Microseismic event detection method and system based on multi-scale feature fusion

Abstract

The invention discloses a microseism event detection method and system based on multi-scale feature fusion, and the method comprises the following steps: collecting microseism data, obtaining time series data, converting the time series data into frequency domain data, and storing the frequency domain data in a spectrogram form; constructing a multi-scale feature extraction module based on a convolutional neural network, and performing feature extraction on the spectrogram by using the multi-scale feature extraction module to sequentially obtain features of different levels; and fusing the features of different levels to obtain a final seismic phase pickup result. According to the method, the multi-scale features are extracted from the spectrogram through the convolutional neural network, different levels of information in the microseismic signals are captured, and then the features are fused, so that effective integration of global features and local features is realized. The method not only optimizes the seismic data feature processing capability, but also improves the precision of microseismic phase pickup.

Description

Technical Field

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

[0002] Hydraulic fracturing technology is widely used in oilfield production increase and shale gas mining monitoring. Liquid is injected into the bottom of the well by a high-pressure pump to create cracks in the formation, thereby releasing the oil and gas resources in the reservoir. In this process, a large number of microseismic events will be generated, and the signals of these events can be collected and recorded by seismic monitoring equipment. By analyzing and processing these microseismic signals, a large amount of valuable information can be extracted to help evaluate and optimize hydraulic fracturing operations.

[0003] In practical applications, microseismic signals usually need to undergo complex analysis and processing to extract features and classify events. Traditional signal processing methods include time-frequency analysis, wavelet transform, etc., but these methods often find it difficult to extract stable and effective features when faced with complex and changeable microseismic signals. In recent years, deep learning technology has gradually demonstrated superior performance in the field of microseismic signal processing due to its powerful feature extraction and learning capabilities. However, some current studies only focus on the single feature extraction level and do not fully consider the fusion of features at different levels, which may lead to the omission of some important information and affect the accuracy of classification and recognition.

[0004] Multi-scale feature fusion technology has been widely used in the field of computer vision and has achieved remarkable results. This technology uses neural networks to extract features at different levels and fuses these features to make full use of the feature information at each level. By fusing these multi-scale features, it is possible to effectively avoid important information that may be missed by single-scale features, thereby improving the classification accuracy of microseismic data. Summary of the invention

[0005] In view of the problem that different levels of features cannot be effectively integrated in existing microseismic signal processing research, the present invention proposes a microseismic event detection method and system based on multi-scale feature fusion, which extracts multi-scale features from the spectrum graph through a convolutional neural network, captures different levels of information in the microseismic signal, and then fuses these features to achieve effective integration of global features and local features. This method not only optimizes the processing capability of seismic data features, but also improves the accuracy of microseismic phase picking.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A microseismic event detection method based on multi-scale feature fusion includes the following steps:

[0008] Acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram;

[0009] Constructing a multi-scale feature extraction module based on a convolutional neural network, and using the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn;

[0010] The features at different levels are fused, and the fusion results are input into the multi-layer perceptron to obtain the final seismic phase picking result.

[0011] Preferably, the method of extracting features from the spectrum graph using the multi-scale feature extraction module to sequentially obtain features at different levels includes:

[0012] Inputting the spectrum graph into a convolutional neural network to obtain features at different levels in sequence, wherein the convolutional neural network comprises four convolutional layers connected in sequence;

[0013] Specifically: the spectrum graph is input into the first convolution layer for preliminary feature extraction, wherein the first convolution layer uses 8 9x9 convolution kernels with a step size of 2;

[0014] The second convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch-normalized, and then convolution is performed using 16 7x7 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the first convolutional layer is added to the output of the second convolutional layer to form a residual connection. The local features in the microseismic signal are obtained and recorded as F L ;

[0015] The third convolutional layer uses 16 5x5 convolution kernels with a step size of 2 to further extract and enhance the local complex features in the microseismic signal, denoted by F M , as the input of the fourth convolutional layer;

[0016] The fourth convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch normalized, and then convolution is performed using 16 3x3 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the third convolutional layer is added to the output of the fourth convolutional layer to form a residual connection. The global feature of the microseismic signal is recorded as F H .

[0017] Preferably, the method of fusing features at different levels to obtain the final seismic phase picking result includes:

[0018] F H After processing by the upsampling module, the upsampled F H With FM Input them into the feature fusion module for fusion, and the initial fusion feature is recorded as F P ;

[0019] The initial fusion feature F P After being processed by the upsampling module, the initial fusion feature F after upsampling P With F L They are input into the feature fusion module for fusion, and the final fusion feature is recorded as F out ;

[0020] The final fusion feature F out After processing by the upsampling module, the final fusion feature F out , input into the multi-layer perceptron and mapped into a probability sequence associated with the P wave and S wave of the seismic signal at each time point, that is, the final seismic phase picking result.

[0021] Preferably, the upsampling module processing method includes:

[0022] Reshape the C channel values ​​at each position of the input feature map into a 1×C vector;

[0023] Perform a linear transformation on the 1×C vector to obtain a 1×12 vector;

[0024] The 1×12 vector is reshaped into a 2×2×3 tensor to obtain a new feature map.

[0025] Preferably, the upsampled F H With F M Input them into the feature fusion module for fusion, and the initial fusion feature is recorded as F P The methods include:

[0026] The upsampled F H Input weight calculation unit to convert into a set of weights W;

[0027] The weight W and F M Multiplying, the expression is: Get the enhanced F M ;

[0028] The enhanced F M That is F M 'With F H Add and fuse to obtain the initial fusion feature F P .

[0029] Preferably, the initial fusion feature F after upsampling P With F L They are input into the feature fusion module for fusion, and the final fusion feature is recorded as Fout The methods include:

[0030] The upsampled preliminary fusion feature F P Input weight calculation unit to convert into a set of weights W';

[0031] The weight W' and F L Multiplying, the expression is: Get the enhanced F L ;

[0032] The enhanced F L That is F L 'With F P Add and fuse to get the final fusion feature F out .

[0033] Preferably, the method in which the weight calculation unit converts the input features into a set of weights includes:

[0034] Assume that the input feature map F∈R C×H×W , where C, H, and W represent the number, length, and width of channels, respectively;

[0035] Through the global average pooling operation, the information of each channel is compressed into a C-dimensional vector, which is recorded as the channel descriptor U;

[0036] The channel descriptor U is input into the fully connected layer to obtain the global channel information, and the output vector is recorded as U g ;

[0037] The channel descriptor U is input into the one-dimensional convolutional layer to obtain the local channel information, and the output vector is recorded as U l ;

[0038] The global channel information and the local channel information are interactively operated through two parallel branches; one branch is U g with U l The transpose multiplication of is expressed as: Another branch is U l with U g The transpose multiplication of is:

[0039] Sum each column of M1 to get a one-dimensional vector U gw As the weight vector of global channel information, the expression is:

[0040] Sum each column of M2 to get a one-dimensional vector U lw As the weight vector of local channel information, the expression is:

[0041] The one-dimensional vector U gw and the one-dimensional vector U lw The result of the addition is activated by the activation function to obtain U w , the expression is: w =σ(U gw +U lw ), where σ represents the sigmoid activation function, which realizes the conversion from feature to weight vector.

[0042] The present invention also provides a microseismic event detection system based on multi-scale feature fusion, comprising: a data conversion module, a feature extraction module and a feature fusion module;

[0043] The data conversion module is used to acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram;

[0044] The feature extraction module is used to construct a multi-scale feature extraction module based on a convolutional neural network, and use the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn;

[0045] The feature fusion module is used to fuse features at different levels, and the fusion result is input into a multi-layer perceptron to obtain the final seismic phase picking result.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention first extracts features from seismic signals through a convolutional neural network, and can extract features at different levels. Then, through an upsampling module and a feature fusion module, features at different levels are fused in stages, thereby effectively integrating local features and global features.

[0048] The weight calculation unit in the feature fusion module can convert the feature map into a set of weights, which can effectively assign the channel weights of the feature map, thereby more accurately representing the importance of each channel.

[0049] The method of the present invention makes full use of the features extracted in each stage, which not only optimizes the processing capability of microseismic data features, but also improves the classification accuracy of microseismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] 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.

[0051] Figure 1 It is a schematic diagram of the steps of a microseismic event detection method based on multi-scale feature fusion in an embodiment of the present invention;

[0052] Figure 2 is an upsampling module diagram in an embodiment of the present invention;

[0053] Figure 3 is a feature fusion module diagram in an embodiment of the present invention;

[0054] Figure 4 Schematic diagram of a weight calculation unit in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] 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.

[0056] 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.

[0057] Embodiment 1

[0058] like Figure 1 As shown, this embodiment provides a microseismic event detection method based on multi-scale feature fusion, including the following steps:

[0059] Acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram;

[0060] Constructing a multi-scale feature extraction module based on a convolutional neural network, and using the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn;

[0061] The features at different levels are fused, and the fusion results are input into the multi-layer perceptron to obtain the final seismic phase picking result.

[0062] In this embodiment, the method of extracting features from the spectrum graph using the multi-scale feature extraction module to sequentially obtain features at different levels includes:

[0063] The spectrum graph is input into the convolutional neural network to obtain the features F at different levels in turn. L ,F M ,F H, wherein the convolutional neural network comprises four convolutional layers, which are connected in sequence;

[0064] Specifically: the spectrum graph is input into the first convolution layer for preliminary feature extraction, wherein the first convolution layer uses 8 9x9 convolution kernels with a step size of 2;

[0065] The second convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch-normalized, and then convolution is performed using 16 7x7 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the first convolutional layer is added to the output of the second convolutional layer to form a residual connection. The local features in the microseismic signal are obtained and recorded as F L , where F L =Conv2(x), x is the spectrum diagram. In microseismic signal processing, local features are usually related to the detailed information of the waveform;

[0066] The third convolutional layer uses 16 5x5 convolution kernels with a step size of 2 to further extract and enhance the local complex features in the microseismic signal, denoted by F M , where F M =Conv3(F L ), as the input of the fourth convolutional layer;

[0067] The fourth convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch normalized, and then convolution is performed using 16 3x3 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the third convolutional layer is added to the output of the fourth convolutional layer to form a residual connection. The global feature of the microseismic signal is recorded as F H , where F H =Conv4(F M ), reflecting the overall pattern and structural relationship of the signal.

[0068] In this embodiment, the method of fusing features at different levels to obtain the final seismic phase picking result includes:

[0069] F H After processing by the upsampling module, the upsampled F H With F M Input them into the feature fusion module for fusion, and the initial fusion feature is recorded as F P ;

[0070] The initial fusion feature F P After being processed by the upsampling module, the initial fusion feature F after upsampling P With F LThey are input into the feature fusion module for fusion, and the final fusion feature is recorded as F out ;

[0071] The final fusion feature F after upsampling out The input is processed by the multi-layer perceptron and mapped into a probability sequence related to the P wave and S wave in the seismic signal at each time point, which are expressed as Ppred and Spred, respectively, which are the final seismic phase picking results. The expression is:

[0072] Ppred = MLP1(F out )

[0073] Spred=MLP2(F out )

[0074] Among them, MLP is a multi-layer perceptron.

[0075] In this embodiment, if Figure 2 As shown, the upsampling module enlarges the input feature map. The expression of the upsampling module is: x'=reshape2(((linear(reshape1(x))), where reshape1 represents the first reshaping operation, linear represents the linear transformation, reshape2 represents the second reshaping operation, and x is the input feature map. Specifically, the upsampling module first reshapes the C channel values ​​at each position of the feature map into a 1×C vector, and then performs a linear transformation on the vector to obtain a 1×12 vector. Next, the 1×12 vector is reshaped into a 2×2×3 tensor. By performing the above operation on each position in the feature map, a new feature map can be obtained, thereby enlarging the scale of the original feature map.

[0076] In this embodiment, if Figure 3 As shown in Figure 2, the feature fusion module is designed to effectively fuse the input features. The expression of the feature fusion module is: Among them, F1 and F2 represent the two features to be fused, and F represents the fused feature. The features extracted by the multi-scale feature extraction module are fused to improve the accuracy and robustness of earthquake event detection. This module contains a weight calculation unit, such as Figure 4 As shown in Figure 1, this unit can convert the input features into a set of weights. The feature fusion module processes the features in two different stages to ensure full fusion and effective use of the features. The first time, the module converts the upsampled F H and F M The specific steps are as follows: First, the upsampled F H Input weight calculation unit to convert into a set of weights W. Then, this set of weights W and F MMultiplying, the expression is: This process is achieved by adjusting F M The importance of each channel is used to enhance useful features and suppress useless features, thereby achieving F M The enhanced F M (i.e. F M ') and F H Add and fuse to obtain the initial fusion feature F P , the expression is: F P =F M '+F H .

[0077] The second time is to combine the upsampled initial fusion features F P and F L The specific steps are as follows: First, the upsampled preliminary fusion features are input into the weight calculation unit to be converted into a set of weights W'. Then, this set of weights W' is combined with F L Multiplying, the expression is: This process is achieved by adjusting F L The weights of each channel are used to enhance useful features and suppress useless features, thereby achieving F L The enhanced F L (i.e. F L ') and F P Add and fuse to get the final fusion feature F out , the expression is: F out＝ F L '+F p .

[0078] The conversion of the above feature map into a set of weights is achieved through the weight calculation unit. The specific operations are as follows:

[0079] U=GAP(F)

[0080]

[0081] Among them, GAP is the global average pooling operation, U is the channel descriptor, FC represents the fully connected layer, Conv1D represents the one-dimensional convolutional layer, σ represents the sigmoid activation function, and W represents the weight of the final output.

[0082] Assume that the input feature map F∈R C×H×W, where C, H, and W represent the number of channels, length, and width, respectively. First, through the global average pooling operation, the information of each channel is compressed into a C-dimensional vector, denoted as the channel descriptor U. Each value of the channel descriptor can effectively represent the channel. The channel descriptor is input into the fully connected layer. The fully connected layer takes into account the information of all input channels, and it can capture the global dependencies between all channels. This global feature extraction capability enables the fully connected layer to integrate and process global information, so inputting the channel descriptor into the fully connected layer can obtain global channel information. The output vector is denoted as U g .

[0083] At the same time, the channel descriptor is input into the one-dimensional convolution layer. Each output node in the one-dimensional convolution layer is only connected to a local area of ​​the input channel. The one-dimensional convolution operation can realize local channel interaction and capture the local dependency between channels, so as to obtain local channel information. The output vector is recorded as U l .

[0084] Next, two parallel branches are used to perform the interaction between global channel information and local channel information. Specifically, one branch is U g with U l The transpose multiplication of is expressed as: Another branch is U l with U g The transpose multiplication of is:

[0085] Next, we sum each column of M1 to get a one-dimensional vector U gw As the weight vector of global channel information, the expression is: Sum each column of M2 to get a one-dimensional vector U lw As the weight vector of local channel information, the expression is:

[0086] The weight vector U representing the global channel information gw and represents the local channel information weight vector U lw After addition and fusion, the activation function is used to activate the weight vector U that integrates the global and local channel information. w , the expression is: w =σ(U gw +U lw ), where σ represents the sigmoid activation function. By effectively fusing global channel information with local channel information, effective channel weight allocation can be performed on the feature map. Through the above operations, the conversion from feature to weight vector is realized.

[0087] Embodiment 2

[0088] The present invention also provides a microseismic event detection system based on multi-scale feature fusion, comprising: a data conversion module, a feature extraction module and a feature fusion module;

[0089] The data conversion module is used to acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram;

[0090] The feature extraction module is used to construct a multi-scale feature extraction module based on a convolutional neural network, and use the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn;

[0091] The feature fusion module is used to fuse features at different levels, and the fusion results are input into the multi-layer perceptron to obtain the final seismic phase picking results.

[0092] 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 microseismic event detection method based on multi-scale feature fusion, characterized in that: The following steps are involved: Acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram; Constructing a multi-scale feature extraction module based on a convolutional neural network, and using the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn; The features at different levels are fused, and the fusion results are input into the multi-layer perceptron to obtain the final seismic phase picking result.

2. The microseismic event detection method based on multi-scale feature fusion according to claim 1 is characterized in that: The method of extracting features from the spectrum graph using the multi-scale feature extraction module to sequentially obtain features at different levels includes: Inputting the spectrum graph into a convolutional neural network to obtain features at different levels in sequence, wherein the convolutional neural network comprises four convolutional layers connected in sequence; Specifically: the spectrum graph is input into the first convolution layer for preliminary feature extraction, wherein the first convolution layer uses 8 9x9 convolution kernels with a step size of 2; The second convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch-normalized, and then convolution is performed using 16 7x7 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the first convolutional layer is added to the output of the second convolutional layer to form a residual connection. The local features in the microseismic signal are obtained and recorded as F L ; The third convolutional layer uses 16 5x5 convolution kernels with a step size of 2 to further extract and enhance the local complex features in the microseismic signal, denoted by F M , as the input of the fourth convolutional layer; The fourth convolutional layer is a two-dimensional convolutional residual block, which contains three repeated modules. Each module is first batch normalized, and then convolution is performed using 16 3x3 convolution kernels with a step size of 1. The three modules are connected in series in sequence, and the output of the third convolutional layer is added to the output of the fourth convolutional layer to form a residual connection. The global feature of the microseismic signal is recorded as F H .

3. The microseismic event detection method based on multi-scale feature fusion according to claim 2 is characterized in that: Methods for fusing features at different levels to obtain the final seismic phase picking results include: F H After processing by the upsampling module, the upsampled F H With F M Input them into the feature fusion module for fusion, and the initial fusion feature is recorded as F P ; The initial fusion feature F P After being processed by the upsampling module, the initial fusion feature F after upsampling P With F L They are input into the feature fusion module for fusion, and the final fusion feature is recorded as F out ; The final fusion feature F out After processing by the upsampling module, the final fusion feature F out , input into the multi-layer perceptron and mapped into a probability sequence associated with the P wave and S wave of the seismic signal at each time point, that is, the final seismic phase picking result.

4. The microseismic event detection method based on multi-scale feature fusion according to claim 3 is characterized in that: The upsampling module processes the following methods: Reshape the C channel values ​​at each position of the input feature map into a 1×C vector; Perform a linear transformation on the 1×C vector to obtain a 1×12 vector; The 1×12 vector is reshaped into a 2×2×3 tensor to obtain a new feature map.

5. The microseismic event detection method based on multi-scale feature fusion according to claim 3 is characterized in that: After upsampling, F H With F M Input them into the feature fusion module for fusion, and the initial fusion feature is recorded as F P The methods include: The upsampled F H Input weight calculation unit to convert into a set of weights W; The weight W and F M Multiplying, the expression is: Get the enhanced F M ; The enhanced F M That is F M 'With F H Add and fuse to obtain the initial fusion feature F P .

6. The microseismic event detection method based on multi-scale feature fusion according to claim 5 is characterized in that: Initial fusion feature F after upsampling P With F L They are input into the feature fusion module for fusion, and the final fusion feature is recorded as F out The methods include: The upsampled preliminary fusion feature F P Input weight calculation unit to convert into a set of weights W'; The weight W' and F L Multiplying, the expression is: Get the enhanced F L ; The enhanced F L That is F L 'With F P Add and fuse to get the final fusion feature F out .

7. The microseismic event detection method based on multi-scale feature fusion according to claim 6 is characterized in that: The method by which the weight calculation unit converts the input features into a set of weights includes: Assume that the input feature map F∈R C×H×W , where C, H, and W represent the number, length, and width of channels, respectively; Through the global average pooling operation, the information of each channel is compressed into a C-dimensional vector, which is recorded as the channel descriptor U; The channel descriptor U is input into the fully connected layer to obtain the global channel information, and the output vector is recorded as U g ; The channel descriptor U is input into the one-dimensional convolutional layer to obtain the local channel information, and the output vector is recorded as U l ; The global channel information and the local channel information are interactively operated through two parallel branches; one branch is U g with U l The transpose multiplication of is expressed as: Another branch is U l with U g The transpose multiplication of is: Sum each column of M1 to get a one-dimensional vector U gw As the weight vector of global channel information, the expression is: Sum each column of M2 to get a one-dimensional vector U lw As the weight vector of local channel information, the expression is: The one-dimensional vector U gw and the one-dimensional vector U lw The result of the addition is activated by the activation function to obtain U w , the expression is: w =σ(U gw +U lw ), where σ represents the sigmoid activation function, which realizes the conversion from feature to weight vector.

8. A microseismic event detection system based on multi-scale feature fusion, characterized in that: include: Data conversion module, feature extraction module and feature fusion module; The data conversion module is used to acquire time series data by collecting microseismic data, convert the time series data into frequency domain data, and save it in the form of a spectrum diagram; The feature extraction module is used to construct a multi-scale feature extraction module based on a convolutional neural network, and use the multi-scale feature extraction module to extract features from the spectrum graph to obtain features at different levels in turn; The feature fusion module is used to fuse features at different levels, and the fusion result is input into a multi-layer perceptron to obtain the final seismic phase picking result.

Images