Self-adaptive microseismic signal noise reduction method
By constructing the CDU-Net model, integrating the U-Net and Transformer modules, and combining characteristic linear modulation and hybrid loss functions, the adaptive noise reduction problem of existing microseismic signal processing methods in complex geological environments is solved, achieving high signal fidelity and noise suppression under different signal-to-noise ratio environments.
Patent Information
- Application Number
- CN202511043462.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-14
AI Technical Summary
Existing microseismic signal processing methods lack adaptive capabilities and struggle to effectively suppress multi-source non-stationary noise in complex geological environments. In particular, their denoising performance degrades significantly in low signal-to-noise ratio environments, affecting the timeliness and effectiveness of disaster early warning.
The CDU-Net model is constructed, integrating the U-Net basic architecture, Transformer module, and feature linear modulation module. It combines a hybrid loss function of multi-scale short-time Fourier transform loss and time-domain mean square error, and enhances the model's adaptive ability through self-attention mechanism and FiLM module. Paired datasets are designed and optimized for training.
It achieves adaptive suppression of non-stationary noise in complex geological environments, improves the identification accuracy and noise reduction performance of microseismic signals, enhances the model's adaptability under different signal-to-noise ratio environments, and ensures high signal fidelity.
Smart Images
Figure CN120950834A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of microseismic signal processing technology, and in particular to an adaptive microseismic signal noise reduction method. Background Technology
[0002] Microseismic monitoring technology, as a high-precision dynamic monitoring method, plays a crucial role in early warning of rockburst disasters in underground engineering, monitoring of rockburst in mine safety, and monitoring of fracture propagation and formation response in oil and gas field development. By capturing and analyzing minute vibration signals of underground rock masses under stress, this technology can detect potential disaster risks in advance, providing strong protection for engineering safety. With the continuous development of deep underground engineering and mineral resources, microseismic monitoring technology faces increasingly complex and variable geological environments, which places higher demands on the accurate identification of signals and the effective suppression of noise.
[0003] Traditional microseismic signal processing methods, such as wavelet transform and empirical mode decomposition, have achieved certain results in specific scenarios, but their denoising performance and recognition accuracy remain insufficient when dealing with non-stationary signals and time-varying noise in complex geological environments. In recent years, the rapid development of deep learning technology has provided a new solution for microseismic signal denoising. By automatically extracting signal features and possessing powerful nonlinear fitting capabilities, deep learning models have shown great potential in microseismic signal denoising under complex noise environments.
[0004] Although deep learning technology has made significant progress in the field of microseismic signal denoising, existing methods employ fixed-pattern denoising strategies and lack the ability to adapt to environmental changes and signal characteristics. Under complex geological conditions, microseismic signals are often affected by the superposition of multi-source non-stationary noise. The spectral characteristics of these noises may overlap with the effective signal, making it difficult for traditional methods and fixed-pattern deep learning models to accurately distinguish and effectively suppress noise. Especially in low signal-to-noise ratio environments, the denoising performance of existing models deteriorates significantly, making it difficult to maintain stable accuracy, thus affecting the timeliness and effectiveness of disaster early warning. Summary of the Invention
[0005] The purpose of this invention is to provide an adaptive microseismic signal denoising method, which aims to solve the technical problem that existing microseismic signal processing methods generally adopt fixed-mode denoising strategies and lack the ability to adapt to environmental changes and signal characteristics.
[0006] To achieve the above objectives, the present invention employs an adaptive microseismic signal noise reduction method, comprising the following steps:
[0007] A conditional denoising model for adaptive noise reduction of microseismic signals is constructed, integrating the U-Net basic architecture, Transformer module, and characteristic linear modulation module conditional modulation mechanism;
[0008] Construct a pairing dataset and preprocess the dataset;
[0009] The CDU-Net model was trained and evaluated to verify its denoising performance and optimize it. Visual analysis was used to provide interpretable evidence for the rationality of the network architecture, and ablation experiments were conducted to analyze the contribution of each component within the model.
[0010] In the step of constructing a conditional denoising model for adaptive noise reduction of microseismic signals, integrating the U-Net infrastructure, Transformer module, and characteristic linear modulation module's conditional modulation mechanism:
[0011] U-Net was chosen as the basic architecture of the model. The encoder-decoder structure and skip connection mechanism were used to preserve the microseismic signal characteristics, and U-Net was improved.
[0012] By embedding a Transformer module in the bottleneck layer of U-Net, a long-range dependency of microseismic signals is established through a self-attention mechanism, capturing global contextual information in the signals and compensating for the limitations of the local receptive field of convolutional neural networks.
[0013] The FiLM module is introduced into the encoder and decoder to dynamically generate modulation parameters based on noise levels or feature differences, and to adaptively adjust the features to enhance the model's adaptability to different signal-to-noise ratio environments.
[0014] We design a hybrid loss function that combines multi-scale short-time Fourier transform loss with time-domain mean square error to ensure that the denoised signal maintains high fidelity with the target signal in both the time and frequency domains, while also constraining frequency domain consistency and time domain energy.
[0015] In the steps of constructing the paired dataset and preprocessing the dataset:
[0016] Collect pure microseismic signals and on-site measured noise data from actual engineering monitoring environments;
[0017] The collected clean microseismic signals are randomly combined with the field measured noise data to generate a large number of noise-denoised sample pairs. Data augmentation techniques are then used to increase data diversity and improve the model's generalization ability.
[0018] The dataset is divided into training and validation sets using a stratified random sampling method to ensure that the distribution ratio of various noise types in the subsets is consistent with that in the overall dataset.
[0019] One hundred representative sample pairs were randomly selected from the dataset as the test set to objectively evaluate the denoising performance of the model.
[0020] In the steps of constructing the paired dataset and preprocessing the dataset:
[0021] Model training is performed within the PyTorch deep learning framework, using the Adam optimizer. Appropriate initial learning rate and batch size are set, and the learning rate is dynamically adjusted based on the loss changes during training to accelerate model convergence.
[0022] A multi-index quantitative evaluation was conducted, using mean squared error, signal-to-noise ratio, peak signal-to-noise ratio, and structural similarity index as comprehensive evaluation indicators to systematically compare the performance of CDU-Net with other denoising models.
[0023] By comparing waveforms, the noise suppression capability and signal fidelity of each algorithm are intuitively evaluated. Waveforms of the signal before and after denoising are plotted to observe the preservation of signal features and the noise suppression effect, and to verify the stability and generalization performance of the CDU-Net model in complex noise environments.
[0024] By visually analyzing the Grad-CAM activation status and waveform spectra before and after denoising of three key sub-modules in CDU-Net on representative low signal-to-noise ratio samples, interpretable evidence is provided for the rationality of the network architecture. The key sub-modules are the encoder, bottleneck and decoder.
[0025] By constructing a pruning variant model, the independent contribution of each sub-module to the final performance of CDU-Net is quantified, and the necessity and effectiveness of each module are verified from two dimensions: the convergence characteristics of the loss function and the improvement effect of the output signal-to-noise ratio.
[0026] Based on the evaluation results, visualization analysis, and ablation experiment results, the model was optimized as necessary, and its denoising performance and generalization ability were improved through continuous iteration.
[0027] In the step of dividing the dataset into training and validation sets using stratified random sampling to ensure that the distribution ratio of various noise types in the subsets remains consistent with that in the overall dataset:
[0028] The ratio of training set to validation set in the dataset is 7:3.
[0029] Among the steps involved in conducting multi-index quantitative evaluation, using mean squared error, signal-to-noise ratio, peak signal-to-noise ratio, and structural similarity index as comprehensive evaluation indicators, and systematically comparing the performance of CDU-Net with other denoising models:
[0030] Other denoising models include wavelet transform, variational mode decomposition, convolutional neural networks, and U-Net.
[0031] Among these steps, the model is optimized as necessary based on evaluation results, visualization analysis, and ablation experiment results, and its denoising performance and generalization ability are improved through continuous iteration:
[0032] Model optimization includes adjusting the network structure, optimizing hyperparameters, and introducing new technical modules.
[0033] This invention discloses an adaptive microseismic signal denoising method. It constructs a conditional denoising model for adaptively denoising microseismic signals, integrating the U-Net infrastructure, Transformer module, and characteristic linear modulation module's conditional modulation mechanism. A paired dataset is constructed and preprocessed. The CDU-Net model is then trained and evaluated to verify its denoising performance and optimize it. Ablation experiments are used to analyze the contribution of each component within the model. Through this method, a microseismic signal denoising model capable of adapting to complex geological environment changes and effectively suppressing non-stationary noise is achieved, solving the problem of adaptive denoising of microseismic signals in complex geological environments. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] Figure 1 This is a flowchart of the adaptive microseismic signal noise reduction method of the present invention.
[0036] Figure 2 This is a flowchart of steps S100 of the present invention.
[0037] Figure 3 This is a flowchart of steps S200 of the present invention.
[0038] Figure 4 This is a flowchart of steps S300 of the present invention.
[0039] Figure 5 This is a schematic diagram of the CDU-Net model architecture of the present invention.
[0040] Figure 6 This invention proposes a low signal-to-noise ratio Grad-CAM and spectral visualization method using CDU-Net processing. Detailed Implementation
[0041] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.
[0042] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0043] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0044] Please see Figures 1-5 This invention provides an adaptive microseismic signal noise reduction method, comprising the following steps:
[0045] S100: Construct a conditional denoising model for adaptive noise reduction of micro-vibration signals, integrating the U-Net infrastructure, Transformer module, and characteristic linear modulation module's conditional modulation mechanism.
[0046] In this embodiment, a CDU-Net model capable of adaptively reducing the noise of microseismic signals is designed and constructed, integrating the U-Net basic architecture, Transformer module, and characteristic linear modulation module conditional modulation mechanism. The specific process is as follows:
[0047] S101: U-Net was chosen as the basic architecture of the model. The encoder-decoder structure and skip connection mechanism were used to preserve the microseismic signal characteristics, and U-Net was improved.
[0048] S102: Embed the Transformer module in the bottleneck layer of U-Net, establish the long-range dependency of microseismic signals through the self-attention mechanism, capture the global context information in the signals, and make up for the local receptive field limitation of the convolutional neural network.
[0049] S103: Introducing the FiLM module in the encoder and decoder, which dynamically generates modulation parameters based on noise level or feature differences, performs channel-level adaptive adjustment of features, and enhances the model's adaptability to different signal-to-noise ratio environments.
[0050] S104: Design a hybrid loss function that combines multi-scale short-time Fourier transform loss with time-domain mean square error to ensure that the denoised signal maintains high fidelity with the target signal in both the time and frequency domains, while constraining frequency domain consistency and time domain energy.
[0051] In the above process, U-Net was chosen as the basic architecture of the model. The encoder-decoder structure and skip connection mechanism are used to preserve the microseismic signal features, avoid information loss during downsampling, and alleviate the gradient vanishing problem of deep networks. At the same time, U-Net is improved to adapt to the microseismic signal denoising task. A Transformer module is embedded in the bottleneck layer of U-Net to establish the long-range dependency of the microseismic signal through the self-attention mechanism, capture the global context information in the signal, and make up for the local receptive field limitation of the convolutional neural network. By introducing the FiLM module in the encoder and decoder, the modulation parameters are dynamically generated according to the noise level or feature difference, and the features are adaptively adjusted at the channel level to enhance the model's adaptability to different signal-to-noise ratio environments. By designing a hybrid loss function of multi-scale short-time Fourier transform loss and time-domain mean square error, it is ensured that the denoised signal maintains high fidelity with the target signal in both the time and frequency domains, while constraining frequency domain consistency and time domain energy.
[0052] In addition, the formulas for calculating the hybrid loss function are shown in equations (1) to (6).
[0053] L total ==L time +αL MSSTFT (1)
[0054]
[0055] in, Different scales of group S Calculate the value of the short-time Fourier transform; ω s The weights for each scale are typically set to 1; λ mag ∈(0,1) adjusts the relative weights of the spectral convergence and amplitude terms; x b For a clean (target) signal, The result is the model denoising, where B is the batch size and T is the length of each signal.
[0056] S200: Construct a paired dataset and preprocess the dataset.
[0057] In this embodiment, a paired dataset containing clean microseismic signals and field-measured noise data is constructed for training and validating the CDU-Net model, ensuring the environmental realism and engineering applicability of the data. The specific process is as follows:
[0058] S201: Collect pure microseismic signals and on-site measured noise data from actual engineering monitoring environments;
[0059] S202: The collected clean microseismic signals are randomly combined with the field measured noise data to generate a large number of noise-denoised sample pairs, and data augmentation techniques are used to increase data diversity and improve the generalization ability of the model.
[0060] S203: The dataset is divided into training and validation sets using a stratified random sampling method to ensure that the distribution ratio of various noise types in the subsets is consistent with that in the overall dataset.
[0061] S204: Randomly select 100 representative sample pairs from the dataset as a test set to objectively evaluate the denoising performance of the model.
[0062] In the above process, pure microseismic signals and field-measured noise data are collected from the actual engineering monitoring environment to ensure the authenticity and representativeness of the data. The collected pure microseismic signals and field-measured noise data are randomly combined to generate a large number of noise-denoised sample pairs. Data augmentation techniques (such as time stretching and pitch transformation) are used to increase data diversity and improve the generalization ability of the model. Then, a stratified random sampling method is used to divide the dataset into training set and validation set to ensure that the distribution ratio of various noises in the subset is consistent with the overall distribution. Finally, 100 representative sample pairs are randomly selected from the dataset as the test set to objectively evaluate the denoising performance of the model.
[0063] S300: The CDU-Net model is trained and evaluated to verify its denoising performance and optimize it. The contribution of each component within the model is analyzed through ablation experiments.
[0064] In this implementation, the denoising performance of the CDU-Net model is verified and necessary optimizations are performed through a systematic training and evaluation process. Furthermore, interpretable evidence of the network architecture's rationality is provided through visualization analysis. Additionally, the contribution of each component within the model is analyzed through ablation experiments. The specific process is as follows:
[0065] S301: Model training is performed in the PyTorch deep learning framework, using the Adam optimizer, setting appropriate initial learning rate and batch size, and dynamically adjusting the learning rate according to the loss changes during training to accelerate model convergence;
[0066] S302: Perform multi-index quantitative evaluation, using mean square error (MSE), signal-to-noise ratio (SNR), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) as comprehensive evaluation indicators to systematically compare the performance of CDU-Net with other denoising models;
[0067] S303: By comparing waveforms, the noise suppression capability and signal fidelity of each algorithm are intuitively evaluated. The waveforms of the signals before and after denoising are plotted to observe the preservation of signal features and the noise suppression effect, and to verify the stability and generalization performance of the CDU-Net model in complex noise environments.
[0068] S304: By visually analyzing the Grad-CAM activation status and waveform spectra before and after denoising of three key sub-modules in CDU-Net on representative low signal-to-noise ratio samples, interpretable evidence is provided for the rationality of the network architecture. The key sub-modules are the encoder, bottleneck and decoder.
[0069] S305: By constructing a pruning variant model, the independent contribution of each sub-module to the final performance of CDU-Net is quantified, and the necessity and effectiveness of each module are verified from two dimensions: the convergence characteristics of the loss function and the improvement effect of the output signal-to-noise ratio.
[0070] S306: Based on the evaluation results, visualization analysis, and ablation experiment results, the model is optimized as necessary, and its noise reduction performance and generalization ability are improved through continuous iteration.
[0071] In the above process, model training is performed within the PyTorch deep learning framework using the Adam optimizer. Appropriate initial learning rates and batch sizes are set, and the learning rate is dynamically adjusted based on loss changes during training to accelerate model convergence. Then, a multi-metric quantitative evaluation is conducted, employing Mean Square Error (MSE), Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM) as comprehensive evaluation metrics. CDU-Net is systematically compared with other denoising models (such as Wavelet Transform (WT), Variational Mode Decomposition (VMD), and Convolutional Neural Networks). The performance of CDU-Net (CNN) and U-Net was evaluated intuitively through waveform comparison, showing the noise suppression capability and signal fidelity of each algorithm. Waveforms of the signals before and after denoising were plotted to observe the preservation of signal features and the noise suppression effect, verifying the stability and generalization performance of the CDU-Net model in complex noise environments. Then, the activation status of Grad-CAM of the three key sub-modules (encoder, bottleneck, and decoder) in CDU-Net on representative low signal-to-noise ratio samples and the waveform spectra before and after denoising were visualized and analyzed, providing interpretable evidence for the rationality of the network architecture. Furthermore, by constructing variant models with reductions (such as the basic model retaining only the U-Net framework and the U-Net-Transformer model that removes the FiLM module but retains the Transformer), the independent contribution of each sub-module (U-Net basic architecture, Transformer module, and FiLM module) to the final performance of CDU-Net is quantified. The necessity and effectiveness of each module are verified from two dimensions: the convergence characteristics of the loss function and the improvement effect of the output signal-to-noise ratio. Finally, based on the evaluation results, visualization analysis, and ablation experiment results, the model is optimized as necessary, such as adjusting the network structure, optimizing hyperparameters, and introducing new technical modules. Through continuous iteration, the denoising performance and generalization ability of the model are improved, making it better suited to real-world application scenarios.
[0072] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0073] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. An adaptive microseismic signal noise reduction method, characterized in that, Includes the following steps: A conditional denoising model for adaptive noise reduction of microseismic signals is constructed, integrating the U-Net basic architecture, Transformer module, and characteristic linear modulation module conditional modulation mechanism; Construct a pairing dataset and preprocess the dataset; The CDU-Net model was trained and evaluated to verify its denoising performance and optimize it. Visual analysis was used to provide interpretable evidence for the rationality of the network architecture, and ablation experiments were conducted to analyze the contribution of each component within the model.
2. The adaptive microseismic signal noise reduction method as described in claim 1, characterized in that, In the steps of constructing a conditional denoising model for adaptive noise reduction of microseismic signals, integrating the U-Net infrastructure, Transformer module, and characteristic linear modulation module's conditional modulation mechanism: U-Net was chosen as the basic architecture of the model. The encoder-decoder structure and skip connection mechanism were used to preserve the microseismic signal characteristics, and U-Net was improved. By embedding a Transformer module in the bottleneck layer of U-Net, a long-range dependency of microseismic signals is established through a self-attention mechanism, capturing global contextual information in the signals and compensating for the limitations of the local receptive field of convolutional neural networks. The FiLM module is introduced into the encoder and decoder to dynamically generate modulation parameters based on noise levels or feature differences, and to perform channel-level adaptive adjustment of features, thereby enhancing the model's adaptability to environments with different signal-to-noise ratios. We design a hybrid loss function that combines multi-scale short-time Fourier transform loss with time-domain mean square error to ensure that the denoised signal maintains high fidelity with the target signal in both the time and frequency domains, while also constraining frequency domain consistency and time domain energy.
3. The adaptive microseismic signal noise reduction method as described in claim 2, characterized in that, In the steps of constructing the paired dataset and preprocessing the dataset: Collect pure microseismic signals and on-site measured noise data from actual engineering monitoring environments; The collected clean microseismic signals are randomly combined with the field measured noise data to generate a large number of noise-denoised sample pairs. Data augmentation techniques are then used to increase data diversity and improve the model's generalization ability. The dataset is divided into training and validation sets using a stratified random sampling method to ensure that the distribution ratio of various noise types in the subsets is consistent with that in the overall dataset. One hundred representative sample pairs were randomly selected from the dataset as the test set to objectively evaluate the denoising performance of the model.
4. The adaptive microseismic signal noise reduction method as described in claim 1, characterized in that, In the steps of constructing the paired dataset and preprocessing the dataset: Model training is performed within the PyTorch deep learning framework, using the Adam optimizer. Appropriate initial learning rate and batch size are set, and the learning rate is dynamically adjusted based on the loss changes during training to accelerate model convergence. A multi-index quantitative evaluation was conducted, using mean squared error, signal-to-noise ratio, peak signal-to-noise ratio, and structural similarity index as comprehensive evaluation indicators to systematically compare the performance of CDU-Net with other denoising models. By comparing waveforms, the noise suppression capability and signal fidelity of each algorithm are intuitively evaluated. Waveforms of the signal before and after denoising are plotted to observe the preservation of signal features and the noise suppression effect, and to verify the stability and generalization performance of the CDU-Net model in complex noise environments. By visually analyzing the Grad-CAM activation status and waveform spectra before and after denoising of three key sub-modules in CDU-Net on representative low signal-to-noise ratio samples, interpretable evidence is provided for the rationality of the network architecture. The key sub-modules are the encoder, bottleneck and decoder. By constructing a pruning variant model, the independent contribution of each sub-module to the final performance of CDU-Net is quantified, and the necessity and effectiveness of each module are verified from two dimensions: the convergence characteristics of the loss function and the improvement effect of the output signal-to-noise ratio. Based on the evaluation results, visualization analysis, and ablation experiment results, the model was optimized as necessary, and its denoising performance and generalization ability were improved through continuous iteration.
5. The adaptive microseismic signal noise reduction method as described in claim 3, characterized in that, In the step of dividing the dataset into training and validation sets using stratified random sampling to ensure that the distribution ratio of various noise types in the subsets remains consistent with that in the overall dataset: The ratio of training set to validation set in the dataset is 7:
3.
6. The adaptive microseismic signal noise reduction method as described in claim 4, characterized in that, In the step of conducting multi-index quantitative evaluation, using mean squared error, signal-to-noise ratio, peak signal-to-noise ratio, and structural similarity index as comprehensive evaluation indicators, and systematically comparing the performance of CDU-Net with other denoising models: Other denoising models include wavelet transform, variational mode decomposition, convolutional neural networks, and U-Net.
7. The adaptive microseismic signal noise reduction method as described in claim 6, characterized in that, Based on the evaluation results, visualization analysis, and ablation experiment results, the model is optimized as necessary, and its denoising performance and generalization ability are improved through continuous iteration: Model optimization includes adjusting the network structure, optimizing hyperparameters, and introducing new technical modules.