A seismic phase picking method and system based on multi-feature fusion

Through multi-feature fusion technology, the problem of low seismic phase picking accuracy in the existing technology is solved by using methods such as semi-channel convolution, dynamic convolution and channel compression modules, and the problem of low seismic phase picking accuracy is significantly improved.

CN118169752BActive Publication Date: 2024-11-08BEIJING INSTITUTE OF PETROCHEMICAL TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410285818.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-13
Publication Date
2024-11-08
Estimated Expiration
2044-03-13

AI Technical Summary

Technical Problem

Existing deep learning-based seismic phase picking models perform well in feature extraction, but fail to fully utilize the features, resulting in poor phase picking accuracy.

Method used

The multi-feature fusion method is adopted, and the semi-channel convolution module, dynamic convolution module and channel compression module are combined with the characteristics of multiple scales to optimize the processing capacity of seismic data and improve the phase picking accuracy.

Benefits of technology

The deep learning model's attention to the global attributes of seismic data has been improved, and the characteristics of each stage are fully utilized, which significantly improves the accuracy of seismic phase pickup.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118169752B_ABST
    Figure CN118169752B_ABST
Patent Text Reader

Abstract

The present invention discloses a seismic phase picking method and system based on multi-feature fusion, comprising the following steps: obtaining time series data through a seismic detector. Inputting the time series data into an encoding module, and using a convolution operation to obtain shallow features of the data. Inputting the shallow features into a semi-channel convolution module, and performing different operations on different parts of the shallow features to obtain deep features. The deep features are respectively input into a dynamic convolution module and a channel compression module to obtain high-level features and compressed features. Finally, the deep features, high-level features, and compressed features are input into a multi-feature fusion module, and the fused features are input into a decoding module, and finally the seismic phase picking results are obtained through a softmax function. The present invention can improve the deep learning model's attention to the global attributes of seismic data, make full use of the features extracted at each stage, not only optimize the processing capability of seismic data features, but also improve the accuracy of seismic phase picking.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of earthquake event detection, and in particular relates to an earthquake phase picking method and system based on multi-feature fusion. Background Art

[0002] Deep learning technology has been widely used in various subfields of seismology, especially in seismic phase picking tasks to solve the problems of earthquake source location and microseismic detection. Existing research based on deep learning only focuses on feature extraction, but does not make full use of features, and is prone to losing key information during feature decoding, resulting in poor phase picking accuracy.

[0003] The idea of ​​multi-feature fusion was first applied in the field of image classification and has developed rapidly in recent years. Existing deep learning models for phase picking perform well in feature extraction, but they are not able to make good use of these features. Multi-feature fusion can combine features at multiple scales to improve the accuracy of phase picking. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a seismic phase picking method and system based on multi-feature fusion, which can improve the deep learning model's attention to the global attributes of seismic data and make full use of the features extracted at each stage. It not only optimizes the processing capability of seismic data features, but also improves the accuracy of seismic phase picking.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A seismic phase picking method based on multi-feature fusion includes the following steps:

[0007] Collect seismic signals through seismic detectors to obtain time series data;

[0008] Input the time series data into the encoding module, and use convolution operation to obtain shallow features of the data;

[0009] The shallow features are input into a half-channel convolution module, and different operations are performed on different parts of the shallow features to obtain deep features;

[0010] Inputting the deep features into a dynamic convolution module and a channel compression module respectively to obtain high-level features and compressed features;

[0011] Inputting the deep features, the high-level features and the compressed features into a multi-feature fusion module to obtain fused features;

[0012] The fused features are input into a decoding module, and the fused features are processed by a deconvolution operation to obtain a final phase picking result.

[0013] Preferably, the method of inputting the shallow features into a half-channel convolution module and performing different operations on different parts of the shallow features to obtain deep features includes:

[0014] Input the shallow features into a half-channel convolution module, and divide the shallow features into four groups of tensors on average along the channel dimension;

[0015] Perform different processing operations on the four groups of tensors respectively, wherein the first group of tensors are sequentially subjected to depthwise separable convolution, GELU, and point convolution operations; the second group of tensors are sequentially subjected to point convolution, GELU, and depthwise separable convolution operations; the third group of tensors are sequentially subjected to dilated convolution with an expansion rate of 5, GELU, and dilated convolution with an expansion rate of 5; and the fourth group of tensors are subjected to identity mapping;

[0016] The four sets of tensor processed features are concatenated to obtain deep features.

[0017] Preferably, the method of concatenating the four sets of tensor-processed features to obtain deep features includes:

[0018] x1,x2,x3,x4=split(x)

[0019] x'1=DW(GELU(PW(x1)))

[0020] x'2=PW(GELU(DW(x2)))

[0021] x'3=Dilated(GELU(Dilated(x3)))

[0022] x'4=Identity(x4)

[0023] x'=Cat(x'1,x'2,x'3,x'4),

[0024] Among them, x represents shallow features; split() means dividing the tensor into four groups along the number of channels; X1 represents the first group of tensors; X2 represents the second group of tensors; X3 represents the third group of tensors; X4 represents the fourth group of tensors; DW() represents depthwise separable convolution; PW() represents point convolution; Dilated() represents dilated convolution; GELU() represents GELU activation function; X1′ represents the first group of tensors after processing, X2′ represents the second group of tensors after processing, X3′ represents the third group of tensors after processing, X4′ represents the fourth group of tensors after processing, and X′ represents the deep features output by the half-channel convolution module.

[0025] Preferably, the method of inputting the deep features into the dynamic convolution module and the channel compression module to obtain high-level features and compressed features includes:

[0026] The deep features are input into the dynamic convolution module and passed through four parallel branches, wherein the first branch is a convolution layer with a convolution kernel size of 1; the second branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 3 and a one-dimensional convolution layer with a convolution kernel size of 1; the third branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and a one-dimensional convolution layer with a convolution kernel size of 3; the fourth branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and two one-dimensional convolution layers with a convolution kernel size of 3;

[0027] The output results of the four branches are spliced ​​to obtain high-level features;

[0028] Inputting the deep features into a channel compression module, and applying a deconvolution operation to upsample the deep features;

[0029] Based on the results of deconvolution, 1×1 convolution is used to perform cross-channel information interaction to obtain compressed features.

[0030] Preferably, the method of inputting the deep features, the high-level features and the compressed features into a multi-feature fusion module to obtain the fused features includes:

[0031] Divide the deep features into four groups {a1, a2, a3, a4} along the channel dimension, divide the high-level features into four groups {b1, b2, b3, b4} along the channel dimension, and call the compressed features c;

[0032] Combine the four groups {a1, a2, a3, a4}, the four groups {b1, b2, b3, b4} and the compressed feature c into four new tensors {a1, b1, c}, {a2, b2, c}, {a3, b3, c}, {a4, b4, c} respectively according to the splitting order;

[0033] Perform dilated convolution operations on the four new tensors {a1,b1,c}, {a2,b2,c}, {a3,b3,c}, and {a4,b4,c} with expansion rates of 7, 5, 3, and 1 respectively;

[0034] The tensors after the four dilated convolution operations are concatenated, and the concatenated features are subjected to a one-dimensional convolution operation with a convolution kernel size of 3 to obtain the fused features.

[0035] Preferably, the expression for obtaining the fused features is:

[0036] a1,a2,a3,a4=split(a)

[0037] b1,b2,b3,b4=split(b)

[0038] t1=Diated(Cat(a1,b1,c))

[0039] t2=Diated(Cat(a2,b2,c))

[0040] t3=Diated(Cat(a3,b3,c))

[0041] t4=Diated(Cat(a4,b4,c))

[0042] t=Conv1D(Cat(t1,t2,t3,t4)),

[0043] Among them, a, b, c, and t represent deep features, high-level features, compressed features, and fused features, respectively; a1 and b1 represent the first group of tensors after segmentation, respectively; a2 and b2 represent the second group of tensors after segmentation, respectively; a3 and b3 represent the third group of tensors after segmentation, respectively; a4 and b4 represent the fourth group of tensors after segmentation, respectively; t1 is the concatenated feature after the dilated convolution with an expansion rate of 7, t2 is the concatenated feature after the dilated convolution with an expansion rate of 5, t3 is the concatenated feature after the dilated convolution with an expansion rate of 3, and t4 is the concatenated feature after the dilated convolution with an expansion rate of 1; Conv1D represents a one-dimensional convolution operation.

[0044] Preferably, the method of inputting the fused features into a decoding module and processing the fused features by a deconvolution operation to obtain a final phase picking result includes:

[0045] The fused features are input into the decoding module and upsampled through three deconvolution layers, wherein the first deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 7, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer; the second deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 5, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer; the third deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 3, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer, to obtain the final seismic phase picking result.

[0046] The present invention also provides a seismic phase picking system based on multi-feature fusion, comprising: a time series data acquisition module, a shallow feature acquisition module, a deep feature acquisition module, an advanced feature and compression feature acquisition module, a fusion feature acquisition module, and a phase picking module;

[0047] The time series data acquisition module is used to collect seismic signals through seismic detectors to obtain time series data;

[0048] The shallow feature acquisition module is used to input the time series data into the encoding module and obtain the shallow features of the data using a convolution operation;

[0049] The deep feature acquisition module is used to input the shallow features into the half-channel convolution module, and perform different operations on different parts of the shallow features to obtain deep features;

[0050] The high-level feature and compression feature acquisition module is used to input the deep features into the dynamic convolution module and the channel compression module respectively to obtain high-level features and compression features;

[0051] The fusion feature acquisition module is used to input the deep features, the high-level features and the compression features into a multi-feature fusion module to obtain fused features;

[0052] The phase picking module is used to input the fused features into the decoding module, and process the fused features using a deconvolution operation to obtain a final phase picking result.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention proposes a semi-channel convolution module to more efficiently process the features extracted by the encoding module; extracts high-level features through a dynamic convolution module; compresses features and obtains the relationship between channels through a channel compression module; and fuses the feature mapping output at each stage through a multi-feature fusion module. The method of the present invention can improve the deep learning model's attention to the global attributes of seismic data, fully utilize the features extracted at each stage, not only optimize the processing capability of seismic data features, but also improve the accuracy of seismic phase picking. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 It is a schematic diagram of the steps of a seismic phase picking method based on multi-feature fusion in an embodiment of the present invention;

[0057] Figure 2 is a schematic diagram of an encoding module in an embodiment of the present invention;

[0058] Figure 3 is a schematic diagram of a half-channel convolution module in an embodiment of the present invention;

[0059] Figure 4is a schematic diagram of a dynamic convolution module in an embodiment of the present invention;

[0060] Figure 5 is a schematic diagram of a channel compression module in an embodiment of the present invention;

[0061] Figure 6 is a schematic diagram of a multi-feature fusion module in an embodiment of the present invention;

[0062] Figure 7 It is a schematic diagram of a decoding module in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Embodiment 1

[0066] like Figure 1 As shown, in this embodiment, a flow chart of a seismic phase picking method based on multi-feature fusion is provided, which includes the following steps:

[0067] Collect the required time series data through seismometers;

[0068] Input the collected time series data into the encoding module to extract shallow features;

[0069] The extracted shallow features are input into the half-channel convolution module to obtain deep features;

[0070] The acquired deep features are input into the dynamic convolution module and the channel compression module to obtain high-level features and compressed features respectively;

[0071] Finally, the deep features, high-level features, and compressed features are input into the multi-feature fusion module to obtain the fused features. Finally, the fused features are passed through the softmax function to obtain the final seismic phase picking results.

[0072] Encoding module such as Figure 2As shown in the figure, the time series data acquired by the seismic detector is processed using the convolution operation, and the intrinsic feature representation contained in the time series data is extracted in a stage-by-stage process. Each stage of the encoding module can enhance the expressiveness of the model, help the model better learn the intrinsic laws and patterns of the data, and thus improve the performance of the model, because a single stage of convolution operation is not enough to capture all the information of the data.

[0073] The encoding module consists of five stages, and the number of channels in each stage gradually increases, namely {8, 16, 24, 32, 48}. The first stage is a one-dimensional convolution with a convolution kernel of 7 connected to a batch normalization layer and a GELU activation layer; the second stage is a one-dimensional convolution with a convolution kernel of 5 connected to a batch normalization layer and a GELU activation layer; the third stage is a one-dimensional convolution with a convolution kernel of 3 connected to a batch normalization layer and a GELU activation layer; the fourth stage is a one-dimensional convolution with a convolution kernel of 5 connected to a batch normalization layer and a GELU activation layer; the fifth stage is a one-dimensional convolution with a convolution kernel of 7 connected to a batch normalization layer and a GELU activation layer. The output of the time series data obtained by the seismic detector after the encoding module is the shallow feature. The shallow feature refers to the preliminary feature representation obtained after the time series data obtained by the seismic detector is processed by the encoding module. It mainly captures the local details of the seismic data and provides an important basis for subsequent processing and analysis.

[0074] The shallow features are specifically expressed as:

[0075] r1=GELU(BatchNorm(Conv1D(r,kernel=7)))

[0076] r2=GELU(BatchNorm(Conv1D(r1,kernel=5)))

[0077] r3=GELU(BatchNorm(Conv1D(r2,kernel=3)))

[0078] r4=GELU(BatchNorm(Conv1D(r3,kernel=5)))

[0079] r'=GELU(BatchNorm(Conv1D(r4,kernel=7)))

[0080] Among them, r represents the time series data obtained by the seismic detector, r1 represents the output of the first stage, r2 represents the output of the second stage, r3 represents the output of the third stage, r4 represents the output of the fourth stage, and r' represents the shallow features of the output of the encoding module; Conv1D() represents a one-dimensional convolution operation, where kernel represents the corresponding convolution kernel size; BatchNorm() represents a batch normalization operation; GELU() represents a GELU activation function.

[0081] The half-channel convolution module is as follows Figure 3 As shown in the figure, the shallow features output by the encoding module are evenly divided into four groups of tensors {x1, x2, x3, x4} along the channel dimension, and different operations are performed on these four groups of tensors. The design of the semi-channel convolution module is to introduce more flexibility and diversity into the neural network, thereby enhancing the model's expressive power and feature extraction capabilities. By grouping the input shallow features along the channel dimension and performing different operations on each group, the model can learn the differences between different groups and better capture the complex structures and patterns in the shallow features.

[0082] The half-channel convolution module performs depthwise separable convolution, GELU, and pointwise convolution operations on the first group of grouped tensors x1; performs pointwise convolution, GELU, and depthwise separable convolution operations on the second group of tensors x2; performs dilated convolution with an expansion rate of 5, GELU, and dilated convolution with an expansion rate of 5 on the third group of tensors x3; and performs identity mapping on the fourth group of tensors x4. The four groups of processed features are then concatenated to obtain deep features, which usually contain richer semantic information and help the model understand the abstract concepts of the data.

[0083] The formula of the half-channel convolution module is expressed as:

[0084] x1,x2,x3,x4=split(x)

[0085] x'1=DW(GELU(PW(x1)))

[0086] x'2=PW(GELU(DW(x2)))

[0087] x'3=Dilated(GELU(Dilated(x3)))

[0088] x'4=Identity(x4)

[0089] x'=Cat(x'1,x'2,x'3,x'4)

[0090] Among them, x represents the shallow features extracted by the encoding module, X′ represents the deep features output by the half-channel convolution module; split() represents dividing the tensor into four groups along the number of channels; DW() represents depthwise separable convolution; PW() represents point convolution; Dilated() represents dilated convolution; GELU() represents GELU activation function; Cat() represents tensor concatenation operation. X1′ represents the first group of tensors after processing, X2′ represents the second group of tensors after processing, X3′ represents the third group of tensors after processing, and X4′ represents the fourth group of tensors after processing. These four different groups of tensors are useful information extracted from seismic signals based on the characteristics of each group of features.

[0091] The deep features are input into the dynamic convolution module. The dynamic convolution module is as follows: Figure 4 As shown. The input deep features first pass through four parallel branches. The first branch is a convolution layer with a convolution kernel size of 1; the second branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 3 and a one-dimensional convolution layer with a convolution kernel size of 1; the third branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and a one-dimensional convolution layer with a convolution kernel size of 3; the fourth branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and two one-dimensional convolution layers with a convolution kernel size of 3. The output results of these four branches are then concatenated to obtain high-level features. The formula for high-level features is expressed as:

[0092] n1=Conv1D(n,kernel=1)

[0093] n2=Conv1D(Conv1D(n,kernel=3),kernel=1)

[0094] n3=Conv1D(Conv1D(n,kernel=1),kernel=3)

[0095] n4=Conv1D(Conv1D(Conv1D(n,kernel=1),kernel=3),kernel=3)

[0096] n'=Cat(n1,n2,n3,n4)

[0097] Among them, n represents the deep features output by the half-channel convolution module; n1, n2, n3, and n4 represent the output results of the four branches respectively; n' represents the high-level features obtained by splicing the output results of these four branches; Conv1D() represents a one-dimensional convolution operation, where kernel represents the corresponding convolution kernel size; Cat() represents a tensor splicing operation.

[0098] Compression modules such as Figure 5As shown. The deep features are input into the channel compression module. First, the deconvolution operation with a convolution kernel size of 5 is applied to upsample the input features to expand the feature map to a larger spatial size. Secondly, for the results after deconvolution, 1×1 convolution is used to perform cross-channel information interaction to obtain compressed features. The compressed features retain the key information in the deep features while reducing the redundancy of the features. The formula of the compression module is expressed as:

[0099] m1=ConvTranspose1D(m,kernel=5)

[0100] m'=Conv1D(m1,kernel=1)

[0101] Among them, m represents the deep features output by the half-channel convolution module; m1 represents the compressed features after cross-channel information interaction; and m' represents the compressed features output by the compression module.

[0102] The multi-feature fusion module accepts three inputs: deep features, high-level features, and compressed features. The multi-feature fusion module is used to overcome the limitations of a single feature. Since a single feature may be interfered by factors such as noise, resulting in a decrease in the performance of the model phase picking, the insufficiency of a single feature can be compensated by fusing features of multiple scales to improve the robustness of the model. Figure 6 As shown, first, the multi-feature fusion module divides the deep features and high-level features into four groups along the channel dimension, where the deep features are divided into four groups {a1, a2, a3, a4} along the channel dimension, and the high-level features are divided into four groups {b1, b2, b3, b4} along the channel dimension. The compressed features are called c, and the compressed features are sequentially concatenated with a group of shallow features and a group of high-level features, that is, they are combined into four new tensors {a1, b1, c}, {a2, b2, c}, {a3, b3, c}, and {a4, b4, c} in the order of splitting. Secondly, the dilated convolution with a convolution kernel size of 3 and an expansion rate of {7, 5, 3, 1} is applied to different groups to obtain four groups of fused features to extract information of different scales. Then, these four groups of fused features are concatenated along the channel dimension, and a one-dimensional convolution operation with a convolution kernel of 3 is performed to interact the global features with the local features to obtain the fused features. The formula of the multi-feature fusion module is as follows:

[0103] a1,a2,a3,a4=split(a)

[0104] b1,b2,b3,b4=split(b)

[0105] t1=Diated(Cat(a1,b1,c))

[0106] t2=Diated(Cat(a2,b2,c))

[0107] t3=Diated(Cat(a3,b3,c))

[0108] t4=Diated(Cat(a4,b4,c))

[0109] t=Conv1D(Cat(t1,t2,t3,t4))

[0110] Among them, a represents the deep features output by the half-channel convolution module, b represents the high-level features output by the dynamic convolution module, c represents the compressed features output by the channel compression module, and t represents the fused features output by the multi-feature fusion module; a1 and b1 represent the first group of tensors after segmentation respectively; a2 and b2 represent the second group of tensors after segmentation respectively; a3 and b3 represent the third group of tensors after segmentation respectively; a4 and b4 represent the fourth group of tensors after segmentation respectively; t1 is the concatenated feature after the dilated convolution with an expansion rate of 7, t2 is the concatenated feature after the dilated convolution with an expansion rate of 5, t3 is the concatenated feature after the dilated convolution with an expansion rate of 3, and t4 is the concatenated feature after the dilated convolution with an expansion rate of 1; Conv1D represents a one-dimensional convolution operation.

[0111] Finally, the fused features are input into the decoding module, such as Figure 7 As shown in the figure, the decoding module upsamples the input through three deconvolution layers. The first deconvolution layer is a stack of transposed convolution with a kernel size of 7 and a stride of 1 and a batch normalization layer; the second deconvolution layer is a stack of transposed convolution with a kernel size of 5 and a stride of 1 and a batch normalization layer; the third deconvolution layer is a stack of transposed convolution with a kernel size of 3 and a stride of 1 and a batch normalization layer. The role of these three deconvolution layers is that the detailed information of the seismic data will be lost due to multiple downsampling operations. The deconvolution layer in the decoding module can restore the lost detailed information of the seismic data and improve the accuracy of phase picking. The output result of the decoding module is passed through the softmax function to obtain the final seismic phase picking result. The formula of the decoding module is as follows:

[0112] h1=BatchNorm(ConvTranspose1D(h,kernel=7))

[0113] h2=BatchNorm(ConvTranspose1D(h1,kernel=5))

[0114] h3=BatchNorm(ConvTranspose1D(h2,kernel=3))

[0115] h'=Softmax(h3))

[0116] Among them, h represents the fusion feature output by the multi-feature fusion module; h1, h2, and h3 represent the deconvolution features; and h' represents the final seismic phase picking result.

[0117] Embodiment 2

[0118] The present invention also provides a seismic phase picking system based on multi-feature fusion, comprising: a time series data acquisition module, a shallow feature acquisition module, a deep feature acquisition module, an advanced feature and compression feature acquisition module, a fusion feature acquisition module, and a phase picking module;

[0119] The time series data acquisition module is used to collect seismic signals through seismic detectors to obtain time series data;

[0120] The shallow feature acquisition module is used to input the time series data into the encoding module and use the convolution operation to obtain the shallow features of the data;

[0121] The deep feature acquisition module is used to input the shallow features into the half-channel convolution module, and perform different operations on different parts of the shallow features to obtain the deep features;

[0122] The high-level feature and compression feature acquisition module is used to input the deep features into the dynamic convolution module and the channel compression module respectively to obtain high-level features and compression features;

[0123] The fusion feature acquisition module is used to input deep features, high-level features and compressed features into the multi-feature fusion module to obtain fused features;

[0124] The phase picking module is used to input the fused features into the decoding module, and process the fused features using a deconvolution operation to obtain a final phase picking result.

[0125] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A seismic phase picking method based on multi-feature fusion, characterized in that: The following steps are involved: Collect seismic signals through seismic detectors to obtain time series data; Input the time series data into the encoding module, and use convolution operation to obtain shallow features of the data; The shallow features are input into a half-channel convolution module, and different operations are performed on different parts of the shallow features to obtain deep features; Inputting the deep features into a dynamic convolution module and a channel compression module respectively to obtain high-level features and compressed features; Inputting the deep features, the high-level features and the compressed features into a multi-feature fusion module to obtain fused features; The fused features are input into a decoding module, and the fused features are processed by a deconvolution operation to obtain a final phase picking result; The method of inputting the shallow features into a half-channel convolution module and performing different operations on different parts of the shallow features to obtain deep features includes: Input the shallow features into a half-channel convolution module, and divide the shallow features into four groups of tensors on average along the channel dimension; Perform different processing operations on the four groups of tensors respectively, wherein the first group of tensors are sequentially subjected to depthwise separable convolution, GELU, and point convolution operations; the second group of tensors are sequentially subjected to point convolution, GELU, and depthwise separable convolution operations; the third group of tensors are sequentially subjected to dilated convolution with an expansion rate of 5, GELU, and dilated convolution with an expansion rate of 5; and the fourth group of tensors are subjected to identity mapping; The features after the four sets of tensor processing are concatenated to obtain the deep features; Methods for concatenating the features after four sets of tensor processing to obtain deep features include: x1,x2,x3,x4=split(x) x'1=DW(GELU(PW(x1))) x'2=PW(GELU(DW(x2))) x'3=Dilated(GELU(Dilated(x3))) x'4=Identity(x4) x'=Cat(x'1,x'2,x'3,x'4), Among them, x represents shallow features; split() means dividing the tensor into four groups along the number of channels; X1 represents the first group of tensors; X2 represents the second group of tensors; X3 represents the third group of tensors; X4 represents the fourth group of tensors; DW() represents depthwise separable convolution; PW() represents point convolution; Dilated() represents dilated convolution; GELU() represents GELU activation function; X1′ represents the first group of tensors after processing, X2′ represents the second group of tensors after processing, X3′ represents the third group of tensors after processing, X4′ represents the fourth group of tensors after processing, and X′ represents the deep features output by the half-channel convolution module.

2. The seismic phase picking method based on multi-feature fusion according to claim 1, characterized in that: The method of inputting the deep features into the dynamic convolution module and the channel compression module to obtain high-level features and compressed features includes: The deep features are input into the dynamic convolution module and passed through four parallel branches, wherein the first branch is a convolution layer with a convolution kernel size of 1; the second branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 3 and a one-dimensional convolution layer with a convolution kernel size of 1; the third branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and a one-dimensional convolution layer with a convolution kernel size of 3; the fourth branch is a stack of a one-dimensional convolution layer with a convolution kernel size of 1 and two one-dimensional convolution layers with a convolution kernel size of 3; The output results of the four branches are spliced ​​to obtain high-level features; Inputting the deep features into a channel compression module, and applying a deconvolution operation to upsample the deep features; Based on the results of deconvolution, 1×1 convolution is used to perform cross-channel information interaction to obtain compressed features.

3. The seismic phase picking method based on multi-feature fusion according to claim 1, characterized in that: The method of inputting the deep features, the high-level features and the compressed features into a multi-feature fusion module to obtain fused features includes: Divide the deep features into four groups {a1, a2, a3, a4} along the channel dimension, divide the high-level features into four groups {b1, b2, b3, b4} along the channel dimension, and call the compressed features c; Combine the four groups {a1, a2, a3, a4}, the four groups {b1, b2, b3, b4} and the compressed feature c into four new tensors {a1, b1, c}, {a2, b2, c}, {a3, b3, c}, {a4, b4, c} respectively according to the splitting order; Perform dilated convolution operations on the four new tensors {a1,b1,c}, {a2,b2,c}, {a3,b3,c}, and {a4,b4,c} with expansion rates of 7, 5, 3, and 1 respectively; The tensors after the four dilated convolution operations are concatenated, and the concatenated features are subjected to a one-dimensional convolution operation with a convolution kernel size of 3 to obtain the fused features.

4. The seismic phase picking method based on multi-feature fusion according to claim 3 is characterized in that: The expression of the fused features is: a1,a2,a3,a4=split(a) b1,b2,b3,b4=split(b) t1=Diated(Cat(a1,b1,c)) t2=Diated(Cat(a2,b2,c)) t3=Diated(Cat(a3,b3,c)) t4=Diated(Cat(a4,b4,c)) t=Conv1D(Cat(t1,t2,t3,t4)), Among them, a, b, c, and t represent deep features, high-level features, compressed features, and fused features, respectively; a1 and b1 represent the first group of tensors after segmentation, respectively; a2 and b2 represent the second group of tensors after segmentation, respectively; a3 and b3 represent the third group of tensors after segmentation, respectively; a4 and b4 represent the fourth group of tensors after segmentation, respectively; t1 is the concatenated feature after the dilated convolution with an expansion rate of 7, t2 is the concatenated feature after the dilated convolution with an expansion rate of 5, t3 is the concatenated feature after the dilated convolution with an expansion rate of 3, and t4 is the concatenated feature after the dilated convolution with an expansion rate of 1; Conv1D represents a one-dimensional convolution operation.

5. The seismic phase picking method based on multi-feature fusion according to claim 1, characterized in that: The method of inputting the fused features into a decoding module and processing the fused features by a deconvolution operation to obtain a final phase picking result includes: The fused features are input into the decoding module and upsampled through three deconvolution layers, wherein the first deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 7, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer; the second deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 5, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer; the third deconvolution layer is a stack of a convolution kernel with a convolution kernel size of 3, a step size of 1, a transposed convolution with a step size of 2, and a batch normalization layer, to obtain the final seismic phase picking result.

6. A seismic phase picking system based on multi-feature fusion, characterized in that: include: Time series data acquisition module, shallow feature acquisition module, deep feature acquisition module, high-level feature and compression feature acquisition module, fusion feature acquisition module, phase picking module; The time series data acquisition module is used to collect seismic signals through seismic detectors to obtain time series data; The shallow feature acquisition module is used to input the time series data into the encoding module and obtain the shallow features of the data using a convolution operation; The deep feature acquisition module is used to input the shallow features into the half-channel convolution module, and perform different operations on different parts of the shallow features to obtain deep features; The high-level feature and compression feature acquisition module is used to input the deep features into the dynamic convolution module and the channel compression module respectively to obtain high-level features and compression features; The fusion feature acquisition module is used to input the deep features, the high-level features and the compression features into a multi-feature fusion module to obtain fused features; The phase picking module is used to input the fused features into the decoding module, and process the fused features by using a deconvolution operation to obtain a final phase picking result; The shallow features are input into the half-channel convolution module, and different operations are performed on different parts of the shallow features to obtain the deep features. The process includes: Input the shallow features into a half-channel convolution module, and divide the shallow features into four groups of tensors on average along the channel dimension; Perform different processing operations on the four groups of tensors respectively, wherein the first group of tensors are sequentially subjected to depthwise separable convolution, GELU, and point convolution operations; the second group of tensors are sequentially subjected to point convolution, GELU, and depthwise separable convolution operations; the third group of tensors are sequentially subjected to dilated convolution with an expansion rate of 5, GELU, and dilated convolution with an expansion rate of 5; and the fourth group of tensors are subjected to identity mapping; The features after the four sets of tensor processing are concatenated to obtain the deep features; The process of concatenating the features after four sets of tensor processing to obtain deep features includes: x1,x2,x3,x4=split(x) x'1=DW(GELU(PW(x1))) x'2=PW(GELU(DW(x2))) x'3=Dilated(GELU(Dilated(x3))) x'4=Identity(x4) x'=Cat(x'1,x'2,x'3,x'4), Among them, x represents shallow features; split() means dividing the tensor into four groups along the number of channels; X1 represents the first group of tensors; X2 represents the second group of tensors; X3 represents the third group of tensors; X4 represents the fourth group of tensors; DW() represents depthwise separable convolution; PW() represents point convolution; Dilated() represents dilated convolution; GELU() represents GELU activation function; X1′ represents the first group of tensors after processing, X2′ represents the second group of tensors after processing, X3′ represents the third group of tensors after processing, X4′ represents the fourth group of tensors after processing, and X′ represents the deep features output by the half-channel convolution module.