Geological information optimization extraction method in seismic wave exploration field

Abstract

The invention provides a geological information optimization extraction method in the field of seismic wave exploration. The geological information optimization extraction method comprises the following steps: step 1, establishing a database; step 2, a 1DSE-ResNeXt network model is established, and the network is trained; step 3, outputting and storing the feature vector before the softmax layer of the network; 4, inputting the feature vector output in the step 3 into an SVM, and training an SVM classifier; and step 5, obtaining a final identification result through the trained SVM classifier. According to the geological information optimization extraction method in the seismic wave exploration field, the classification capability is improved, the support vector machine classifier can further improve the classification precision, the identification precision of oil-gas exploration is improved, a new reference is provided for application of distributed optical fiber sensing in the oil-gas exploration field, and the method has a wide application prospect.

Description

Technical Field

[0001] The invention relates to the technical field of seismic wave exploration, and in particular to a method for optimizing and extracting geological information in the field of seismic wave exploration. Background Art

[0002] Oil and natural gas are important strategic resources in my country. As the first key link in the exploitation of oil and gas resources, the development and improvement of oil and gas exploration technology is the focus of researchers in the field. In oil and gas exploration, the seismic wave exploration method is the most important method in physical exploration, which is based on artificially creating strong vibrations and analyzing the refracted waves returned by the rock formation to obtain the geological structure and oil and gas circles. In recent years, various seismic sensors based on seismic wave measurement have been widely studied. However, traditional seismic sensors such as piezoelectric sensors and dynamic coil sensors are mostly composed of electronic components, which are extremely susceptible to electromagnetic interference and are difficult to meet the accuracy and reliability requirements in practical applications.

[0003] Distributed acoustic sensing (DAS) has many advantages such as easy bending, small size, high sensitivity, intrinsic safety, etc., and can realize long-distance, distributed, and real-time acquisition of dynamic strain along the optical fiber. It is currently widely used in many fields such as oil and gas field exploration. However, due to the very high sensitivity of optical fiber, the complex exploration objects, and the changeable environment. Therefore, there is an urgent need for effective recognition algorithms to effectively distinguish the categories of geological information, so as to obtain more accurate resource information. In recent years, with the development of computer technology and artificial intelligence, machine learning and deep learning have been widely used in the field of oil and gas DAS pattern recognition. Convolutional neural network (CNN) is frequently used in the recognition of seismic signals because of its powerful feature learning ability. However, the continuous deepening of the number of network layers to obtain higher recognition accuracy has always been a key issue restricting the application of convolutional neural networks. Most CNN classification algorithms are not highly sensitive to the information of key features, and are easily interfered by non-key feature information, resulting in inaccurate classification results. Therefore, it is urgent to build a more effective convolutional neural network model.

[0004] In the Chinese patent application with application number: CN202110752889.9, a high-dimensional closed-loop network seismic inversion method under logging constraints is involved. The method specifically includes the following steps: Step 1: Build a convolutional neural network; including a forward modeling network and an inversion network; Step 2: Prepare training data; including logging wave impedance data, interpolated wave impedance data and synthetic seismic data; Step 3: Train the network and fine-tune: train the forward modeling network and the inversion network, and then apply the one-dimensional logging data to the two-dimensional model and the three-dimensional model, and fine-tune them; Step 4: Prediction and evaluation: first invert the seismic data; then forward model the inversion results to obtain reconstructed seismic data; finally, use the reconstructed seismic data to evaluate the effectiveness of the inversion results. This method does not require additional collection of input and reference images, can ensure good lateral continuity, and has higher accuracy than traditional and other deep learning inversion methods.

[0005] In the Chinese patent application with application number: CN201310414827.2, a supervised waveform classification method in three-dimensional seismic signals is involved, which includes the steps of data preprocessing, feature selection and classification identification. The beneficial effect is that the present invention is based on three-dimensional seismic signal data and logging data information, optimizes the extracted attribute features through genetic algorithms, and uses SVM classification algorithms to classify the waveforms of the analyzed three-dimensional seismic target layer data, identify different seismic phases, and then provide reliable support for subsequent seismic data interpretation, improve the reliability of lithology prediction, sand body prediction, fractured oil and gas reservoir prediction and concealed oil and gas reservoir prediction, etc. Compared with using only SVM to design a classifier, the addition of genetic algorithms for feature selection reduces the design complexity of the SVM classifier, thereby improving the efficiency of waveform classification processing.

[0006] In the Chinese patent application with application number: CN202211206267.7, a real-time microseismic magnitude calculation method and device based on deep learning are involved, including: constructing a DAS-based horizontal well microseismic monitoring system; constructing a training data set; constructing a magnitude calculation module: the magnitude calculation module includes two input branches, frequency spectrum and time waveform, and the two input branches use a 3-layer convolution structure to extract the frequency characteristics and waveform characteristics of microseismic events respectively, and then perform model fusion, and then pass through 2 fully connected layers, and finally output the calculated magnitude; the input of the magnitude calculation module is a single DAS microseismic strain data, and the output is the calculated magnitude, and the error between the calculated magnitude and the theoretical magnitude Gaussian distribution is used to update the network parameters; the magnitude calculation module is trained using the training data set; and the actual data is analyzed and processed. The magnitude calculation method of the invention improves the ability to quickly estimate the microseismic magnitude, without the need to convert the strain data, and improves the accuracy of microseismic magnitude estimation.

[0007] In the Chinese patent application with application number: CN202110830865.0, a surface wave intelligent inversion exploration method, storage medium and terminal device are involved. The method includes the following steps: using a forward algorithm to generate samples for training, training CNN based on the samples to form a classifier, the trained classifier can directly predict the type of the formation model through the Rayleigh wave spectrum, and automatically set the initial model of the formation model according to the classification result of the Rayleigh wave spectrum and its corresponding physical parameter range; according to the initial model and the dispersion curve extracted from the Rayleigh wave spectrum, the initial model is iteratively adjusted using the particle swarm method, and finally the S-wave velocity model result of the formation is obtained. The classifier obtained by training in this invention can directly process two-dimensional matrix data, thereby avoiding manual design of features; the classifier can directly predict the dynamic information of the formation through the Rayleigh wave spectrum and has high accuracy and computational efficiency. This invention method has good application prospects in detecting low-velocity layers.

[0008] The above existing technologies are all significantly different from the present invention and fail to solve the technical problem we want to solve. For this reason, we have invented a new method for optimizing and extracting geological information in the field of seismic wave exploration. Summary of the invention

[0009] The purpose of the present invention is to provide a method for optimizing the extraction of geological information in the field of seismic wave exploration, which can effectively distinguish the types of geological information and thus obtain more accurate resource information.

[0010] The purpose of the present invention can be achieved by the following technical measures: a method for optimizing the extraction of geological information in the field of seismic wave exploration, the method for optimizing the extraction of geological information in the field of seismic wave exploration comprising:

[0011] Step 1: Build a database;

[0012] Step 2: Build a 1DSE-ResNeXt network model and train the network;

[0013] Step 3, save the feature vector output before the network softmax layer;

[0014] Step 4, input the feature vector output in step 3 into SVM to train the SVM classifier;

[0015] Step 5: Get the final recognition result through the trained SVM classifier.

[0016] The purpose of the present invention can also be achieved by the following technical measures:

[0017] In step 1, data is collected through field experiments of blasting seismic waves, and the collected data are segmented and preprocessed to build a database.

[0018] In step 1, the oilfield geological information data and common oil-free geological signals are collected by artificial blasting data collection at the oilfield site, and the collected data are segmented and preprocessed to build a database.

[0019] In step 2, the main network structure consists of four SE-ResNeXt structures, including residual structures and SE blocks; ResNeXt is equivalent to a simplified version of InceptionNet, the difference is that the topological structure of the branches of ResNeXt is the same, and the inter-layer residual jump connection is added at the same time; the SE module mainly includes two parts, compression and excitation, which explicitly model the weights on each channel, and then weight the original feature map so that each channel has different importance; the entire network is mainly divided into four parts: input layer, convolution layer, pooling layer, and output layer.

[0020] Step 2 includes:

[0021] Step 21, the original 1D vibration signal is directly used as input and sent to the convolutional layer for adaptive feature extraction;

[0022] Step 22, apply global average pooling to reduce data dimensions and training parameters;

[0023] In step 23, the network uses softmax as a classifier.

[0024] In step 21, a one-dimensional convolution kernel is used to replace the original two-dimensional convolution kernel. The one-dimensional convolution layer can be specifically expressed as:

[0025]

[0026] Where: l represents the number of network layers; M represents the input x l-1 The number of channels; ω i represents the weight of the convolution kernel of the i-th layer; b represents the bias function; δ is a nonlinear activation function, the purpose of which is to make the model approximate any nonlinear function; * represents the convolution operation.

[0027] In step 21, to avoid the gradient vanishing problem, the residual features on the branch are scaled before being added; as a squeezing operation, the global average pooling layer uses two fully connected layers to form a bottleneck structure to simulate the correlation between channels and outputs the same number of weights as the input features.

[0028] In step 21, the SE-ResNeXt structure can be expressed as:

[0029]

[0030] Where x is the input, y is the output, L is the topology, and G is the number of branches.

[0031] In step 22, global average pooling is applied to reduce the data dimension and training parameters, which is defined as follows:

[0032]

[0033] Among them, P is the output result after pooling; Avg means taking the average of all elements in the entire feature map and outputting it to the next layer; K is the pooling kernel size; C l+1 is the output; s is the pooling kernel compensation; L is the pooling kernel size; a l (n) is the lth neuron.

[0034] In step 23, the network uses softmax as a classifier, which is expressed as:

[0035]

[0036] where z i The output value of the i-th node, C is the number of output nodes, that is, the number of categories of the classification; the output value of the multi-classification can be converted into a probability distribution in the range of [0,1] and 1 through the Softmax function; the Adam optimizer is used as the training optimizer, and the cross entropy is used as the loss function; the cross entropy is defined as:

[0037]

[0038] r is the true value; is an estimated value.

[0039] In step 4, the feature vector before the softmax layer is saved and used as the input of the SVM classifier for classification and recognition. The SVM classifier is used in the final classification step to improve the classification ability; the SVM classifier is trained using the feature vector saved by the convolutional neural network training.

[0040] In step 5, the data to be identified is first subjected to feature extraction through the trained 1DSE-ResNeXt network, and then the final identification result is obtained through the trained SVM classifier.

[0041] The method for optimizing and extracting geological information in the field of seismic wave exploration in the present invention is applied to the pattern recognition algorithm of oil and gas field exploration, and belongs to the field of seismic wave exploration of distributed acoustic wave sensors. When DAS is applied to fields such as oil and gas field exploration, due to the very high sensitivity of optical fiber, the complex exploration object, the changeable environment and many other reasons. Therefore, an effective recognition algorithm is constructed to effectively distinguish the category of geological information, so as to obtain more accurate resource information. The 1DSE-ResNeXt recognition algorithm proposed in the present invention combines one-dimensional convolution with ResNeXt, combines the SE structure, introduces inter-layer jump connection, directly adds the unit input to the unit output, and then activates it. This operation effectively improves the gradient dissipation problem in the process of network back propagation, thereby making it easy to train deep networks and effectively extract deep-level features of data. The SE structure has an attention mechanism of squeezing and excitation blocks that improve the classification ability by screening feature information without being affected by non-critical information, and the support vector machine classifier can further improve the classification accuracy.

[0042] The 1DSE-ResNeXt recognition algorithm proposed in the present invention uses a strong classifier SVM to replace the softmax layer used for classification in the original network, thereby improving the recognition accuracy of oil and gas exploration. The present invention provides a new reference for the application of distributed optical fiber sensing in the field of oil and gas exploration and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is an overall structural diagram of 1DSE-ResNeXt in a specific embodiment of the present invention;

[0044] Figure 2 A schematic diagram of a CCRN network structure in a specific embodiment of the present invention;

[0045] Figure 3 The present invention is a flowchart of a specific embodiment of the method for optimizing and extracting geological information in the field of seismic wave exploration. DETAILED DESCRIPTION

[0046] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0048] like Figure 3 As shown, Figure 3 The flowchart of the method for optimizing and extracting geological information in the field of seismic wave exploration of the present invention. The method for optimizing and extracting geological information in the field of seismic wave exploration comprises:

[0049] Step 101, collect data through on-site experiments of blasting seismic waves, segment and pre-process the collected data, and build a database.

[0050] Step 102, build a 1DSE-ResNeXt network model and train the network.

[0051] Step 103, save the feature vector output before the network softmax layer.

[0052] Step 104, input the feature vector output in step 104 into SVM to train the SVM classifier.

[0053] Step 105, the data to be identified is first subjected to feature extraction through the trained 1DSE-ResNeXt network, and then subjected to the trained SVM classifier to obtain the final identification result.

[0054] The following are several specific embodiments of the present invention.

[0055] Example 1

[0056] In a specific embodiment 1 of the present invention, the method for optimizing the extraction of geological information in the field of seismic wave exploration includes the following steps:

[0057] (1) Data is collected through field experiments of blasting seismic waves, and the collected data is segmented and preprocessed to build a database.

[0058] (2) Build a 1DSE-ResNeXt network model and train the network.

[0059] (3) Save the feature vector output before the network softmax layer.

[0060] (4) Input the feature vector output in step (3) into SVM and train the SVM classifier.

[0061] (5) The data to be identified is firstly subjected to feature extraction through the trained 1DSE-ResNeXt network, and then the final identification result is obtained through the trained SVM classifier.

[0062] The step (1) is data collection, segmentation and preprocessing of the collected data, and construction of a database. The oilfield geological information data and common oil-free geological signals are collected by artificial blasting data collection at the oilfield site.

[0063] The step (2) is to build a 1DSE-ResNeXt network and perform training. The 1DSE-ResNeXt network model mainly consists of four parts, which will be described in detail below:

[0064] A. The main structure of the network consists of four SE-ResNeXt structures, including residual structures and SE blocks. ResNeXt is equivalent to a simplified version of InceptionNet. The difference is that the topological structure of the branches of ResNeXt is the same, and the inter-layer residual jump connection is added at the same time. The SE module mainly includes two parts: compression and excitation. The weights on each channel are explicitly modeled, and then the original feature map is weighted so that each channel has different importance. The entire network is mainly divided into four parts: input layer, convolution layer, pooling layer, and output layer.

[0065] B. The original 1D vibration signal is directly used as input and sent to the convolution layer for adaptive feature extraction. Using a one-dimensional convolution kernel instead of the original two-dimensional convolution kernel, the one-dimensional convolution layer can be specifically expressed as:

[0066]

[0067] Where: l represents the number of network layers; M represents the input x l-1 The number of channels; ω i represents the weight of the convolution kernel of the i-th layer; b represents the bias function; δ is a nonlinear activation function, the purpose of which is to make the model approximate any nonlinear function; * represents the convolution operation.

[0068] To avoid the gradient vanishing problem, the residual features on the branch are scaled before being added. As a squeezing operation, the global average pooling layer uses two fully connected layers to form a bottleneck structure to model the correlation between channels and output the same number of weights as the input features. Compared with direct fully connected layers, these operations have the advantage of more nonlinearity, allowing better fitting of complex corrective relationships between channels and significantly reducing the number of parameters and calculations.

[0069] The SE-ResNeXt structure can be expressed as:

[0070]

[0071] Where x is the input, y is the output, L is the topology, and G is the number of branches.

[0072] C. Apply global average pooling to reduce data dimensions and training parameters, defined as follows:

[0073]

[0074] Among them, C l+1 is the output; s is the pooling kernel compensation; L is the pooling kernel size; a l (n) is the lth neuron.

[0075] D. The network uses softmax as a classifier, and its expression is:

[0076]

[0077] where z i The output value of the i-th node, C is the number of output nodes, that is, the number of categories. The Softmax function can be used to convert the output values ​​of multiple categories into a probability distribution ranging from [0,1] and 1. The Adam optimizer is used as the training optimizer, and the cross entropy is used as the loss function. The cross entropy is defined as:

[0078]

[0079] r is the true value; For estimated value

[0080] In the step (4), the feature vector before the softmax layer is saved and used as the input of the SVM classifier for classification and recognition. In the present invention, the SVM classifier is used in the final classification step to improve the classification ability. The SVM classifier is trained using the feature vector saved by the convolutional neural network training. The input data to be tested is first subjected to feature extraction by the trained SE-ResNeXt network to output the feature vector, and then the final recognition result is output by the trained SVM classifier.

[0081] For the specific network structure diagram, see Figure 1 .

[0082] The overall process of 1DSE-ResNeXt+SVM recognition algorithm is shown in Figure 2 .

[0083] The basic structure of 1DSE-ResNeXt is as follows Figure 1 As shown in

[15] , it is based on the residual block in ResNeXt, but adds a new “Squeeze-and-Excitation (SE)” mechanism to enhance the learning ability of feature importance.

[0084] In the SE mechanism, the output of each residual block is fed into a small neural network called the SE module, which adaptively scales the input features by learning their importance. Specifically, the SE module contains two fully connected layers, where the first layer compresses the input feature vectors to a small size and then scales them back to their original size through a sigmoid activation function. In this way, each feature dimension is assigned a weight that is adjusted according to its relative importance.

[0085] The 1DSE-ResNeXt network uses multiple residual blocks and SE modules to build the entire model. Each residual block consists of multiple parallel branches that use a concept similar to "cardinality" in ResNeXt to improve the effect of the model. At the same time, each branch also contains an SE module to further enhance the learning ability of the features.

[0086] Figure 2 The specific operation process of this network model is to first collect seismic wave signal data through the DAS system, pre-process the collected data and divide it into training set, validation set and test set. The training set is used to train the 1DSE-ResNeXt network, and the validation set is used to update the network parameters. Finally, the trained 1DSE-ResNeXt network is obtained. The feature vector obtained by convolution before the network softmax layer is output and saved to train the SVM classifier, and finally the test machine to be identified is input to the trained SVM classifier to obtain the final recognition result.

[0087] Example 2

[0088] In the specific embodiment 2 of the present invention, the specific training process of the SVM in step 104 is described in detail. The sample feature information of the sample vibration signal is input into the initial support vector machine classifier, and the second event recognition result output by the initial support vector machine classifier is obtained; according to the second event recognition result and the event label, the second training loss value of the initial type prediction network is calculated; according to the second training loss value and the second convergence condition, the initial type prediction network is trained for event type prediction.

[0089] Among them, the second convergence condition can be the same as the first convergence condition, or can be better than the first convergence condition, and the embodiment of the present application does not limit this.

[0090] As an example, a Radial Basis Function (RBF) is used in the initial support vector machine classifier training process.

[0091] Among them, the function expression of the radial basis kernel function is shown in the following formula:

[0092]

[0093] In the formula, K(x,x′) is defined as a monotonic function of the Euclidean distance between any point x in space and a center x′, ||xx′|| 2 It can be regarded as the squared Euclidean distance between two feature vectors; σ is a free parameter. During training, the number of iterations is set to infinite until the optimal recognition rate is found.

[0094] Example 3

[0095] In the specific embodiment 3 of the present invention, a preprocessing stage after data collection in step 101 is described in detail, and the data first arrival time is confirmed by the endpoint detection method and segmentation preprocessing is performed. The signal range is determined by the endpoint detection method, and the specific process is as follows:

[0096] ① Smoothing

[0097] Signals often contain noise. In order to remove the noise and retain the main information of the signal, the signal can be smoothed. There are many methods for smoothing, one of which is to use a moving average filter. The moving average filter can smooth the signal by calculating the mean of the signal within a certain time window. For a signal of length N, let the smoothed signal be y(n), and the expression of the moving average filter is:

[0098] y(n)=(1 / M)*sum(x(n-M+1:n))

[0099] Among them, x(n-M+1:n) represents the subsequence from n-M+1 to n of the signal x, and M is the filter window length.

[0100] ②Short-time energy calculation

[0101] The signal is divided into a series of time windows, and the energy of the signal in each time window is calculated. The energy of the signal is defined as the sum of the squares of the signal. For a signal of length N, let the i-th time window be xi(n), the window length be L, and the calculation formula of the energy Ei is:

[0102] Ei=sum(x_i(n)^2)

[0103] Where i = 1, 2, 3, ..., N / L. By analyzing the changes in the energy curve, the mutation and change points of the signal are detected, thereby determining the start and end positions of the signal.

[0104] ③ Calculation of short-time autocorrelation function

[0105] The signal is divided into a series of time windows, and the autocorrelation function of the signal in each time window is calculated. The autocorrelation function of a signal is defined as the inner product of the signal and itself at different time delays. For a signal of length N, let the i-th time window be xi(n), the window length be L, and the calculation formula of the autocorrelation function Ri(k) is:

[0106] Ri(k)=sum(x_i(n) * x_i(nk))

[0107] Where i = 1, 2, 3, ..., N / L, k = 0, 1, 2, ..., L-1. By analyzing the changes in the autocorrelation function curve, the mutation and change points of the signal are detected, thereby determining the start and end positions of the signal.

[0108] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0109] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A method for optimizing and extracting geological information in the field of seismic wave exploration, characterized in that: The method for optimizing the extraction of geological information in the field of seismic wave exploration includes: Step 1: Build a database; Step 2: Build a 1DSE-ResNeXt network model and train the network; Step 3, save the feature vector output before the network softmax layer; Step 4, input the feature vector output in step 3 into SVM to train the SVM classifier; Step 5: Get the final recognition result through the trained SVM classifier.

2. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 1, characterized in that: In step 1, data is collected through field experiments of blasting seismic waves, and the collected data are segmented and preprocessed to build a database.

3. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 2, characterized in that: In step 1, the oilfield geological information data and common oil-free geological signals are collected by artificial blasting data collection at the oilfield site, and the collected data are segmented and preprocessed to build a database.

4. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 1, characterized in that: In step 2, the main network structure consists of four SE-ResNeXt structures, including residual structures and SE blocks; ResNeXt is equivalent to a simplified version of InceptionNet, the difference is that the topological structure of the branches of ResNeXt is the same, and the inter-layer residual jump connection is added at the same time; the SE module mainly includes two parts, compression and excitation, which explicitly model the weights on each channel, and then weight the original feature map so that each channel has different importance; the entire network is mainly divided into four parts: input layer, convolution layer, pooling layer, and output layer.

5. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 4, characterized in that: Step 2 includes: Step 21, the original 1D vibration signal is directly used as input and sent to the convolutional layer for adaptive feature extraction; Step 22, apply global average pooling to reduce data dimensions and training parameters; In step 23, the network uses softmax as a classifier.

6. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 5, characterized in that: In step 21, a one-dimensional convolution kernel is used to replace the original two-dimensional convolution kernel. The one-dimensional convolution layer can be specifically expressed as: Where: l represents the number of network layers; M represents the input x l-1 The number of channels; ω i represents the weight of the convolution kernel of the i-th layer; b represents the bias function; δ is a nonlinear activation function, the purpose of which is to make the model approximate any nonlinear function; * represents the convolution operation.

7. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 6, characterized in that: In step 21, to avoid the gradient vanishing problem, the residual features on the branch are scaled before adding; As a squeezing operation, the global average pooling layer uses two fully connected layers to form a bottleneck structure to model the correlation between channels and outputs the same number of weights as the input features.

8. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 7, characterized in that: In step 21, the SE-ResNeXt structure can be expressed as: Where x is the input, y is the output, L is the topology, and G is the number of branches.

9. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 5, characterized in that: In step 22, global average pooling is applied to reduce the data dimension and training parameters, which is defined as follows: Among them, P is the output result after pooling; Avg means taking the average of all elements in the entire feature map and outputting it to the next layer; K is the pooling kernel size; C l+1 is the output; s is the pooling kernel compensation; L is the pooling kernel size; a l (n) is the lth neuron.

10. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 5, characterized in that: In step 23, the network uses softmax as a classifier, which is expressed as: Where z i The output value of the i-th node, C is the number of output nodes, that is, the number of categories of the classification; the output value of the multi-classification can be converted into a probability distribution in the range of [0,1] and 1 through the Softmax function; the Adam optimizer is used as the training optimizer, and the cross entropy is used as the loss function; the cross entropy is defined as: r is the true value; is an estimated value.

11. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 1, characterized in that: In step 4, the feature vector before the softmax layer is saved and used as the input of the SVM classifier for classification and recognition. The SVM classifier is used in the final classification step to improve the classification ability; the SVM classifier is trained using the feature vector saved by the convolutional neural network training.

12. The method for optimizing and extracting geological information in the field of seismic wave exploration according to claim 1, characterized in that: In step 5, the data to be identified is first subjected to feature extraction through the trained 1DSE-ResNeXt network, and then the final identification result is obtained through the trained SVM classifier.

Images