Microseismic signal processing method for waveform high-fidelity restoration and first arrival reservation
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2026-05-22
- Publication Date
- 2026-08-07
AI Technical Summary
[0003]本发明的目的在于克服传统去噪方法在面对低信噪比微震信号时易导致初至拾取不准确、尾波细节丢失以及传统自注意力机制存在计算复杂度较高的问题，提出一种基于波形高保真还原与初至保留的微震信号处理方法
[0019]（1）微弱初至事件识别能力提升：方法对低幅值、低能量的初至波具有更高敏感性，在复杂噪声背景下仍能保持初至到时特征的精确性，为震相拾取与速度反演提供更可靠输入；
Smart Images
Figure CN122525635A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mine safety monitoring and geophysical signal processing technology, specifically relating to a microseismic signal processing method that achieves high-fidelity waveform restoration and first-arrival preservation. This method extracts residual features from noisy signals using a deep learning model, achieving high-fidelity restoration of the first arrival and wake waves of microseismic signals. Background Technology
[0002] In fields such as mining production, oil extraction, and geological disaster early warning, microseismic monitoring technology is a crucial means of ensuring safe production. By monitoring microseismic signals generated by rock mass fracturing in real time, researchers can analyze the source location, magnitude, and focal mechanism, thus providing early warning information for dynamic disasters such as rockbursts, roof collapses, and water hazards. However, in actual mine monitoring environments, microseismic signals are often surrounded by strong background noise such as mechanical vibrations and electromagnetic interference, resulting in extremely low signal-to-noise ratios. Existing models cannot accurately identify the first arrival of microseismic signals during denoising, affecting the accuracy of subsequent analysis and positioning. Because the tailwave energy of microseismic signals is weak and decays rapidly, it is easily misidentified as noise and over-suppressed, leading to waveform damage. Therefore, there is an urgent need for an intelligent microseismic signal processing method that can effectively suppress complex background noise, faithfully reproduce the first arrival waveform and tailwave details, and possess efficient computational performance. Summary of the Invention
[0003] The purpose of this invention is to overcome the problems of inaccurate first arrival picking and loss of wake wave details in traditional denoising methods when dealing with low signal-to-noise ratio microseismic signals, as well as the high computational complexity of traditional self-attention mechanisms. This invention proposes a microseismic signal processing method based on high-fidelity waveform restoration and first arrival preservation. This method achieves high-fidelity denoising of microseismic signals by fusing dual-path feature extraction, a linear attention mechanism, and a multi-scale feature enhancement module, thereby improving the scientific reliability and real-time processing efficiency of microseismic monitoring data. This invention includes the following steps:
[0004] 1. A method for high-fidelity waveform restoration and first-arrival preservation in microseismic signal processing, characterized by the following specific steps:
[0005] S1. Import the noisy micro-vibration signal and normalize the signal;
[0006] S2. By constructing a U-Net microseismic signal processing network that enables high-fidelity waveform restoration and first-arrival preservation, multi-scale features are extracted using a dual-path encoder and a multi-scale feature enhancement module (MSFE). Combined with a linear attention mechanism and a high-fidelity shallow feature replacement mechanism, high-fidelity denoising of microseismic signals under complex noise environments is achieved. The specific steps are as follows:
[0007] S2.1. A dual-path shallow encoder is used to extract initial features from the input signal. The first branch Enc0 uses a one-dimensional convolutional layer with a kernel size of 3 and combines grouped normalization (GN) and ReLU activation functions to extract local fine features. The second branch Enc1 uses a one-dimensional convolutional layer with a kernel size of 7 to capture long-term temporal context information. Then, a residual attention module is connected in series to achieve multi-scale feature enhancement.
[0008] S2.2. A hierarchical encoder is used to extract multi-level deep feature representations step by step. Encoders Enc2, Enc3, and Enc4 are used respectively to increase the number of feature channels from 48 to 96, 192, and finally to 288. Enc2 and Enc3 are configured with two residual attention modules for feature enhancement, and Enc4 is configured with one residual attention module for feature enhancement. An average pooling layer with a stride of 2 is used between each encoder to achieve downsampling operation and complete scale compression.
[0009] S2.3. The residual attention module first extracts local features through convolutional layers and grouping normalization operations, then combines channel attention and spatial attention mechanisms to achieve adaptive enhancement of feature response, and finally fuses with input features through residual connections;
[0010] S2.4. Features are weighted by channel attention and spatial attention. Channel attention generates channel weights based on multilayer perceptron (MLP), and spatial attention generates spatial weights based on channel statistical features and through convolution and sigmoid function, so as to enhance key channels and important temporal positions.
[0011] S2.5. By introducing a linear attention mechanism into the bottleneck layer, it is possible to efficiently capture long-term temporal context information;
[0012] S2.6. Use a multi-scale feature enhancement module to perform multi-scale feature fusion on bottleneck features and Enc3, Enc2, and Enc1 skip connection features. Capture information such as first arrival time, periodic structure, and tail wave attenuation through a multi-branch structure, and combine convolution, grouping normalization, ReLU function activation, and residual connection to achieve feature enhancement.
[0013] S2.7. Finally, a multi-level decoder is used to recover the signal step by step. Each level of the decoder achieves feature recovery through convolutional upsampling and uses linear interpolation to complete the time dimension alignment. Then, the upsampled features are concatenated with the skip connection features enhanced by MSFE. Feature extraction is completed through convolution, group normalization, ReLU function activation and residual attention module. In the final decoder layer Dec1, the high-fidelity shallow features extracted by Enc0 are used to replace the Enc1 features in order to preserve the fine structural information of the initial time step.
[0014] S2.8. The output features at the end of the decoder are first integrated through a convolutional layer, and then processed by grouping normalization and ReLU activation in sequence;
[0015] S2.9. Construct a microseismic signal dataset. After preprocessing the original microseismic signals, generate microseismic signals through the convolution method, and introduce real noise and synthetic noise to simulate real microseismic signals in order to construct sample data under different signal-to-noise ratio conditions. Divide the dataset into training set, validation set and test set according to a certain ratio to improve the generalization ability of the network model.
[0016] S3. The constructed residual learning network is used to perform forward propagation calculation on the training samples. The difference between the network output and the target signal is evaluated by joint constraints. The backpropagation algorithm is used to iteratively update and optimize the network parameters until the model converges. Finally, the optimal denoising model with high-fidelity waveform reconstruction and first arrival preservation is obtained.
[0017] S4. Input the noisy micro-vibration signal into the trained optimal denoising model for denoising processing, and output the corresponding high-fidelity waveform and first-arrival preserved signal.
[0018] The microseismic signal processing method of the present invention, which provides high-fidelity waveform restoration and first-arrival preservation, has the following advantages:
[0019] (1) Improved ability to identify weak first arrival events: The method is more sensitive to low amplitude and low energy first arrival waves, and can still maintain the accuracy of first arrival time characteristics in complex noise background, providing more reliable input for phase picking and velocity inversion;
[0020] (2) High fidelity of waveform restoration: By combining residual learning with waveform constraint design for the first arrival region and the tail wave region, the problems of inaccurate first arrival picking, loss of waveform details and tail wave energy attenuation that are easy to occur in the traditional microseismic denoising process are effectively improved. The denoised signal maintains a high degree of consistency with the original microseismic signal in terms of waveform shape, amplitude characteristics and energy distribution, thereby improving the accuracy of microseismic event location and energy assessment.
[0021] (3) High efficiency of real-time monitoring: The introduction of the linear attention mechanism enables the model to maintain linear complexity when processing high-frequency sampling long sequences, effectively reducing computational complexity and storage requirements, and meeting the real-time requirements of mine monitoring;
[0022] (4) Physical consistency of output results: Through residual reconstruction and inverse normalization steps, the output denoised waveform is ensured to not only eliminate background noise, but also to maintain consistency with the real microseismic signal in terms of initial arrival, waveform shape and spectral distribution, thus achieving high-fidelity recovery. Attached Figure Description
[0023] To more clearly illustrate the technical solutions and embodiments of the present invention, the accompanying drawings used in the technical description and embodiments are briefly introduced below. The drawings are provided to further understand the embodiments of the present invention and constitute a part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention.
[0024] Figure 1 This is a flowchart of a microseismic signal processing method that achieves high-fidelity waveform restoration and first arrival preservation.
[0025] Figure 2 This is a schematic diagram of the U-Net network structure for high-fidelity waveform restoration and first-arrival preservation.
[0026] Figure 3 This is a schematic diagram of training and validation loss, as well as validation set SNR.
[0027] Figure 4 This is a schematic diagram illustrating the denoising effect on the synthesized signal. Detailed Implementation
[0028] To more clearly illustrate the technical solution and beneficial effects of the micro-vibration signal denoising method with high-fidelity waveform restoration and first-arrival preservation proposed in this invention, the embodiments of this invention are described in detail with reference to the accompanying drawings. The specific implementation methods of this invention are as follows:
[0029] 1. A method for high-fidelity waveform restoration and first-arrival preservation in microseismic signal processing, characterized by the following specific steps:
[0030] S1. Import the noisy micro-vibration signal and normalize the signal;
[0031] S2. By constructing a U-Net microseismic signal processing network that enables high-fidelity waveform restoration and first-arrival preservation, multi-scale features are extracted using a dual-path encoder and a multi-scale feature enhancement module (MSFE). Combined with a linear attention mechanism and a high-fidelity shallow feature replacement mechanism, high-fidelity denoising of microseismic signals under complex noise environments is achieved. The specific steps are as follows:
[0032] S2.1. Adopt as follows Figure 2 The dual-path shallow encoder in the model performs initial feature extraction on the input signal. The first branch, Enc0, uses a one-dimensional convolutional layer with a kernel size of 3 and combines grouped normalization (GN) and ReLU activation functions to extract local fine features. The second branch, Enc1, uses a one-dimensional convolutional layer with a kernel size of 7 to capture long-term temporal context information. Then, a residual attention module is connected in series to achieve multi-scale feature enhancement.
[0033] S2.2. A hierarchical encoder is used to extract multi-level deep feature representations step by step. Encoders Enc2, Enc3, and Enc4 are used respectively to increase the number of feature channels from 48 to 96, 192, and finally to 288. Enc2 and Enc3 are configured with two residual attention modules for feature enhancement, and Enc4 is configured with one residual attention module for feature enhancement. An average pooling layer with a stride of 2 is used between each encoder to achieve downsampling operation and complete scale compression.
[0034] S2.3. The residual attention module first extracts local features through convolutional layers and grouping normalization operations, then combines channel attention and spatial attention mechanisms to achieve adaptive enhancement of feature response, and finally fuses with input features through residual connections;
[0035] S2.4. Features are weighted by channel attention and spatial attention. Channel attention generates channel weights based on multilayer perceptron (MLP), and spatial attention generates spatial weights based on channel statistical features and through convolution and sigmoid function, so as to enhance key channels and important temporal positions.
[0036] S2.5. By introducing a linear attention mechanism into the bottleneck layer, it is possible to efficiently capture long-term temporal context information;
[0037] S2.6. Use a multi-scale feature enhancement module to perform multi-scale feature fusion on bottleneck features and Enc3, Enc2, and Enc1 skip connection features. Capture information such as first arrival time, periodic structure, and tail wave attenuation through a multi-branch structure, and combine convolution, grouping normalization, ReLU function activation, and residual connection to achieve feature enhancement.
[0038] S2.7. Finally, a multi-level decoder is used to recover the signal step by step. Each level of the decoder achieves feature recovery through convolutional upsampling and uses linear interpolation to complete the time dimension alignment. Then, the upsampled features are concatenated with the skip connection features enhanced by MSFE. Feature extraction is completed through convolution, group normalization, ReLU function activation and residual attention module. In the final decoder layer Dec1, the high-fidelity shallow features extracted by Enc0 are used to replace the Enc1 features in order to preserve the fine structural information of the initial time step.
[0039] S2.8. The output features at the end of the decoder are first integrated through a convolutional layer, and then processed by grouping normalization and ReLU activation in sequence;
[0040] S2.9. Construct a microseismic signal dataset. After preprocessing the original microseismic signals, generate microseismic signals through the convolution method, and introduce real noise and synthetic noise to simulate real microseismic signals in order to construct sample data under different signal-to-noise ratio conditions. Divide the dataset into training set, validation set and test set according to a certain ratio to improve the generalization ability of the network model.
[0041] S3. The constructed residual learning network is used to perform forward propagation calculation on the training samples. The difference between the network output and the target signal is evaluated by joint constraints. The backpropagation algorithm is used to iteratively update and optimize the network parameters until the model converges. Finally, the optimal denoising model with high-fidelity waveform reconstruction and first arrival preservation is obtained.
[0042] S4. Input the noisy micro-vibration signal into the trained optimal denoising model for denoising processing, and output the corresponding high-fidelity waveform and first-arrival preserved signal.
[0043] Implementation examples of the present invention:
[0044] When training the deep learning model, this invention selects 550 clean microseismic signals, generates microseismic signals through the convolution method, and introduces real noise and synthetic noise to simulate real microseismic signals in order to construct sample data under different signal-to-noise ratio conditions. In the model testing stage, 55 microseismic signals that were not used in training are selected to construct an independent test set to verify the denoising performance and generalization ability of the model.
[0045] Figure 1 This is a schematic diagram of the overall structure of the U-Net microseismic signal processing network based on high-fidelity waveform restoration and first-arrival preservation proposed in this invention.
[0046] Figure 2 The diagram shows the U-Net network structure for high-fidelity waveform reconstruction and first arrival preservation. As can be seen from the figure, the network consists of a dual-path shallow encoder, a hierarchical deep encoder, a linear attention bottleneck layer, a multi-scale feature enhancement module (MSFE), and a multi-level decoder, achieving efficient feature modeling and noise suppression of microseismic signals.
[0047] Figure 3 Figure 1 shows the loss function curve during the model training process. In Figure 2(a), the blue line is the loss curve during the training phase. As the number of iterations increases, the loss function gradually decreases and tends to stabilize, indicating that the model can effectively learn the mapping relationship between signal and noise. In Figure 3(a), the yellow line is the loss curve during the validation phase, and Figure 4(b) is the curve of the validation set SNR. As can be seen from the figures, the validation loss changes in the same trend as the training loss, and there is no obvious overfitting or oscillation phenomenon, indicating that the model has good generalization ability.
[0048] Figure 4Figure 1 shows a comparison of the denoising effects. Figure 2 shows the original signal, Figure 3 shows the time spectrum of the original signal after generalized S-transformation, Figure 4 shows the noise signal after adding 10dB Gaussian white noise, Figure 5 shows the time spectrum of the noise signal after generalized S-transformation, Figure 6 shows the signal after denoising with the optimal model, and Figure 7 shows the time spectrum of the denoised signal after generalized S-transformation. The method of this invention can effectively recover the main waveform structure of the microseismic signal, including the first arrival time and the main energy distribution. Compared with traditional methods, this method can better preserve the first arrival time and waveform and other main features of the signal while suppressing noise.
[0049] The above embodiments are only used to illustrate the present invention. The implementation steps of the method can be varied. Any equivalent transformations and improvements made on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.
Claims
1. A method for microseismic signal processing with high-fidelity waveform restoration and first arrival preservation, characterized in that... The following specific steps are adopted: S1. Import the noisy micro-vibration signal and normalize the signal; S2. By constructing a U-Net microseismic signal processing network that enables high-fidelity waveform restoration and first-arrival preservation, multi-scale features are extracted using a dual-path encoder and a multi-scale feature enhancement module (MSFE). Combined with a linear attention mechanism and a high-fidelity shallow feature replacement mechanism, high-fidelity denoising of microseismic signals under complex noise environments is achieved. The specific steps are as follows: S2.
1. A dual-path shallow encoder is used to extract initial features from the input signal. The first branch Enc0 uses a one-dimensional convolutional layer with a kernel size of 3, and combines grouped normalization (GN) and ReLU activation function to extract local fine features. The second branch, Enc1, uses a one-dimensional convolutional layer with a kernel size of 7 to capture long-term temporal context information, and then connects a residual attention module to achieve multi-scale feature enhancement. S2.
2. A hierarchical encoder is used to extract multi-level deep feature representations step by step. Encoders Enc2, Enc3, and Enc4 are used respectively to increase the number of feature channels from 48 to 96, 192, and finally to 288. Enc2 and Enc3 are configured with two residual attention modules for feature enhancement, and Enc4 is configured with one residual attention module for feature enhancement. An average pooling layer with a stride of 2 is used between each encoder to achieve downsampling operation and complete scale compression. S2.
3. The residual attention module first extracts local features through convolutional layers and grouping normalization operations, then combines channel attention and spatial attention mechanisms to achieve adaptive enhancement of feature response, and finally fuses with input features through residual connections; S2.
4. Features are weighted by channel attention and spatial attention. Channel attention generates channel weights based on multilayer perceptron (MLP), and spatial attention generates spatial weights based on channel statistical features and through convolution and sigmoid function, so as to enhance key channels and important temporal positions. S2.
5. By introducing a linear attention mechanism into the bottleneck layer, it is possible to efficiently capture long-term temporal context information; S2.
6. Use a multi-scale feature enhancement module to perform multi-scale feature fusion on bottleneck features and Enc3, Enc2, and Enc1 skip connection features. Capture information such as first arrival time, periodic structure, and tail wave attenuation through a multi-branch structure, and combine convolution, grouping normalization, ReLU function activation, and residual connection to achieve feature enhancement. S2.
7. Finally, a multi-level decoder is used to recover the signal step by step. Each level of the decoder achieves feature recovery through convolutional upsampling and uses linear interpolation to complete the time dimension alignment. Then, the upsampled features are concatenated with the skip connection features enhanced by MSFE. Feature extraction is completed through convolution, group normalization, ReLU function activation and residual attention module. In the final decoder layer Dec1, the high-fidelity shallow features extracted by Enc0 are used to replace the Enc1 features in order to preserve the fine structural information of the initial time step. S2.
8. The output features at the end of the decoder are first integrated through a convolutional layer, and then processed by grouping normalization and ReLU activation in sequence; S2.
9. Construct a microseismic signal dataset. After preprocessing the original microseismic signals, generate microseismic signals through the convolution method, and introduce real noise and synthetic noise to simulate real microseismic signals in order to construct sample data under different signal-to-noise ratio conditions. Divide the dataset into training set, validation set and test set according to a certain ratio to improve the generalization ability of the network model. S3. The constructed residual learning network is used to perform forward propagation calculation on the training samples. The difference between the network output and the target signal is evaluated by joint constraints. The backpropagation algorithm is used to iteratively update and optimize the network parameters until the model converges. Finally, the optimal denoising model with high-fidelity waveform reconstruction and first arrival preservation is obtained. S4. Input the noisy micro-vibration signal into the trained optimal denoising model for denoising processing, and output the corresponding high-fidelity waveform and first-arrival preserved signal.