A method for seismic phase picking based on MFU-Net model

Abstract

The present invention relates to the field of seismic exploration, and in particular to a seismic phase picking method based on the MFU-Net model. The overall design of the present invention significantly improves the accuracy of seismic phase picking. Compared with the traditional U-Net model, MFU-Net achieves more accurate P- and S-wave detection by optimizing skip connections, enhancing the key feature extraction capabilities of the bottleneck layer, and addressing dataset imbalance. This improves the accuracy of seismic phase picking.

Description

Technical Field

[0001] The present invention relates to fields related to seismic exploration, and in particular to a seismic phase picking method based on an MFU-Net model. Background Art

[0002] In existing earthquake phase picking, some researchers use the encoder-decoder structure of the U-Net model to process seismic waveform data. However, the traditional U-Net's skip connection directly splices the feature maps of the encoder and decoder (a simple concat operation), resulting in insufficient fusion of semantic information and spatial information at different levels, limiting the model's ability to synergistically utilize local details and global features of seismic waveforms. In the earthquake phase picking process, the bottleneck layer of the U-Net model is responsible for capturing global contextual information, but the standard convolution operation treats all features equally and cannot distinguish between key signals (P waves / S waves) and background noise, resulting in important information being submerged, which easily affects the accuracy of the earthquake phase picking process. In addition, most time points in seismic data are noise, and there are very few P-wave and S-wave samples. Traditional loss functions, such as cross-entropy, are biased towards the majority class, that is, noise, resulting in the model's low recognition sensitivity to the minority class (P waves / S waves), which in turn easily affects the accuracy of earthquake phase picking. Summary of the Invention

[0003] Therefore, in order to solve the above shortcomings, the present invention provides an earthquake phase picking method based on the MFU-Net model to solve the above technical problems.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for picking earthquake phases based on the MFU-Net model, comprising the following steps:

[0005] S1. Build an MFU-Net network model with an encoder-decoder structure. The MFU-Net network model is introduced based on U-Net:

[0006] A multi-scale feature fusion module, which is embedded in the skip connection and is used to fuse multi-level features between the encoder and decoder;

[0007] Multi-head attention mechanism, integrated into the bottleneck layer, enhances the perception of key signal areas;

[0008] Weighted loss function to address the class imbalance problem in training data;

[0009] S2, obtaining raw seismic signal data, and inputting it into the MFU-Net model after preprocessing;

[0010] S3, downsample the input signal layer by layer through the encoder to extract multi-scale features;

[0011] S4. Apply a multi-head attention mechanism at the bottleneck layer to calculate the correlation between queries, keys, and values, generate attention weights, and strengthen the representation of key phase areas;

[0012] S5, the decoder gradually restores the resolution through transposed convolution operations and uses the MFFM module to fuse the multi-scale features from the encoder;

[0013] S6, the final output layer uses the SoftMax classifier to perform three-category prediction, corresponding to P wave, S wave and noise respectively;

[0014] S7. Use a weighted loss function to train the model, where the class-specific weights are inversely proportional to the frequency of each class sample to alleviate the training bias caused by the scarcity of P-wave and S-wave samples.

[0015] Preferably, the specific implementation process of the multi-scale feature fusion module includes:

[0016] S11. Perform concatenation on the two input features and merge the two inputs along the channel dimension. The formula is expressed as: ;

[0017] where concat represents the concatenation function applied along the channel dimension, and Represents the two features input to the multi-scale feature fusion module, Represents two input features and The result of the concatenation operation along the channel dimension;

[0018] S12, apply three convolution branches with different receptive fields to extract corresponding multi-scale features;

[0019] S13. Fusion of the features extracted by these three branches by element-by-element addition: , , , to produce the final multi-scale feature representation, the formula is: ;

[0020] in, It represents the multi-scale fusion features finally output after module processing.

[0021] Preferably, the specific implementation process of the multi-head attention mechanism includes:

[0022] S41. Compute the query (Q), key (K), and value (V) through independent linear transformations, formally expressed as:

[0023] ;

[0024] ;

[0025] ;

[0026] Where (X) is the input feature sequence, 、 、 is the projection matrix of the h-th head;

[0027] S42. For each linearly transformed group Qi, Ki, and Vi, perform a scaled dot product attention operation as follows:

[0028] Compute the dot product between Qi and Ki to obtain the raw similarity scores, and scale these scores by a factor of 1 / dk (where dk represents the dimension of Ki);

[0029] After applying the optional mask operation, the scores are normalized by the SoftMax function to generate attention weights;

[0030] Multiplying the attention weight with Vi produces the output of the i-th attention head. The scaled dot product attention can be formally expressed as: ;

[0031] S43. Concatenate the outputs of all attention heads and project the combined features to the desired output dimension through a linear transformation to produce the final output of the multi-head attention mechanism.

[0032] Preferably, the specific implementation process of the weighted loss function includes:

[0033] S71. Define a category set C. For each category c in the set, , setting its weight inversely proportional to the category frequency;

[0034] S72. Assign higher weights to P waves and S waves to solve the problem of sample imbalance. A class-weighted loss function is introduced into the loss function:

[0035] , , Inversely proportional to category frequency;

[0036] in Represents a collection of categories, represents the loss function, N represents the total number of samples, Indicates the P wave category, Indicates the S wave category, Indicates the noise category, is a class-specific weight that is inversely proportional to the class frequency, which means that classes with few samples (P-wave and S-wave) are given higher weights (larger wc) in the loss calculation, while classes with many samples (noise) are given lower weights (smaller wc). and are the true label and predicted probability of category c for sample i, respectively.

[0037] Preferably, the encoder consists of multiple stacked one-dimensional convolutional layers, each of which is sequentially connected to a batch normalization layer and an activation function layer; some convolutional layers use a larger step size to achieve fast spatial dimension compression, improve computational efficiency and enhance the ability to extract high-level semantic features.

[0038] Preferably, the decoder is composed of multiple transposed convolutional layers, each layer is used to gradually restore the temporal resolution of the signal. During the decoding process of each layer, the feature map from the corresponding level of the encoder is introduced, and feature fusion is performed through a multi-scale feature fusion module to improve positioning accuracy and phase recognition ability.

[0039] Preferably, the original seismic signal is a three-component signal, and the input size is , representing the seismic waveform data in the X, Y, and Z directions. After being processed by the MFU-Net model, it outputs a three-category label sequence of the same length, representing the probability distribution of P waves, S waves, and noise, respectively, for automatic picking of subsequent seismic phase arrival times.

[0040] Beneficial effects of the present invention:

[0041] 1. The present invention embeds a multi-scale feature fusion module into the skip connection. The multi-scale feature fusion module uses convolution kernels of different sizes to capture multi-scale features. Different convolution kernel sizes represent different receptive areas. Smaller convolution kernels focus on local details, while larger convolution kernels extract global or large target features. By fusing these multidimensional features, the model can simultaneously utilize fine-grained and coarse-grained information, thereby enhancing its ability to represent complex scenes and diverse targets, thereby improving the model's ability to represent complex seismic waveform scenes and diverse targets.

[0042] 2. The present invention integrates the multi-head attention mechanism into the bottleneck layer. By calculating the query, key and value through independent linear transformations, and then performing a scaled dot product attention operation, it can calculate the correlation of each part in the input feature sequence and generate attention weights, thereby highlighting the characteristics of the key seismic phase area and enhancing the perception of key signals, that is, enhancing the perception of P waves and S waves, improving the ability to extract and identify key seismic phase features, and helping to improve the accuracy of the earthquake phase picking process.

[0043] 3. The present invention introduces a weighted loss function to assign higher weights to P waves and S waves, so that categories with scarce samples (P waves and S waves) are given higher weights in the loss calculation, while categories with numerous samples (noise) are given lower weights. In this way, during the model training process, the training bias caused by the scarcity of P-wave and S-wave samples can be alleviated, and the model's recognition ability for minority classes, that is, the ability to recognize P waves and S waves, is improved, which is conducive to improving the accuracy of earthquake phase picking.

[0044] 4. The overall design of the present invention greatly improves the accuracy of earthquake phase picking. Compared with the traditional U-Net model, MFU-Net achieves more accurate P-wave and S-wave detection by optimizing skip connections, strengthening the key feature extraction capability of the bottleneck layer, and solving the problem of data set imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 1 is a flow chart of the earthquake phase picking method based on the MFU-Net model of the present invention;

[0046] Figure 2 It is a schematic diagram of the overall architecture of the MFU-Net model of the present invention;

[0047] Figure 3 Schematic diagram of the scale feature fusion structure of the present invention;

[0048] Figure 4 It is a flowchart of the multi-head attention mechanism of the present invention. DETAILED DESCRIPTION

[0049] In order to further explain the technical solution of the present invention, specific embodiments are described in detail below.

[0050] like Figures 1 to 4 As shown, the present invention provides an earthquake phase picking method based on the MFU-Net model, comprising the following steps:

[0051] S1. Build an MFU-Net network model with an encoder-decoder structure. The MFU-Net network model is introduced based on U-Net:

[0052] A multi-scale feature fusion module is embedded in the skip connection and is used to fuse multi-level features between the encoder and decoder. The specific implementation process of the multi-scale feature fusion module includes:

[0053] S11. Perform concatenation on the two input features and merge the two inputs along the channel dimension. The formula is expressed as: ;

[0054] where concat represents the concatenation function applied along the channel dimension, and Represents the two features input to the multi-scale feature fusion module, Represents two input features and The result of the splicing operation along the channel dimension

[0055] S12, apply three convolution branches with different receptive fields to extract corresponding multi-scale features;

[0056] S13. Fusion of the features extracted by these three branches by element-by-element addition: , , , to produce the final multi-scale feature representation, the formula is: ;in, Represents the multi-scale fusion features finally output after module processing;

[0057] The core idea of ​​the multi-scale feature fusion module is to use convolution kernels of different sizes to capture multi-scale features. Different convolution kernel sizes represent different receptive areas. Smaller convolution kernels focus on local details, while larger convolution kernels extract global or large-scale target features. By fusing these multi-dimensional features, the model can simultaneously utilize fine-grained and coarse-grained information, thereby enhancing its ability to represent complex scenes and diverse targets.

[0058] Multi-head attention mechanism, integrated into the bottleneck layer, enhances the perception of key signal areas;

[0059] Weighted loss function to address the class imbalance problem in training data;

[0060] The overall architecture of the MFU-Net model is shown in the attached manual. Figure 2 As described above, k represents the convolution kernel size, s represents the stride, BN represents batch normalization, (i, j) represents the signal length and the number of channels, Conv represents a one-dimensional standard convolution layer, DeConv refers to transposed convolution (deconvolution), MAM represents a multi-head attention mechanism, and MFFM represents a multi-scale feature fusion module;

[0061] In the actual work of the MFU-Net model, the encoder on the left gradually compresses the spatial dimension by downsampling with a stride of 2, such as input 6000×3 → bottleneck layer 94×64, extracts multi-scale features, and introduces a multi-head attention mechanism (MAM) at the bottom bottleneck layer to enhance the perception of key areas. On the right, the decoder transposed convolution (DeConv) gradually upsamples to restore the resolution, and the skip connection is replaced by the MFFM module (see details). Figure 3), the multi-scale features of the encoder and decoder are integrated, and finally the output layer uses Softmax to perform three classifications (P wave / S wave / noise);

[0062] Furthermore, the encoder consists of multiple stacked one-dimensional convolutional layers, each of which is sequentially connected to a batch normalization layer and an activation function layer; some convolutional layers use a larger step size to achieve fast spatial dimension compression, improve computational efficiency, and enhance the ability to extract high-level semantic features;

[0063] Furthermore, the decoder is composed of multiple transposed convolutional layers, each of which is used to gradually restore the temporal resolution of the signal. During the decoding process of each layer, the feature map from the corresponding level of the encoder is introduced and feature fusion is performed through a multi-scale feature fusion module to improve positioning accuracy and phase recognition capabilities.

[0064] S2, obtaining raw seismic signal data, and inputting it into the MFU-Net model after preprocessing;

[0065] The original seismic signal is a three-component signal, and the input size is , represents the seismic waveform data in the X, Y, and Z directions. After being processed by the MFU-Net model, it outputs a three-category label sequence of the same length, representing the probability distribution of P waves, S waves, and noise, respectively, for the automatic picking of subsequent seismic phase arrival times;

[0066] S3, downsample the input signal layer by layer through the encoder to extract multi-scale features;

[0067] S4. Apply the multi-head attention mechanism at the bottleneck layer to calculate the correlation between the query, key, and value, generate attention weights, and strengthen the representation of the key phase area. The specific implementation process of the multi-head attention mechanism (MAM) includes:

[0068] S41. Compute the query (Q), key (K), and value (V) through independent linear transformations. These linear projections map the original input to the appropriate dimensions for attention calculation, formally expressed as:

[0069] ;

[0070] ;

[0071] ;

[0072] Where (X) is the input feature sequence, 、 、 is the projection matrix of the h-th head;

[0073] S42. For each linearly transformed group Qi, Ki, and Vi, perform a scaled dot product attention operation as follows:

[0074] Compute the dot product between Qi and Ki to obtain the raw similarity scores, and scale these scores by a factor of 1 / dk (where dk represents the dimension of Ki);

[0075] After applying the optional mask operation, the scores are normalized by the SoftMax function to generate attention weights;

[0076] Multiplying the attention weight with Vi produces the output of the i-th attention head. The scaled dot product attention can be formally expressed as: ;

[0077] S43. Concatenate the outputs of all attention heads and project the combined features to the desired output dimension through a linear transformation to produce the final output of the multi-head attention mechanism.

[0078] The multi-head attention mechanism generates attention weights by calculating the correlation between queries, keys, and values. This can highlight key signals such as P-waves and S-waves and suppress noise interference, thereby enhancing the perception of key signal areas. This allows the MFU-Net model to more accurately capture key information about seismic phases and improve the accuracy of earthquake phase detection.

[0079] S5, the decoder gradually restores the resolution through transposed convolution operations and uses the MFFM module to fuse the multi-scale features from the encoder;

[0080] S6, the final output layer uses the SoftMax classifier to perform three-category prediction, corresponding to P wave, S wave and noise respectively;

[0081] S7. Use a weighted loss function to train the model, where the class-specific weights are inversely proportional to the frequency of each class sample to mitigate the training bias caused by the scarcity of P-wave and S-wave samples. The specific implementation of the weighted loss function includes:

[0082] S71. Define a category set C. For each category c in the set, , setting its weight inversely proportional to the category frequency;

[0083] S72. Assign higher weights to P waves and S waves to solve the problem of sample imbalance. A class-weighted loss function is introduced into the loss function:

[0084]

[0085] ( , Inversely proportional to category frequency)

[0086] in Represents a collection of categories, represents the loss function, N represents the total number of samples, Indicates the P wave category, Indicates the S wave category, Indicates the noise category, is a class-specific weight that is inversely proportional to the class frequency, which means that classes with few samples (P-wave and S-wave) are given higher weights (larger wc) in the loss calculation, while classes with many samples (noise) are given lower weights (smaller wc). and are the true label and predicted probability of category c for sample i, respectively.

[0087] The present invention provides an earthquake phase picking method based on the MFU-Net model, which greatly improves the accuracy of earthquake phase picking. Compared with the traditional U-Net model, MFU-Net achieves more accurate P-wave and S-wave detection by optimizing skip connections, strengthening the key feature extraction capability of the bottleneck layer, and solving the problem of data set imbalance.

[0088] The above description is merely 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, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A seismic phase picking method based on the MFU-Net model, characterized in that: The following steps are involved: S1. Build an MFU-Net network model with an encoder-decoder structure. The MFU-Net network model is introduced based on U-Net: A multi-scale feature fusion module, which is embedded in the skip connection and is used to fuse multi-level features between the encoder and decoder; Multi-head attention mechanism, integrated into the bottleneck layer, enhances the perception of key signal areas; Weighted loss function to address the class imbalance problem in training data; S2, obtaining raw seismic signal data, and inputting it into the MFU-Net model after preprocessing; S3, downsample the input signal layer by layer through the encoder to extract multi-scale features; S4. Apply a multi-head attention mechanism at the bottleneck layer to calculate the correlation between queries, keys, and values, generate attention weights, and strengthen the representation of key phase areas. The specific implementation process of the multi-head attention mechanism includes: S41. Calculate the query Q, key K, and value V through independent linear transformations, formally expressed as: ； ； ； Where X is the input feature sequence, 、 、 is the projection matrix of the h-th head; S42. For each linearly transformed group Qi, Ki, and Vi, perform a scaled dot product attention operation as follows: Compute the dot product between Qi and Ki to obtain the raw similarity scores, and scale these scores by a factor 1 / dk, where dk represents the dimension of Ki; After applying the optional mask operation, the scores are normalized by the SoftMax function to generate attention weights; Multiplying the attention weight with Vi produces the output of the i-th attention head. The scaled dot product attention is expressed as: ; S43. Concatenate the outputs of all attention heads and project the combined features to the desired output dimension through a linear transformation to produce the final output of the multi-head attention mechanism. S5, the decoder gradually restores the resolution through transposed convolution operations and uses the MFFM module to fuse the multi-scale features from the encoder; S6, the final output layer uses the SoftMax classifier to perform three-category prediction, corresponding to P wave, S wave and noise respectively; S7. Use a weighted loss function to train the model, where the class weight is inversely proportional to the frequency of each class sample to alleviate the training bias caused by the scarcity of P-wave and S-wave samples. The specific implementation process of the weighted loss function includes: S71. Define a category set C. For each category c in the set, , setting its weight inversely proportional to the category frequency; S72. Assign higher weights to P waves and S waves to solve the problem of sample imbalance. A class-weighted loss function is introduced into the loss function: ， ； in Represents a collection of categories, represents the loss function, N represents the total number of samples, Indicates the P wave category, Indicates the S wave category, Indicates the noise category, is the weight of the category, which is inversely proportional to the category frequency. P-wave and S-wave are given high weights in the loss calculation, and noise is given low weights. and are the true label and predicted probability of category c for sample i, respectively.

2. The earthquake phase picking method based on the MFU-Net model according to claim 1, characterized in that: The specific implementation process of the multi-scale feature fusion module includes: S11. Perform a concatenation operation on the two input features and merge the two inputs along the channel dimension. The formula is expressed as: ; where concat represents the concatenation function applied along the channel dimension, and Represents the two features input to the multi-scale feature fusion module, Represents two input features and The result of the concatenation operation along the channel dimension; S12, apply three convolution branches with different receptive fields to extract corresponding multi-scale features; S13. Fusion of the features extracted by these three branches by element-by-element addition: , , , to produce the final multi-scale feature representation, the formula is: ； in, It represents the multi-scale fusion features finally output after module processing.

3. The earthquake phase picking method based on the MFU-Net model according to claim 1, characterized in that: The encoder consists of multiple stacked one-dimensional convolutional layers, each of which is sequentially followed by a batch normalization layer and an activation function layer. Some convolutional layers use large strides to achieve fast spatial dimension compression, improve computational efficiency, and enhance the ability to extract high-level semantic features.

4. The earthquake phase picking method based on the MFU-Net model according to claim 1, characterized in that: The decoder consists of multiple transposed convolutional layers, each of which is used to gradually restore the temporal resolution of the signal. During the decoding process of each layer, the feature map from the corresponding level of the encoder is introduced, and feature fusion is performed through a multi-scale feature fusion module to improve positioning accuracy and phase recognition capabilities.

5. The earthquake phase picking method based on the MFU-Net model according to claim 1, characterized in that: The original seismic signal is a three-component signal, and the input size is , representing the seismic waveform data in the X, Y, and Z directions. After being processed by the MFU-Net model, it outputs a three-category label sequence of the same length, representing the probability distribution of P waves, S waves, and noise, respectively, for automatic picking of subsequent seismic phase arrival times.

Images