Microseismic signal denoising method fusing multi-scale features and global-local attention

Abstract

The invention discloses a microseismic signal denoising method fusing multi-scale features and global-local attention. The method comprises the following steps: constructing a large-scale pure microseismic signal data set based on elastic wave field simulation; carrying out noise-containing sample set generation and data standardization preprocessing in combination with real environment noise; carrying out scientific division and preparation on the noisy / pure sample pair; an MSF-GLANet layered coding and decoding network framework based on the U-Net + + topology is constructed; constructing a composite loss function which takes the Charbonier loss as a core, and combining a multi-level depth supervision strategy; performing initialization and iterative training based on a composite loss function on the MSF-GLANet network by using the training set; the model performance is monitored in real time through the verification set, and the optimal model weight is stored; a real environment microseismic signal to be denoised is preprocessed and then input into the trained model for reasoning; and performing element-by-element addition reconstruction on a residual signal predicted by the network through a residual connection mechanism to realize high-fidelity recovery of a de-noised signal.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a technology for the accurate recovery of weak microseismic signals in low signal-to-noise ratio environments. Specifically, it relates to a microseismic signal denoising method that integrates multi-scale features and global-local attention for the identification, localization, and subsequent analysis of weak elastic wave signals generated by rock fractures in geological structures. Background Technology

[0002] Microseismic monitoring technology, which detects weak elastic wave signals generated by rock fractures in geological structures, has become a key technology in fields such as oil and gas reservoir exploration and hydraulic fracturing monitoring. In this technology, noise suppression of microseismic signals is a crucial step in data processing. Improving the signal-to-noise ratio (SNR) of acoustic signals can significantly enhance the accuracy of microseismic event identification, localization, and subsequent analysis. Traditional methods based on frequency domain, time-frequency analysis, or wavelet transform struggle to effectively handle large-scale, low-SNR, non-stationary noise data, limiting their accuracy and adaptability. In recent years, deep learning technology has made progress in signal denoising, particularly performing well with characteristic acoustic signals. However, the complex acquisition environment of microseismic signals introduces a large amount of non-stationary and nonlinear noise, and the inherently low SNR of the signals makes it difficult for existing deep learning denoising methods to achieve ideal results.

[0003] Patent document CN120195735A discloses a method for denoising and reconstructing microseismic signals based on multi-scale Shapelet decomposition and U-Net. Its feature extraction process relies on the pre-defined local matching pattern of Shapelet decomposition, inherently lacking end-to-end adaptive learning capabilities and struggling to robustly handle complex ambient noise. Furthermore, the model is deficient in global information modeling, primarily relying on locally stacked TCN and traditional U-Net structures, lacking global context awareness and failing to effectively suppress non-stationary noise across time series. Simultaneously, its skip connections lack intelligent gating and filtering mechanisms, easily transmitting redundant information indiscriminately, further interfering with the accurate reconstruction of weak signals. Summary of the Invention

[0004] The purpose of this invention is to address the technical problem that existing deep learning techniques for denoising microseismic signals are difficult to accurately recover microseismic signals in low signal-to-noise ratio environments due to insufficient multi-scale feature capture, limited long-range dependency modeling, and noise introduced by skip connections. The proposed method is a microseismic signal denoising method that integrates multi-scale features and global-local attention.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for denoising microseismic signals by integrating multi-scale features and global-local attention includes the following steps: Step 1: Constructing a large-scale clean microseismic signal dataset based on elastic wave field simulation; Step 2: Generating a noisy sample set combined with real environmental noise and performing data standardization preprocessing; Step 3: Scientifically dividing and preparing noisy / clean sample pairs according to statistical principles; Step 4: Constructing an MSF-GLANet hierarchical encoder-decoder network skeleton based on U-Net++ topology; Step 5: Integrating an MSFF multi-scale feature fusion module into the encoder path to capture multi-scale signal features; Step 6: Configuring a GLFB global-local fusion block at the network bottleneck layer to collaboratively model long-range dependencies and local context; 7. Replace traditional skip connections with SEGA squeeze and excitation-gated attention modules to achieve intelligent feature selection; 8. Construct a composite loss function with Charbonnier loss as the core and combine it with a multi-level deep supervision strategy; 9. Initialize the MSF-GLANet network using the training set and iteratively train it based on the composite loss function; 10. Monitor the model performance in real time using the validation set and save the best model weights; 11. Input the preprocessed real-world microseismic signal to be denoised into the trained model for inference; 12. Reconstruct the residual signal predicted by the network by adding it element-wise through the residual connection mechanism to achieve high-fidelity recovery of the denoised signal.

[0006] In step 1, based on the actual geological conditions and exploration background parameters, a large-scale clean microseismic signal dataset with different sources, propagation paths and structural characteristics is generated using elastic wave field forward modeling technology.

[0007] In step 2, the real environmental noise of the target work area is collected, and according to the preset signal-to-noise ratio distribution, it is superimposed with the clean signal generated in step 1 to form a noisy / clean sample pair; then, uniform normalization and other standardized preprocessing operations are performed on the entire dataset to eliminate differences in data scale.

[0008] In step 3, to ensure the stability of model training and the fairness of evaluation, the preprocessed noisy / clean sample pairs are strictly divided into independent training sets, validation sets and test sets to ensure the model's generalization ability under different data distributions.

[0009] In step 4, the MSF-GLANet hierarchical encoder-decoder network structure based on the U-Net++ topology includes an MSFF multi-scale feature fusion module, a GLFB global-local fusion block, and a SEGA squeezing and excitation-gated attention module; the network structure is specifically as follows: Input signal After convolution, the resulting features are used as initial features to connect to the first layer. Module input; first layer, first module Module output characteristics ; For encoder paths, After the first layer of downsampling Then input to the first one of the second layer Module, first in the second layer Module output characteristics ;feature After the second layer of downsampling Then input to the first one of the third layer Module, first in the third layer Module output characteristics ;feature After the third layer of downsampling Then input to the first one of the fourth layer Module, first in the fourth layer Module output characteristics ;feature After the fourth layer of downsampling Then enter Module, Module output bottleneck layer features .

[0010] Decoder path from features Upsampling and feature fusion begin sequentially. In the main network structure, the features of each decoder node are formed by concatenating and fusing upsampled features from the next layer and multiple predecessor features from the same layer. Feature fusion in skip connections employs... The module performs gating and filtering.

[0011] Specifically for the decoder: The first level of decoding is as follows: feature After upsampling, with features Passing by together The features processed by the module are the second feature of the fourth layer. The input features of the module, the second one in the fourth layer. Module output characteristics ;feature After upsampling, with features Passing by together The features processed by the module are used as the second feature of the third layer. The input features of the module, the second one in the third layer Module output characteristics ;feature After upsampling, with features Passing by together The features processed by the module are used as the second layer. The input features of the module, the second layer, the second Module output characteristics ;feature After upsampling, with features Passing by together The features processed by the module are used as the first and second layers. The input features of the module, the first layer and the second layer Module output characteristics ; The second-level decoding is as follows: feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the third layer. The input features of the module, the third layer, the third... Module output characteristics ;feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the second and third layers. The input features of the module, the second and third layers Module output characteristics ;feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the first and third layers. The input features of the module, the third layer of the first layer Module output characteristics ; The third-level decoding specifically involves: feature After upsampling, with features ,feature and characteristics Passing by together The features processed by the module are used as the fourth feature of the second layer. The input features of the module, the fourth one in the second layer. Module output characteristics ;feature After upsampling, with features ,feature and characteristics Passing by together The features processed by the module are the fourth in the first layer. The module's input features, the fourth one in the first layer. Module output characteristics ; The fourth level of decoding is as follows: feature After upsampling, with features ,feature ,feature and characteristics Passing by together The features processed by the module are the fifth in the first layer. The module's input features, the fifth one in the first layer. Module output characteristics ; In the main output path, features After the fourth convolutional module Output and input signals Perform residual connections to obtain the main output information. .

[0012] In step 5, an MSFF module is integrated into the initial feature extraction layer at the network front end and each downsampling layer in the encoder path. This module synchronously captures multi-scale features of the signal using parallel multi-size convolutional kernels and employs SE squeezing and excitation mechanisms to dynamically weight the fused features at the channel level, thereby highlighting effective signal features. Specifically: Module receives input features And simultaneously input them into three parallel feature extraction branches; The specific steps of the first branch are as follows: Input features The input to the first 3×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 3×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 3×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs the first branch multi-scale features. ; The specific steps for the second branch are as follows: Input features The input to the first 5×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 5×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 5×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs the second branch multi-scale features. ; The specific steps for the third branch are as follows: Input features The input to the first 7×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 7×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 7×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs third-branch multi-scale features. ; Then we move to the main branch, which contains the multi-scale features of these three branches. , and By piecing them together, a fusion feature is obtained. and will As Input characteristics of the extrusion and excitation modules; Weighted characteristics of the output channel of the extrusion and excitation module Channel weighted features Channel weighting characteristics Connect the input of the first 3×1 convolutional layer in the main branch, and use the output features of the first 3×1 convolutional layer as the first batch normalization layer. The input features, the first batch normalization layer Output features as the first The input features of the activation function, the first The input features of the activation function are used as the input features of the second 3×1 convolutional layer, and the output features of the second 3×1 convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer Output features as the second The input features of the activation function, the second The output features of the activation function and Residual connection, to obtain The module's final output.

[0013] exist In the module, fused features After global average pooling The output after the operation is squeezed is connected to the first fully connected layer. The input, the first fully connected layer Output connection The input to the activation function, The activation function output connects to the second fully connected layer. The input, the second fully connected layer Output connection The input to the activation function, Activation function output connected channel multiplication Input for the operation The operation will apply the learned channel weights. Channel weighted features are obtained. .

[0014] In step 6, at the highest abstraction layer of the network, i.e., the bottleneck layer, configure... The global-local fusion block employs a dual-branch parallel structure; where the local branch is implemented through... Obtain multi-granularity local context and integrate global branches The encoder explicitly models the long-range dependencies of the signal; specifically: Module receives input features and connect it to Local branches and Input for the global branch; exist In local branches, Feature connected to The inputs to the five parallel submodules are: 1×1 convolution path: The feature is connected to the input of a 1×1 convolutional layer, the output of which is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the first feature path. ; Path with expansion rate r=6: The feature connects to the input of a 3×1 convolutional layer with a dilation of 6, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the second-path features. ; Path with expansion rate r=12: The feature connects to the input of a 3×1 convolutional layer with a dilation of 12, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the third-path feature. ; Path with expansion rate r=18: The feature connects to the input of a 3×1 convolutional layer with a dilation of 18, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the third-path feature. ; Global pooling path: Feature connection global average pooling The input to the operation is global average pooling. The operation output is connected to the input of a 1×1 convolutional layer, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function output is connected to the input of the upsampling operation, and the output of the upsampling operation is the fifth feature. ; Will , , , and The five features are concatenated, and the concatenated features are connected to the input of a 1×1 output convolutional layer. This output convolutional layer outputs local branches of the features. ; exist In the global branch: Features are obtained after positional encoding. , The input is connected to the multi-head self-attention mechanism, in which... The features are first connected to three parallel linear transformation layers: Query the linear layer: Features are connected to the first linear layer The linear layer takes the input as input and outputs the query features. ; Bond linear layer: Features are connected to the second linear layer The linear layer takes the input as input and outputs the key features. ; Value linear layer: Features are connected to the third linear layer The input of the linear layer is the feature value of the output value. ; Query features Bond features Connect to matrix multiplication Input for the operation Operation output connected to scaling The input to the operation and the output of the scaling operation are connected to the mask. Input for the operation Operation output connected to The input to the activation function, The activation function outputs an attention weight matrix; the attention weight matrix is ​​then concatenated into a second matrix multiplication. The input for the operation, The operation combines the attention weight matrix with the value features. Multiply the results to output the scaled dot product attention result; this scaled dot product attention result is connected to the concatenated input; the concatenation layer fuses the attention results obtained from multi-head computation along the channel dimension; the concatenated features are then connected to the fourth linear layer. The input is the linear layer, and the output of this linear layer is the multi-head self-attention output; this output is then processed... Operation and After addition, connect to the first normalization layer. Input; The output is connected to the input of the feedforward network. ;exist internal: The output is connected to the first linear layer. Input, Output connection The input to the activation function, Output connection to the second The input, the second The output connects to the second linear layer. The input, the second Output connection to the third The input, get Output ; go through After operation and the first layer The input features are added together and used as the second layer. Input features, second layer The output yields the global branch output characteristics. ; In the final fusion stage, Local branch output features and Global branch output features The features are concatenated and then connected to the input of a 1×1 fusion convolutional layer. The output of this fusion convolutional layer is the... The module's final output; In step 7, the traditional non-selective skip connection is replaced with The module utilizes semantic information from the deeper layers of the decoder as a gating signal to achieve dynamic recalibration and filtering of shallow encoder features, effectively suppressing redundant noise propagated non-selectively in skip connections; specifically: The module receives features upsampled from the decoder. and from the same level layer of the encoder Two input signals; first, Connect to the first The module's input, Weighted gating signal of module output channel ;at the same time, Connect to the second The module's input, Weighted skip connection feature of module output channel ; Next, in the feature projection path, the channel-weighted gated signal The input is connected to the first 1×1 convolutional layer, and the output of that convolutional layer is connected to the batch normalization layer. Obtain gated projection features Simultaneously, channel-weighted skip connection characteristics The input is connected to a second 1×1 convolutional layer, the output of which is connected to a batch normalization layer. To obtain the jump projection features ; Gated projection features and jump projection features Element-level fusion is performed through addition, and the added features are connected to... The input to the activation function, The output is connected to the input of the attention coefficient generator; in the attention coefficient generator, The output of the first convolutional layer is connected to the input of the third 1×1 convolutional layer, and the output of the third convolutional layer is connected to the batch normalization layer. normalization layer Output connected to The input to the activation function, The activation function outputs an attention coefficient matrix. ; Finally, the attention coefficient matrix Connected to the input of the channel-level multiplication operation. Channel-weighted skip connectivity features Perform dot product to obtain gating and recalibration. The module's final output; In step 8, a composite loss function is constructed with the Charbonnier loss function, which is robust to outliers, as its core. Furthermore, a deep supervision strategy is employed, adding auxiliary supervision signals at different levels of the decoder to ensure effective gradient backpropagation and accelerate the optimization of multi-scale features. Specifically: In the main network architecture, the second layer of the first layer Module output characteristics After the first convolutional module Get auxiliary output The third one on the first floor Module output characteristics After the second convolutional module Get auxiliary output The fourth one on the first floor Module output characteristics After the third convolutional module Get auxiliary output .

[0015] Compared with the prior art, the present invention has the following technical effects: 1) The method and network proposed in this application can robustly capture multi-scale features, accurately model long-range dependencies, and intelligently filter redundant noise during feature fusion, especially achieving accurate recovery of weak signals in low signal-to-noise ratio environments; 2) In this invention, the network model integrates the MSFF multi-scale feature fusion module, the GLFB global-local fusion block, and the squeezing and SEGA excitation gating attention mechanism to construct MSF-GLANet based on the improved U-Net++ architecture for denoising complex low signal-to-noise ratio microseismic signals. Specifically, by employing the MSFF multi-scale feature fusion module, multiple convolutional kernels of different sizes are used in parallel to simultaneously capture the local details and macroscopic trends of the signal. Simultaneously, the SE squeezing and excitation mechanism is incorporated to dynamically weight the fused features at the channel level, effectively highlighting the target signal features and suppressing redundant information between channels, overcoming the insufficient adaptability of traditional single-kernel convolution to complex multi-scale features. 3) This invention introduces a GLFB global-local fusion block in the bottleneck layer and adopts a dual-branch parallel architecture. Local features obtain multi-granularity context through the ASPP hollow spatial pyramid pooling module; global dependencies are explicitly captured through the Transformer encoder and its MHSA multi-head self-attention mechanism to capture long-range dependencies, which significantly improves the model's robustness in suppressing non-stationary and nonlinear complex noise across time series. 4) This invention innovatively constructs a SEGA squeezing and excitation-gated attention module as an intelligent filter in skip connections. This module uses deep semantic information as a gating signal to dynamically recalibrate and filter shallow encoder features, ensuring that only the information most valuable for signal reconstruction is transmitted. This solves the problem of traditional skip connections transmitting noise and redundant information without selection, and ensures the accurate recovery of weak signals. 5) This invention employs a deep supervision strategy, adding auxiliary supervision signals at multiple levels of the decoder to ensure effective gradient backpropagation, accelerate convergence, and optimize multi-scale feature learning. The core loss function is the Charbonnier loss, which is robust to outliers, combined with residual connection design, allowing the model to focus on learning noise and preserving the structure and details of the original microseismic signal to the greatest extent possible. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the overall workflow of the present invention; Figure 2 yes Figure 1 MSF-GLANet model structure diagram; Figure 3 yes Figure 1 Flowchart of the MSFF multi-scale feature fusion module; Figure 4 yes Figure 1 Flowchart of the global-local fusion block module in GLFB; Figure 5 yes Figure 1Flowchart of SEGA squeeze and excitation gating attention module; Figure 6 This is a waveform diagram of a single-channel pure micro-vibration signal in the embodiment; Figure 7 This is a waveform diagram of a single-channel noisy micro-vibration signal in the embodiment; Figure 8 This is a waveform diagram of the denoising result of the U-Net network in the embodiment; Figure 9 This is a waveform diagram of the denoising result of the MSF-GLANet network in the embodiment; Figure 10 This is a time-frequency diagram of a single-channel pure microseismic signal in the embodiment; Figure 11 This is a time-frequency diagram of a single-channel noisy micro-vibration signal in the embodiment; Figure 12 This is a time-frequency diagram of the denoising results of the U-Net network in the embodiment; Figure 13 This is a time-frequency diagram of the denoising results of the MSF-GLANet network in the embodiment; Figure 14 This is a cross-sectional view of multiple pure microseismic signals in the embodiment; Figure 15 This is a cross-sectional view of the multi-channel noisy micro-vibration signal in the embodiment; Figure 16 This is a cross-sectional view of the denoising results of the U-Net network in the embodiment; Figure 17 This is a cross-sectional view of the denoising results of the MSF-GLANet network in the embodiment. Detailed Implementation

[0017] like Figure 1As shown, an intelligent denoising method for microseismic signals based on the MSF-GLANet model includes the following steps: Step 1: Constructing a large-scale clean microseismic signal dataset based on elastic wave field simulation; Step 2: Generating a noisy sample set combined with real environmental noise and data standardization preprocessing; Step 3: Scientifically dividing and preparing noisy / clean sample pairs according to statistical principles; Step 4: Constructing an MSF-GLANet hierarchical encoder-decoder network skeleton based on U-Net++ topology; Step 5: Integrating an MSFF multi-scale feature fusion module in the encoder path to capture multi-scale signal features; Step 6: Configuring a GLFB global-local fusion block in the network bottleneck layer to collaboratively model long-range dependencies and local context. Step 7: Replace traditional skip connections with SEGA squeeze and excitation-gated attention modules to achieve intelligent feature selection; Step 8: Construct a composite loss function with Charbonnier loss as the core and combine it with a multi-level deep supervision strategy; Step 9: Initialize the MSF-GLANet network using the training set and iteratively train it based on the composite loss function; Step 10: Monitor the model performance in real time using the validation set and save the best model weights; Step 11: Input the preprocessed real-world micro-seismic signal to be denoised into the trained model for inference; Step 12: Reconstruct the residual signal predicted by the network by adding it element-wise through the residual connection mechanism to achieve high-fidelity recovery of the denoised signal.

[0018] In step 1, the specific procedure is as follows: based on the actual geological conditions and exploration background parameters, elastic wave field forward modeling technology is used to generate a large-scale pure microseismic signal dataset with different sources, propagation paths and structural characteristics.

[0019] In step 2, the specific procedure is as follows: the system collects the real environmental noise of the target work area, and according to the preset signal-to-noise ratio distribution, it superimposes it with the clean signal generated in step 1 to form a noisy / clean sample pair; then, it performs uniform normalization and other standardized preprocessing operations on the entire dataset to eliminate the difference in data scale.

[0020] In step 3, the specific approach is as follows: to ensure the stability of model training and the fairness of evaluation, the preprocessed noisy / clean sample pairs are strictly divided into independent training sets, validation sets, and test sets to ensure the model's generalization ability under different data distributions.

[0021] In step 4, a hierarchical encoder-decoder network skeleton based on the U-Net++ dense skip connection topology is constructed. The core of this network integrates the MSFF multi-scale feature fusion module, the GLFB global-local fusion block, and the SEGA squeeze and excitation-gated attention module, specifically as follows: First, input signal After convolution, the resulting features are used as initial features to connect to the first layer. Module input; first layer, first module Module output characteristics .

[0022] For encoder paths, After the first layer of downsampling Then input to the first one of the second layer Module, first in the second layer Module output characteristics ;feature After the second layer of downsampling Then input to the first one of the third layer Module, first in the third layer Module output characteristics ;feature After the third layer of downsampling Then input to the first one of the fourth layer Module, first in the fourth layer Module output characteristics ;feature After the fourth layer of downsampling Then enter Module, Module output bottleneck layer features ; Decoder path from Upsampling and feature fusion begin sequentially. In the main network structure, the features of each decoder node are formed by concatenating and fusing upsampled features from the next layer and multiple predecessor features from the same layer. Feature fusion in skip connections employs... The module performs gating and filtering.

[0023] The specific steps for the first-level decoding are as follows: After upsampling, and Passing by together The features processed by the module are the second feature of the fourth layer. The input features of the module, the second one in the fourth layer. Module output characteristics . After upsampling, and Passing by together The features processed by the module are used as the second feature of the third layer. The input features of the module, the second one in the third layer Module output characteristics . After upsampling, and Passing by together The features processed by the module are used as the second layer. The input features of the module, the second layer, the second Module output characteristics . After upsampling, and Passing by together The features processed by the module are used as the first and second layers. The input features of the module, the first layer and the second layer Module output characteristics .

[0024] The specific steps for the second-level decoding are as follows: After upsampling, and and Passing by together The features processed by the module are used as the third layer. The input features of the module, the third layer, the third... Module output characteristics . After upsampling, and and Passing by together The features processed by the module are used as the second and third layers. The input features of the module, the second and third layers Module output characteristics . After upsampling, and and Passing by together The features processed by the module are used as the first and third layers. The input features of the module, the third layer of the first layer Module output characteristics .

[0025] The specific steps for the third-level decoding are as follows: After upsampling, and , and Passing by together The features processed by the module are used as the fourth feature of the second layer. The input features of the module, the fourth one in the second layer. Module output characteristics . After upsampling, and , and Passing by together The features processed by the module are the fourth in the first layer. The module's input features, the fourth one in the first layer. Module output characteristics .

[0026] The specific steps for Level 4 decoding are as follows: After upsampling, and , , and Passing by together The features processed by the module are the fifth in the first layer. The module's input features, the fifth one in the first layer. Module output characteristics .

[0027] In the main output path, After the fourth convolutional module Output and input signals Perform residual connections to obtain the main output information. .

[0028] In step 5, an MSFF module is integrated into the initial feature extraction layer at the network front end and each downsampling layer in the encoder path. This module synchronously captures multi-scale features of the signal using parallel multi-size convolutional kernels and employs SE squeezing and excitation mechanisms to dynamically weight the fused features at the channel level, thereby highlighting effective signal features. Specifically: Module receives input features And simultaneously input it into three parallel feature extraction branches: The specific steps for the first branch are as follows: Input features The input to the first 3×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 3×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 3×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs the first branch multi-scale features. .

[0029] The specific steps for the second branch are as follows: Input features The input to the first 5×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 5×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 5×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs the second branch multi-scale features. .

[0030] The specific steps for the third branch are as follows: Input features The input to the first 7×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 7×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 7×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs third-branch multi-scale features. .

[0031] Then we move to the main branch, which contains the multi-scale features of these three branches. , and By piecing them together, a fusion feature is obtained. and will As Input characteristics of the extrusion and excitation modules. In the module, fused features After global average pooling The output after the operation is squeezed is connected to the first fully connected layer. The input, the first fully connected layer Output connection The input to the activation function, The activation function output connects to the second fully connected layer. The input, the second fully connected layer Output connection The input to the activation function, Activation function output connected channel multiplication Input for the operation The operation will apply the learned channel weights. Channel weighted features are obtained. Channel weighting characteristics Connect the input of the first 3×1 convolutional layer in the main branch, and use the output features of the first 3×1 convolutional layer as the first batch normalization layer. The input features, the first batch normalization layer Output features as the first The input features of the activation function, the first The input features of the activation function are used as the input features of the second 3×1 convolutional layer, and the output features of the second 3×1 convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer Output features as the second The input features of the activation function, the second The output features of the activation function and Residual connection, to obtain The module's final output.

[0032] In step 6, at the highest abstraction layer of the network, i.e., the bottleneck layer, configure... The global-local fusion block employs a dual-branch parallel structure; where the local branch is implemented through... Obtain multi-granularity local context and integrate global branches The encoder explicitly models the long-range dependencies of the signal. Specifically: Module receives input features and connect it to Local branches and Input for the global branch.

[0033] exist In local branches, Feature connected to The inputs to the five parallel submodules are: 1×1 convolution path: The feature is connected to the input of a 1×1 convolutional layer, the output of which is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the first feature path. .

[0034] Path with expansion rate r=6: The feature connects to the input of a 3×1 convolutional layer with a dilation of 6, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the second-path features. .

[0035] Path with expansion rate r=12: The feature connects to the input of a 3×1 convolutional layer with a dilation of 12, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the third-path feature. .

[0036] Path with expansion rate r=18: The feature connects to the input of a 3×1 convolutional layer with a dilation of 18, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function outputs the third-path feature. .

[0037] Global pooling path: Feature connection global average pooling The input to the operation is global average pooling. The operation output is connected to the input of a 1×1 convolutional layer, and the output of this convolutional layer is connected to a batch normalization layer. normalization layer Output connected to Activation function The activation function output is connected to the input of the upsampling operation, and the output of the upsampling operation is the fifth feature. .

[0038] Will , , , and The five features are concatenated, and the concatenated features are connected to the input of a 1×1 output convolutional layer. This output convolutional layer outputs local branches of the features. .

[0039] exist In the global branch: Features are obtained after positional encoding. , The input is connected to the multi-head self-attention mechanism, in which... The features are first connected to three parallel linear transformation layers: Query the linear layer: Features are connected to the first linear layer The linear layer takes the input as input and outputs the query features. .

[0040] Bond linear layer: Features are connected to the second linear layer The linear layer takes the input as input and outputs the key features. .

[0041] Value linear layer: Features are connected to the third linear layer The input of the linear layer is the feature value of the output value. .

[0042] Query features Bond features Connect to matrix multiplication Input for the operation Operation output connected to scaling The input to the operation and the output of the scaling operation are connected to the mask. Input for the operation Operation output connected to The input to the activation function, The activation function outputs an attention weight matrix. This attention weight matrix is ​​then concatenated with a second matrix multiplication. The input for the operation, The operation combines the attention weight matrix with the value features. Multiply the results to output the scaled dot product attention result. This scaled dot product attention result is then connected to the concatenated input. The concatenation layer fuses the attention results obtained from the multi-head computation along the channel dimension. The concatenated features are then connected to the fourth linear layer. The input is the linear layer, and the output of this linear layer is the multi-head self-attention output. This output is then processed... Operation and After addition, connect to the first normalization layer. Input.

[0043] The output is connected to the input of the feedforward network. .exist internal: The output is connected to the first linear layer. Input, Output connection The input to the activation function, Output connection to the second The input, the second The output connects to the second linear layer. The input, the second Output connection to the third The input, get Output .

[0044] go through After operation and the first layer The input features are added together and used as the second layer. Input features, second layer The output yields the global branch output characteristics. .

[0045] In the final fusion stage, Local branch output features and Global branch output features The features are concatenated and then connected to the input of a 1×1 fusion convolutional layer. The output of this fusion convolutional layer is the... The module's final output.

[0046] In step 7, the traditional non-selective skip connection is replaced with This module utilizes semantic information from deeper layers of the decoder as gating signals to dynamically recalibrate and filter features from the shallow encoder, effectively suppressing redundant noise propagated non-selectively in skip connections. Specifically: The module receives features upsampled from the decoder. and from the same level layer of the encoder Two input signals. First, Connect to the first The module's input, Weighted gating signal of module output channel .at the same time, Connect to the second The module's input, Weighted skip connection feature of module output channel .

[0047] Next, in the feature projection path, the channel-weighted gated signal The input is connected to the first 1×1 convolutional layer, and the output of that convolutional layer is connected to the batch normalization layer. Obtain gated projection features Meanwhile, channel-weighted skip connectivity features The input is connected to a second 1×1 convolutional layer, the output of which is connected to a batch normalization layer. To obtain the jump projection features .

[0048] Gated projection features and jump projection features Element-level fusion is performed through addition, and the added features are connected to... The input to the activation function, The output is connected to the input of the attention coefficient generator. In the attention coefficient generator, The output of the first convolutional layer is connected to the input of the third 1×1 convolutional layer, and the output of the third convolutional layer is connected to the batch normalization layer. normalization layer Output connected to The input to the activation function, The activation function outputs an attention coefficient matrix. .

[0049] Finally, the attention coefficient matrix Connected to the input of the channel-level multiplication operation. Channel-weighted skip connectivity features Perform dot product to obtain gating and recalibration. The module's final output.

[0050] In step 8, a composite loss function is constructed with the Charbonnier loss function, which is robust to outliers, at its core. Furthermore, a deep supervision strategy is employed, adding auxiliary supervision signals at different levels of the decoder to ensure effective gradient backpropagation and accelerate the optimization of multi-scale features. Specifically: In the main network architecture, the second layer of the first layer Module output characteristics After the first convolutional module Get auxiliary output The third one on the first floor Module output characteristics After the second convolutional module Get auxiliary output The fourth one on the first floor Module output characteristics After the third convolutional module Get auxiliary output .

[0051] In step 9, the specific steps are as follows: the MSF-GLANet network is initialized and iteratively trained using the training set, and the gradient is calculated based on the composite loss function using the optimization algorithm and backpropagation is performed to continuously optimize the network parameters.

[0052] In step 10, the specific approach is as follows: monitor the model's performance metrics on unseen data in real time using the validation set. When the model's performance reaches its optimal level and no longer improves, stop training and save the model weights with the best performance at that stage.

[0053] In step 11, the specific procedure is as follows: the real environmental microseismic signal to be denoised is preprocessed in the same way as the training set, and then input into the trained MSF-GLANet model for inference.

[0054] In step 12, the specific method is as follows: using the residual connection mechanism of the final output layer of the network, the network predicts the residual signal used to compensate for noise, and reconstructs it by adding it element by element with the original noisy signal, and finally outputs a denoised micro-vibration signal with a high signal-to-noise ratio, so as to achieve accurate recovery of weak effective signals.

[0055] Example: The MSF-GLANet model structure is shown below. Figure 2 The specific implementation details are as follows: This network uses a hierarchical encoder-decoder structure, achieving cross-layer feature interaction through dense skip connections. Specifically, the network's feature maps consist of... It means that, among them Represents the downsampling level. Represents the convolutional layer; input signal First, an initial feature map is generated through a multi-scale feature fusion layer. For subsequent nodes in the encoder path, their features The features are obtained by downsampling the features from the previous layer; formulas (1) and (2) are the specific feature generation process; (1); (2); In formula (1) (2), This is a multi-scale feature fusion module. This refers to the convolution operation. This is a maximum pooling operation. Network bottleneck layer Adopting customization The global-local fusion block is used to model long-range dependencies and multi-scale contextual information at the highest level of abstraction, as shown in formula (3).

[0056] (3); In the decoder path, each node is densely skipped. The features are all formed by fusing features from all its sibling and lower-level predecessor nodes and upsampled features from lower levels of the encoder path. Figure 2 As can be seen from the various colored dashed lines, these feature flows are integrated by squeezing and stimulating the gating attention module, as detailed in formula (4).

[0057] (4); in, Representative feature concatenation. For upsampling operation, This represents the concatenation of all elements within the set. This is for squeezing and stimulating gating attention modules.

[0058] See the flowchart for the multi-scale feature fusion module. Figure 3 Its core innovation lies in achieving simultaneous capture of signal features under different receptive fields through parallel multi-scale convolutional topology, thereby effectively solving the problem of insufficient adaptability of fixed kernel convolution to multi-scale features of microseismic signals. This module contains three parallel convolutional branches, using one-dimensional convolutional kernels with kernel sizes of 3, 5, and 7 respectively. Each branch consists of three consecutive convolutional layers, a batch normalization layer, and an activation function. The output feature maps from these three parallel branches... The data is spliced ​​along the channel dimension to form a comprehensive multi-scale feature tensor. The specific process is shown in formula (5): (5); in, It is the input feature map. Represents a complete convolution branch operation, defining a kernel with a size of Basic convolutional blocks ,in For batch normalization operations, then The internal structure is defined in formula (6); (6); in, This is the LeakyReLU activation function (Leaky Rectified Linear Unit).

[0059] To address the potential redundancy and channel imbalance issues introduced by simple feature splicing, a feature was introduced after splicing. The squeezing and excitation module performs a channel-level attention mechanism on the fused features. First, the "squeezing" operation... Global average pooling compresses the concatenated multi-scale features along the spatial dimension to obtain channel statistical vectors. ,in This represents the number of channels. This step can be considered as encoding the global information for each channel, as shown in formula (7): (7); Among them, the Descriptors of each channel The specific calculation method is shown in formula (8): (8); in The length of the feature map is given; next, the "activation" step utilizes a lightweight fully connected network based on channel statistical vectors. Generate a set of channel-level attention weights This process consists of two fully connected layers and two activation functions, aiming to model the nonlinear dependencies between channels. The specific process is shown in formula (9): (9); in, and It is the weight matrix of the fully connected layer. The activation function is Sigmoid; to more clearly represent the complete process described above, the entire squeezing and excitation module is defined as a single function. For specific operations, see formula (10); (10); in This represents element-wise multiplication. Ultimately, these attention weights are applied channel-by-channel to the multi-scale feature tensor, dynamically recalibrating the importance of each channel to highlight effective signal features and suppress noise. After processing by the squeezing and activation module, the feature tensor undergoes further feature extraction through two basic convolutional blocks constructed from convolution-batch normalization-activation functions. Finally, the output is obtained by residual concatenation with the original input of the module. ;set up Therefore, this final output process can be formally represented as formula (11). (11); The global-local fusion block adopts a dual-branch parallel architecture; see the flowchart for details. Figure 4Local branches focus on capturing local features at different granularities; their core is... The dilated spatial pyramid pooling module effectively expands the receptive field of the convolution kernel by using dilated convolutions with different dilation rates in parallel, without increasing parameters or computational cost. The implementation employs five parallel branches: one convolution branch with a kernel size of 1, and three dilated convolution branches with a kernel size of 3, with dilation rates of 6, 12, and 18, respectively. In addition, the module includes a global average pooling branch to capture global contextual information. This process is detailed in formula (12). (12); in, These are the input features of the GLFB module. Is the expansion rate Hollow convolution.

[0060] In parallel, the global branch is used to explicitly model the long-range dependencies of the signal. At the heart of this branch is the Transformer encoder block, which utilizes... The multi-head self-attention mechanism computes the association between each location in the feature map and all other locations. This mechanism breaks the limitations of the local receptive field of convolutional layers, capturing global contextual information across the entire time series. This branch also integrates key components such as positional encoding and layer normalization to ensure stable training of the Transformer and accurate modeling of temporal information.

[0061] The internal operations of a Transformer encoder can be further decomposed. For input features... First, add position encoding. This is to provide the model with positional information for each element in the sequence when there is a lack of recurrent or convolutional structures. See formula (13) for details. (13); Subsequently, the feature is fed into a multi-head self-attention layer, which maps the input feature to a query, key, and value matrix and performs a scaled dot product attention operation; see formula (14) for details. (14); in, , , These represent the query, key, and value matrices, respectively. It is a scaling factor. Softmax is the activation function; the multi-head self-attention layer further divides the input features into multiple subspaces, performs attention calculations independently in each subspace, and then concatenates the outputs of all subspaces together, so that the model can pay attention to the features from different angles. The final attention output is the concatenation result of all heads; see formula (15) for details. (15); in , This is the learnable output projection weight matrix.

[0062] After the attention layer, residual connections and layer normalization are used. Residual connections help alleviate the vanishing gradient problem in deep networks, allowing information to flow smoothly in the network; while layer normalization is used to stabilize training and accelerate model convergence; see formula (16) for details. (16); in, For layer normalization operations, This is a regularization technique for random inactivation.

[0063] at last, Sent to a Feedforward Network The output is then combined with Perform residual connections and layer normalization to obtain the final global branch output. For the specific process, see formulas (17) and (18). (17); (18); in, It is a fully connected linear layer. The ReLU activation function; output from the global branch. Then, the output of the local branch. The layers are spliced ​​together and finally fused using a convolutional layer with a kernel size of 1. This parallel design ensures that the network simultaneously receives signals from various sources. It benefits from the fine local details captured and the wide-area global context information modeled by Transformer, making it highly robust in suppressing complex, scattered noise patterns; the final fusion process can be expressed as Equation (19).

[0064] (19); See the flowchart for the squeeze and excitation gating attention module. Figure 5 Its design aims to adaptively filter features in skip connections to achieve more optimized feature fusion. The structure of this module is as follows: Figure 4 As shown, it has two key inputs: Skip Connection characteristics and paths from the decoder Gating Signal; the core idea is that the gating signal (from the deeper, more semantically informative decoder path) should guide and filter skip connection features (from the shallower, more spatially detailed encoder path); specifically, the two input signals first pass through an independent squeezing and excitation module to dynamically adjust the importance of their channel-level features; this preprocessing step can be formalized as formula (20). (20); Subsequently, the two feature maps were processed by the SE module. and Each element is fed into a feature projection submodule, which consists of a convolutional layer with a kernel size of 1. It consists of a batch normalization layer. This step aims to align the feature dimensions of the two inputs for subsequent fusion operations. Then, the two projected features are added together and fed into a LeakyReLU activation function; see formula (21) for details. (twenty one); Next, the processed features The input is fed into an attention coefficient generator. This generator consists of a convolutional layer, a batch normalization layer, and a sigmoid activation function, used to generate an attention weight map. The specific process is shown in formula (22); (twenty two); Finally, the attention weight map is generated. Element-wise multiplication back to the original skip connection feature The above process enables dynamic recalibration of skip connection features; this process effectively suppresses features unrelated to the gating signal and highlights features that contribute to the denoising task, ultimately outputting the filtered features. The specific process can be seen in formula (23); (twenty three); This dual-attention strategy ensures that MSF-GLANet actively suppresses redundant information during the feature transfer process of skip connections, transmitting only the key information most valuable for signal reconstruction, thereby significantly improving denoising performance, especially in recovering weak signals that are heavily submerged in noise.

[0065] The vanishing gradient problem is a common challenge during the training of deep neural networks, especially in deep architectures with long skip connections like U-Net. Information originating from deep layers struggles to be effectively backpropagated to shallower layers, leading to slow convergence or poor performance. To address this challenge and accelerate model convergence, a deep supervision strategy is introduced. The core idea of ​​this strategy is to add auxiliary supervision signals at different levels of the decoder path, forcing the network to generate high-quality denoised predictions in the middle stages of feature recovery. This multi-point supervision mechanism ensures effective gradient backpropagation at each stage of the network, thereby promoting the model's collaborative learning and optimization of multi-scale features.

[0066] According to the architecture design of the present invention, the auxiliary output is led out from the intermediate output of the decoder path. Figure 2 As can be seen, from the second to the fourth layer of the decoder Three auxiliary outputs are derived, which, along with the final main output, are supervised. Each auxiliary output passes through an independent 1x1 convolutional layer (Conv) to reduce the number of channels to 1, and then the loss is calculated with the original clean signal, thus corresponding to... , and This design allows the model to be optimized at multiple granular levels during signal recovery.

[0067] To uniformly optimize all outputs, a combined loss function was constructed as the overall optimization objective of the model. This loss function is the main loss of the final output. The total loss is a weighted sum of the outputs from these three auxiliary tasks. This design not only balances the features learned by the network at different depths but also ensures that optimization of each auxiliary task directly contributes to improving the performance of the main task. The definition is given in formula (24); (twenty four); in, It is the final denoised signal Compared with the original clean signal The principal loss is calculated between them; It is the loss from the i-th auxiliary output; These are the weights used to control the contribution of each auxiliary loss. In this invention, to emphasize the importance of deep auxiliary tasks, incremental weights are set, specifically... .

[0068] Considering the possible outliers and sudden noise in the microseismic signal, the Mean Squared Error (MSE) loss function, which is sensitive to outliers, was not used. Instead, the Charbonnier loss function, which is more robust to outliers, was chosen. This loss function can handle large errors more smoothly, and its mathematical expression is shown in formula (25). (25); in, and These represent the clean signal and the denoised signal predicted by the network, respectively. It is the length of the signal; It is a local minimum (taken here) The main loss and auxiliary loss are used to ensure that the loss function remains stable when the gradient is close to zero. The complete forms of the main loss and auxiliary loss can be expressed as Equation (26).

[0069] (26); Furthermore, to further enhance the model's robustness and detail recovery capabilities, the final output layer of this invention employs a residual connection strategy. Unlike directly predicting the denoised signal, this method allows the network to learn only the noise components that need to be removed from the noisy signal, i.e., the noise residuals. Figure 1 As shown, the final denoised signal is obtained by adding the original noisy input signal to the denoised signal predicted by the network (i.e., the noise residual) element by element. This process can be formalized as formula (27).

[0070] (27); in, This represents all the module steps before the final residual connection; this residual connection method significantly simplifies the model's learning task, reducing it from learning the complete mapping function to learning only a residual function. In denoising tasks, noise signals are usually small perturbations in the original noisy signal, so learning the noise-compensated residual signal is easier to implement than learning the complete clean signal, and can effectively preserve the structure and details of the original signal.

[0071] To train the model's basic denoising capabilities and evaluate its performance limits at low signal-to-noise ratios (SNR), this invention constructs a noisy training dataset by superimposing conventionally distributed Gaussian white noise onto a clean microseismic signal. Under SNR conditions of -1dB, -5dB, and -10dB noise levels, this invention compares MSF-GLANet with traditional denoising methods such as bandpass filter (BPF), wavelet transform (WT), classic models like DnCNN and U-Net, and innovative models proposed in recent years such as DeepDenoiser and SeismicSuperResolution (SeisSR). Objective analysis is performed using three key performance indicators: signal-to-noise ratio (SNR), correlation coefficient (CC), and root mean square error (RMSE). Specific data are shown in Table 1.

[0072] Table 1: Comparison of denoising performance of various methods for synthesized signals containing Gaussian noise at different signal-to-noise ratio levels

[0073] Table 1 consistently demonstrates that the MSF-GLANet model proposed in this invention significantly outperforms all comparable methods at all tested signal-to-noise ratio (SNR) levels, establishing optimal performance in the three key metrics of SNR, CC, and RMSE. Specifically, compared to the U-Net baseline model, MSF-GLANet achieves improvements in SNR of over 74.76%, 69.65%, and 67.29%, respectively. Simultaneously, MSF-GLANet consistently maintains the lowest RMSE and the highest CC value, directly proving that the model not only achieves high amplitude fidelity during denoising but also fully preserves the waveform morphology and phase characteristics of the signal.

[0074] from Figures 6 to 13 In a comprehensive comparative analysis of experimental results, the MSF-GLANet method of this invention demonstrates significant performance advantages over the traditional U-Net network in complex microseismic signal denoising tasks. In time-domain analysis (… Figures 6-9 In frequency domain analysis, U-Net denoising often results in amplitude attenuation and over-smoothing in the instantaneous peak region of the signal, leading to severe waveform distortion. In contrast, MSF-GLANet, with its precise structure, can achieve lossless, high-fidelity reconstruction of weak signals, restoring the amplitude and instantaneous structure of the original signal to the greatest extent, effectively ensuring the integrity and interpretability of the signal. Figures 10-13 In terms of noise reduction, the time-frequency background after U-Net denoising still retains non-stationary and chaotic noise energy; while the time-frequency map generated by MSF-GLANet has a highly pure background, and the target signal energy clusters have sharp boundaries and concentrated energy, which fully demonstrates its robust and thorough suppression capability against non-stationary and nonlinear noise.

[0075] Figures 14 to 17 The overall effect comparison of denoising on microseismic signal profiles is clearly demonstrated. Facing Figure 15 As shown in the heavily noise-contaminated profile, the MSF-GLANet method significantly outperforms the traditional U-Net network. While U-Net reduces overall noise, significant random noise remains in the profile background, and the recovered phase axes exhibit breaks and energy dispersion, lacking necessary spatial coherence. In contrast, MSF-GLANet successfully removes and focuses more weak but effective signals from the high-intensity noise background, resulting in a high concentration of signal energy clusters along the trace number direction. Its recovered phase axes demonstrate good continuity and structural integrity between adjacent traces, and the profile background is closer to a clean state. This fully demonstrates that MSF-GLANet has a thorough ability to suppress random noise in the profile and effectively maintains the spatial structural characteristics of the microseismic signal, greatly improving data interpretability and local signal-to-noise ratio.

[0076] In summary, this invention addresses the core shortcomings of existing deep learning denoising methods in processing complex microseismic signals, such as insufficient multi-scale feature capture, limited long-range dependency modeling, and noise introduced by skip connections. It proposes a novel intelligent denoising network, MSF-GLANet (Multi-Scale Feature Fusion and Global-Local Attention Network), based on an improved U-Net++ architecture. The innovations and key structures of this method include: (1) The network designs a multi-scale feature fusion module, MSFF (Multi-Scale Feature Fusion), which uses parallel multi-size convolutional kernels to synchronously process input features. The module incorporates a squeeze-and-excitation mechanism (SE) to perform channel-level feature weighting on the concatenated multi-channel features. (2) The network introduces a Global-Local Fusion Block (GLFB) at the bottleneck layer and adopts a dual-branch structure: the local branch uses Atrous Spatial Pyramid Pooling (ASPP) to obtain multi-granular local features through dilated convolutions with different dilation rates; the global branch integrates a Transformer encoder and uses Multi-Head Self-Attention (MHSA) to explicitly capture the long-range dependencies of the signal. (3) To address the problem that skip connections in U-Net are prone to introducing noise, the network constructs a Squeeze and Excitation Gated Attention (SEGA) module to replace the skip connections in the traditional U-Net. The gated attention mechanism in the module uses semantic information from the deep layers of the decoder as a gate signal to dynamically recalibrate and filter the features of the shallow encoder. (4) This method introduces a deep supervision and composite loss strategy to add auxiliary supervision signals at different levels of the decoder. The core optimization employs Charbonnier loss, which is robust to outliers, and combines it with a residual connection design, enabling the network to focus on learning and predicting noisy residuals.

Claims

1. A microseismic signal denoising method fusing multi-scale features and global-local attention, characterized in that, The method comprises the following steps: Step 1: Construct a large-scale pure microseismic signal dataset based on elastic wave field simulation; Step 2: Generate data standardization preprocessing combined with noisy sample set of real environment noise; Step 3: Scientifically divide and prepare the noisy / pure sample pairs; Step 4: Construct the MSF-GLANet hierarchical encoding-decoding network skeleton based on the U-Net++ topology; Step 5: Integrate the MSFF multi-scale feature fusion module in the encoder path to capture multi-scale signal features; Step 6: Configure the GLFB global-local fusion block at the network bottleneck layer to cooperatively model long-range dependencies and local context; Step 7: Implement intelligent filtering of features through the SEGA squeezing and excitation gate attention module; Step 8: Construct a composite loss function centered on the Charbonnier loss and combine a multi-level deep supervision strategy; Step 9: Use the training set to initialize and iteratively train the MSF-GLANet network based on the composite loss function; Step 10: Monitor the model performance in real time through the validation set and save the best model weights; Step 11: Preprocess the real environment microseismic signal to be denoised and input it into the trained model for inference; Step 12: Add the residual signals predicted by the network through the residual connection mechanism to reconstruct the denoised signal with high fidelity.

2. The method of claim 1, wherein, In step 1, according to the actual geological conditions and exploration background parameters, an elastic wave field forward modeling technique is used to generate a large-scale pure microseismic signal dataset with different sources, propagation paths, and structural characteristics.

3. The method of claim 1, wherein, In step 2, real environment noise is collected in the target work area, and according to the preset signal-to-noise ratio distribution, it is superimposed with the pure signal generated in step 1 to form a noisy / pure sample pair; then, uniform normalization and standardization preprocessing operations are performed on the entire dataset to eliminate differences in data scales.

4. The method of claim 1, wherein, In step 3, to ensure the stability of model training and the fairness of evaluation, the preprocessed noisy / pure sample pairs are strictly divided into independent training, validation, and test sets to ensure the model's generalization ability under different data distributions.

5. The method of claim 1, wherein, In step 4, the MSF-GLANet hierarchical encoding-decoding network structure based on the U-Net++ topology includes the MSFF multi-scale feature fusion module, the GLFB global-local fusion block, and the SEGA squeezing and excitation gate attention module; the network structure is specifically as follows: input signal after convolution operation as initial feature connection first layer first module input; first layer first module output feature ; For the encoder path, After first layer down-sampling is input to a second layer first module, the second layer first module outputs a feature ; the feature After second layer down-sampling is input to a third layer first module, the third layer first module outputs a feature ; the feature After third layer down-sampling is input to a fourth layer first module, the fourth layer first module outputs a feature ; the feature After fourth layer down-sampling is input to a module, the module outputs a bottleneck layer feature .

6. The method of claim 5, wherein, The decoder path starts with features starting with upsampling and feature fusion step by step; In the main network structure, the features of each decoder node are formed by concatenating and fusing the up-sampled features from the next level and multiple predecessor features of the same level, wherein the feature fusion in the skip connection adopts The module performs gated screening.

7. The method of claim 6, wherein, For the decoder, specifically: The first-level decoding is specifically: Features after up-sampling operations, along with features processed by the module, as input features to a second module of the fourth layer, a second module of the fourth layer outputs features ; features after up-sampling operations, along with features processed by the module, as input features to a second module of the third layer, a second module of the third layer outputs features ; features after up-sampling operations, along with features processed by the module, as input features to a second module of the second layer, a second module of the second layer outputs features ; features after up-sampling operations, along with features processed by the module, as input features to a second module of the first layer, a second module of the first layer outputs features ;​​​​ The second-level decoding is specifically: feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the third layer. The input features of the module, the third layer, the third... Module output characteristics ;feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the second and third layers. The input features of the module, the second and third layers Module output characteristics ;feature After upsampling, with features and characteristics Passing by together The features processed by the module are used as the first and third layers. The input features of the module, the third layer of the first layer Module output characteristics ; The third-level decoding is specifically: feature After upsampling, with features ,feature and characteristics Passing by together The features processed by the module are used as the fourth feature of the second layer. The input features of the module, the fourth one in the second layer. Module output characteristics ;feature After upsampling, with features ,feature and characteristics Passing by together The features processed by the module are the fourth in the first layer. The module's input features, the fourth one in the first layer. Module output characteristics ; The fourth-level decoding is specifically: feature After upsampling, with features ,feature ,feature and characteristics Passing by together The features processed by the module are the fifth in the first layer. The module's input features, the fifth one in the first layer. Module output characteristics ; In the main output path, the feature output of the fourth convolution module is connected with the input signal to obtain the main output information .

8. The method according to one of claims 1 to 7, characterized in that, In step 5, the MSFF module is integrated in the initial feature extraction layer at the front end of the network and each down-sampling layer in the encoder path, which synchronously captures multi-scale features of the signal through parallel multi-size convolution kernels, and dynamically weights the fused features at the channel level using the SE squeezing and excitation mechanism to highlight effective signal features, specifically as follows: The module receives input features and inputs them simultaneously to three parallel feature extraction branches; The specific steps of the first branch are as follows: input feature of the first 3x1 convolutional layer output feature of the first convolutional layer as the input feature of the first batch normalization layer output feature of the first batch normalization layer output feature of the second batch normalization layer as the input feature of the second 3x1 convolutional layer output feature of the second convolutional layer as the input feature of the second batch normalization layer output feature of the third batch normalization layer as the input feature of the third 3x1 convolutional layer output feature of the third convolutional layer as the input feature of the third batch normalization layer output feature of the third batch normalization layer input feature of the activation function output first branch multi-scale feature of the activation function ; The specific steps of the second branch are as follows: Input features The input to the first 5×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 5×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 5×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs the second branch multi-scale features. ; The specific steps of the third branch are as follows: Input features The input to the first 7×1 convolutional layer is connected, and the output features of the first convolutional layer are used as the first batch normalization layer. The input features, the first batch normalization layer The output features are used as the input features of the second 7×1 convolutional layer, and the output features of the second convolutional layer are used as the second batch normalization layer. The input features, the second batch normalization layer The output features are used as the input features of the third 7×1 convolutional layer, and the output features of the third convolutional layer are used as the third batch normalization layer. The input features, the third batch normalization layer Output features as Input features of the activation function The activation function outputs a third branch multi-scale feature. ; Subsequently into the total branch, the three branches of multi-scale features 、 and Spliced together to get fusion features , and As squeeze and excitation module input features; Squeeze and excitation module output channel weighting features , channel weighting features ; Channel weighting features Connect the input of the first 3x1 convolution layer of the total branch, the output features of the first 3x1 convolution layer as the input features of the first batch normalization layer , the output features of the first batch normalization layer As the input features of the first Activation function, the input features of the first Activation function as the input features of the second 3x1 convolution layer, the output features of the second 3x1 convolution layer as the input features of the second batch normalization layer , the output features of the second batch normalization layer As the input features of the second Activation function, the output features of the second Activation function and Residual connection, get The final output of the module.

9. The method of claim 8, wherein, exist In the module, fused features After global average pooling The output after the operation is squeezed is connected to the first fully connected layer. The input, the first fully connected layer Output connection The input to the activation function, The activation function output connects to the second fully connected layer. The input, the second fully connected layer Output connection The input to the activation function, Activation function output connected channel multiplication Input for the operation The operation will apply the learned channel weights. Channel weighted features are obtained. .

10. The method according to one of claims 1 to 7, characterized in that, In step 6, the network is configured at the highest abstraction level, i.e. the bottleneck layer The global-local fusion block adopts a double-branch parallel structure; wherein the local branch is integrated by Multi-granularity local context is obtained, and the global branch is integrated The long-range dependence relationship of the signal is explicitly modeled by the encoder; specifically: The module receives input features and connects them simultaneously to the input of a local branch and the input of a global branch; In In the local branch, The feature is connected to five parallel sub-module inputs: 1x1 convolution path: The input of the feature connection 1x1 convolution layer is connected to the output of the batch normalization layer The output of the batch normalization layer is connected to the input of the activation function The output of the activation function is the first path feature ; Expansion rate r = 6 path: The input of the feature connection 3 x 1 expansion rate 6 convolution layer, the output of the convolution layer is connected to the batch normalization layer , the batch normalization layer The output is connected to The activation function, The output of the activation function is the second path feature ; Expansion rate r = 12 path: The input of the feature connection 3 x 1 expansion rate 12 convolution layer, the output of the convolution layer is connected to the batch normalization layer , the batch normalization layer The output is connected to The activation function, The output of the activation function is the third path feature ; Expansion rate r = 18 path: The input of the feature connection 3 x 1 expansion rate 18 convolution layer, the output of the convolution layer is connected to the batch normalization layer , the batch normalization layer The output is connected to The activation function, The output of the activation function is the third road feature ; Global pooling path: Feature concatenation global average pooling Input of the operation, global average pooling Input of the operation output, 1x1 convolutional layer whose output is connected to a batch normalization layer Batch normalization layer Output is connected to Activation function Activation function output is connected to the input of the up-sampling operation, the output of which is the fifth feature ; The five-way features of , , , and are spliced, and the spliced features are connected to the input of a 1x1 output convolutional layer, and the output convolutional layer outputs local branch output features . In In global branches: Features after position encoding , Input to the multi-headed self-attention mechanism, in which, Features are first connected to three parallel linear transformation layers: query linear layer: The features are concatenated to the input of a first linear layer which outputs the query features ; key linear layer: the input of a second linear layer which outputs the key feature ; Values linear layer: The feature is connected to the input of a third linear layer which outputs values feature ; query features and key features connected to a matrix multiplication input of an operation, operation output connected to a scaling input of an operation, scaling operation output connected to a masking input of an operation, operation output connected to input of an activation function, activation function output attention weight matrix; The attention weight matrix is connected to a second matrix multiplication An input to the operation, the The operation multiplies the attention weight matrix with the value feature and outputs a result of scaled dot-product attention; The scaling dot product attention result is connected to the concatenated input. The concatenation layer fuses the attention results obtained from multi-head computation along the channel dimension. The concatenated features are then connected to the fourth linear layer. The input is the linear layer, and the output of this linear layer is the multi-head self-attention output; this output is then processed... Operation and After addition, connect to the first normalization layer. Input; output is connected to the input of the feedforward network ; in the interior: output is connected to the input of the first linear layer , the output is connected to the input of the activation function , the output is connected to the input of the second linear layer, the output is connected to the input of the second activation function, the output is connected to the input of the third linear layer, the output is connected to the input of the third activation function, the output is connected to the input of the fourth linear layer, the output is connected to the input of the fourth activation function, the output is connected to the input of the fifth linear layer, the output is connected to the input of the fifth activation function, the output is connected to the input of the sixth linear layer, the output is connected to the input of the sixth go through After operation and the first layer The input features are added together and used as the second layer. Input features, second layer The output yields the global branch output characteristics. ; In the final fusion stage, Local branch output feature With Global branch output feature Spliced, the spliced feature is connected to the input of a 1x1 fusion convolution layer, and the output of the fusion convolution layer is the final output of the module. Module In step 7, the traditional non-selective skip connection is replaced by The module uses semantic information from the deep layer of the decoder as a gating signal to realize dynamic re-scaling and screening of the shallow layer encoder features, so as to effectively suppress the redundant noise in the non-selective transmission in the skip connection; specifically, it is: The module receives features upsampled from the decoder. and from the same level layer of the encoder Two input signals; first, Connect to the first The module's input, Weighted gating signal of module output channel ;at the same time, Connect to the second The module's input, Weighted skip connection feature of module output channel ; Then, in the feature projection path, the channel-weighted gating signal connected to the input of a first 1x1 convolutional layer whose output is connected to a batch normalization layer to obtain a gating projection feature ; meanwhile, the channel-weighted skip-connection feature connected to the input of a second 1x1 convolutional layer whose output is connected to a batch normalization layer to obtain a skip-connection projection feature ; Gated projection features and skip projection features Element-wise fusion by addition operation, the added features are concatenated to the input of an activation function, the output is connected to the input of an attention coefficient generator; in the attention coefficient generator, the output is connected to the input of a third 1x1 convolutional layer, the output of which is connected to a batch normalization layer , the batch normalization layer outputs are connected to the input of an activation function, the activation function outputs an attention coefficient matrix ; Finally, the attention coefficient matrix connected to the input of the channel-wise multiplication operation, with channel-weighted skip-connection features point-wise multiplication, resulting in the gated and re-scaled final output of the module; In step 8, a composite loss function is constructed with the Charbonnier loss function as the core, which is robust to outliers; and a depth supervision strategy is combined to add auxiliary supervision signals at different levels of the decoder, ensuring effective gradient backpropagation and accelerating the optimization of multi-scale features. Specifically: In the main network architecture, the first layer second module output features through the first convolutional module get auxiliary output , the first layer third module output features through the second convolutional module get auxiliary output , the first layer fourth module output features through the third convolutional module get auxiliary output .

Images