the feature information extracted from the scattering coefficient of the third layer of the scattering transform using the first-level DenseBlock group, and the concatenated feature information is input into the third-level DenseBlock group for a third feature fusion and extraction to obtain the third-level feature information; The third-level feature information is further fused and extracted through the fully connected layer, and the probability of the seismic data sampling point belonging to each type of seismic phase is calculated through the classifier, and the seismic phase category with the largest probability is selected as the seismic phase classification result of the seismic data corresponding to the seismic data sampling point.

2. The method according to claim 1, characterized in that The method further comprises: Acquire two-dimensional post-stack seismic data and randomly select a number of two-dimensional seismic data with seismic phase labels to form a training sample set; any training sample in the training sample set includes: two-dimensional seismic data corresponding to a sampling point and a seismic phase label, wherein the seismic phase label is used to characterize the seismic phase category to which the sampling point belongs; Constructing a scattering network and using a two-dimensional wavelet scattering transform to extract multi-scale feature information with local deformation stability and translation invariance of the two-dimensional seismic data corresponding to the sampling points in the training sample set; The initial densely connected convolutional network is trained using the multi-scale feature information of the two-dimensional seismic data corresponding to the sampling points in the extracted training sample set and the seismic phase labels, thereby obtaining a densely connected convolutional network that can perform multi-resolution processing and fusion of the multi-scale features to predict the seismic phase category to which the sampling points belong.

3. The method according to claim 2, characterized in that The generation of the training sample set includes: Obtain sample 2D seismic data and preprocess them to obtain 2D post-stack seismic data Y, Y∈R N×M ; Where R represents a real number set, M represents the number of seismic traces of 2D seismic data, N represents the number of sampling points of each seismic data trace, R N×M represents a two-dimensional real number space of size N×M; Select several two-dimensional seismic data from the two-dimensional post-stack seismic data and obtain their corresponding seismic phase labels to form a training sample set Y train , Y train ∈R N×L , where L represents the number of seismic traces with known seismic phase labels, N also represents the number of sampling points for each seismic data, and R N×L represents a two-dimensional real number space of size N×L; The two-dimensional seismic data corresponding to the sampling points are as follows: the two-dimensional post-stack seismic data are discretized into N×M sampling points, and for any of the sampling points, the two-dimensional post-stack seismic data are divided using a square window with the sampling point as the center, and the two-dimensional seismic data of all the sampling points within the square window are used as the two-dimensional seismic data corresponding to the sampling point.

4. The method according to claim 2, characterized in that Constructing the scattering network includes: A two-dimensional wavelet transform function family is constructed by rotating and scaling the bandpass filter Ψ(u): Among them, u is a two-dimensional variable, q∈Z is a scale parameter, and Z represents an integer set; Δ -θ ∈H is the rotation parameter, which means that the angle of rotation of variable u along the instantaneous clockwise direction is θ, and H represents the two-dimensional real number set R 2 A discrete rotation set of -θ The relationship between and the rotation angle θ can be expressed as: When the maximum scale is 2 Q When , the two-dimensional seismic data of a training sample input is a two-dimensional seismic data F, and the wavelet transform of the two-dimensional seismic data F is expressed as: Among them, * represents convolution operation, W Q represents the wavelet operator, represents a low-pass filter, and {} represents a set of filters of different scales; make Easy to know Then the two-dimensional wavelet transform modulo expression is: Among them, M Q represents the wavelet modulus operator, For bandpass filters parameter The set of all values, represents a low-pass filter, and {} represents a set of filters of different scales; Thus, the 0th-order scattering coefficient and the 1st-order wavelet modulus coefficient of the first layer of the scattering network can be obtained. The scattering coefficient of the scattering network is used to characterize the low-frequency characteristics of the two-dimensional seismic data F, which has local deformation stability and translation invariance. The wavelet modulus coefficient is used to characterize the high-frequency characteristics of the two-dimensional seismic data F; wherein the 0th-order scattering coefficient is expressed as: The first-level wavelet modulus coefficient is expressed as: Where Ξ1 is the bandpass filter used by the first-order wavelet mode coefficient parameter The set of all values; The 0th-order scattering coefficient is used as the output of the first layer of the scattering network, and the first-order wavelet mode coefficient is input into the second layer of the scattering network to calculate the first-order scattering coefficient and the second-order wavelet mode coefficient of the scattering network: Wherein, Ξ2 is the bandpass filter used by the second-order wavelet mode coefficient parameter The set of all values; Repeat the above wavelet convolution, nonlinear modulus and low-pass filtering operation process, take the scattering coefficient of each layer of the scattering network as the output of the layer, and input the wavelet modulus coefficient of the layer into the next layer of the scattering network for calculation, so as to iterate step by step to obtain scattering coefficients and wavelet modulus coefficients of different scales; The scattering coefficients output by the first several layers of the scattering network are combined to obtain the multi-scale features corresponding to the two-dimensional seismic data F.

5. The method according to claim 2, characterized in that The training of the densely connected convolutional network includes: Build a densely connected convolutional network and set the initial network parameters; Inputting the multi-scale features of the two-dimensional seismic data corresponding to the sampling points in the training sample set extracted by the scattering network into the densely connected convolutional network, performing multi-resolution processing and feature fusion, and obtaining the predicted seismic phase category corresponding to the sampling points; Calculating an objective function according to the predicted seismic phase category of the sampling point and the seismic phase label corresponding to the sampling point, and performing reverse gradient update on the network parameters of the densely connected convolutional network based on the objective function to train the densely connected convolutional network and update the parameters; The seismic phase label is represented as a binary vector obtained by one-hot encoding, and the objective function is: Among them, y kj Indicates the kth value of the one-hot encoding corresponding to the seismic phase label of the jth sampling point, It represents the kth value of the one-hot code corresponding to the predicted seismic phase category of the jth sampling point, and J represents the number of sampling points.

6. The method according to claim 5, characterized in that The initial learning rate of the densely connected convolutional network is set to lr initial =10 -3 , the learning decay constant is set to lr decay =10 -2 , the maximum number of iterations of the network is set to N max =50, then the Nth i The learning rate of the secondary network can be expressed as: in, Indicates the Nth i The learning rate of times.

7. A seismic phase classification device, characterized in that: The device comprises: A data acquisition module is used to acquire two-dimensional seismic data of a target area and pre-process the two-dimensional seismic data to obtain two-dimensional seismic data corresponding to each seismic data sampling point; A multi-scale feature extraction module is configured to perform a two-dimensional wavelet scattering transform on the two-dimensional seismic data using a pre-set scattering network to obtain multi-scale features of the seismic data corresponding to the seismic data sampling points; the multi-scale features are scattering coefficients corresponding to different layers of the scattering transform that have local deformation stability and translation invariance; A seismic phase classification prediction module is configured to input the multi-scale features of the seismic data corresponding to the seismic data sampling points into a preset densely connected convolutional network, perform multi-resolution processing and feature fusion, obtain seismic phase classification results of the seismic data corresponding to the seismic data sampling points, and perform deblocking processing on the seismic phase classification results to obtain a seismic phase profile of the target area; The densely connected convolutional network includes a connected first-level DenseBlock group, a second-level DenseBlock group, a third-level DenseBlock group, a fully connected layer and a classifier, wherein any DenseBlock group includes several DenseBlocks and convolutional layers of different depths, and two DenseBlocks are connected through the convolutional layer; The multi-scale features are input into a preset densely connected convolutional network to perform multi-resolution processing and feature fusion to obtain corresponding seismic phase classification results, including: Inputting the scattering coefficients of different layers of the scattering transform into the first-level DenseBlock group of the densely connected convolutional network for first feature extraction to obtain first-level feature information; The feature information extracted from the scattering coefficients of the first two layers of the scattering transformation using the first-level DenseBlock group is spliced ​​and input into the second-level DenseBlock group for a second feature fusion and extraction to obtain the second-level feature information; The second-level feature information is concatenated with the feature information extracted from the scattering coefficient of the third layer of the scattering transform using the first-level DenseBlock group, and the concatenated feature information is input into the third-level DenseBlock group for a third feature fusion and extraction to obtain the third-level feature information; The third-level feature information is further fused and extracted through the fully connected layer, and the probability of the seismic data sampling point belonging to each type of seismic phase is calculated through the classifier, and the seismic phase category with the largest probability is selected as the seismic phase classification result of the seismic data corresponding to the seismic data sampling point.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images