A mine low signal-to-noise ratio microseismic signal denoising method

Abstract

The application discloses a mine low signal-to-noise ratio microseismic signal denoising method, and relates to the technical field of signal data processing. First, mine original microseismic signals are collected, and Gaussian noise is synthesized to jointly construct a mixed signal sample set with labels. Secondly, a U-BiLSTM-CA-DDNet model is constructed, a joint loss function is constructed, and the model is iteratively trained. Finally, the microseismic signal to be processed is input into the model to obtain a predicted clean signal and a predicted noise signal. The U-BiLSTM-CA-DDNet model is based on the improvement of the U-Net. On the basis of multi-scale feature extraction and time sequence dependence modeling, the channel attention mechanism and the double decoder architecture are introduced. Through multi-layer feature fusion, adaptive frequency band regulation and explicit signal separation mechanism, explicit separation and joint optimization of signals and noise under low signal-to-noise ratio conditions are realized, so that the denoising strength and waveform fidelity are simultaneously ensured under the complex non-stationary noise environment.

Description

Technical Field

[0001] This invention relates to the field of signal data processing technology, specifically to a method for denoising low signal-to-noise ratio microseismic signals in mines. This method is used to intelligently denoise and preserve features of microseismic signals with strong noise interference in complex mining environments, significantly improving the reliability of microseismic monitoring and the accuracy of seismic source location. Background Technology

[0002] Microseismic monitoring technology is an important tool for assessing the stability of rock masses and providing early warning of disasters in deep mines. By collecting and analyzing microseismic signals caused by rock mass fracturing, the evolution of rock mass damage and stress concentration characteristics can be revealed, thereby enabling early warning of disasters such as rock bursts and roof falls. However, in actual mining environments, microseismic signals are often severely interfered with by multi-source noise, including mechanical vibration, blasting vibration, electromagnetic noise, and fluid noise. These interference signals significantly reduce the signal-to-noise ratio (SNR), mask key seismic phase characteristics such as P-waves and S-waves, and lead to a decrease in the accuracy of seismic source location.

[0003] Traditional noise reduction methods, such as bandpass filtering, Gaussian filtering, wavelet thresholding, and empirical mode decomposition (EMD), are generally based on the assumption of noise stationarity, suppressing noise through fixed thresholds or frequency band segmentation. For example, Chinese invention patent application number 202311437967.1 discloses a microseismic signal denoising method based on improved wavelet denoising combined with EMD. This method decomposes the signal into multiple IMF components, then uses a correlation weighting method to exclude weakly correlated IMF components, and after denoising the remaining IMF components, reconstructs the signal to obtain the denoised signal. However, mine noise typically exhibits non-stationary and co-band characteristics. When the noise spectrum overlaps with the effective signal frequency band, traditional methods easily lead to waveform distortion or loss of dominant frequency energy.

[0004] In recent years, deep learning models have been widely introduced into the field of earthquake and microseismic signal processing. Models such as U-Net, convolutional autoencoders, and variational modal networks can improve denoising performance to a certain extent. For example, Chinese invention patent application number 202510335761.0 discloses a microseismic signal denoising and reconstruction method and system based on multi-scale decomposition, which constructs a multi-scale feature matrix and a U-Net microseismic signal denoising model to effectively identify and remove noise, thereby achieving denoising of microseismic monitoring data. However, existing models still have the following key problems:

[0005] (1) Co-band interference in the spectrum is not effectively distinguished: Single-channel networks have difficulty accurately separating effective signals from non-stationary noise in the feature space;

[0006] (2) Insufficient coupling between short-term mutations and long-term decay features: Convolutional neural networks are good at capturing local structures, but ignore temporal dependencies; Recurrent neural networks can model global trends, but are prone to losing local burst features;

[0007] (3) Significant energy distortion and phase shift problems: Multi-layer downsampling and fixed threshold lead to signal energy attenuation and waveform shift, affecting the reconstruction of phase characteristics.

[0008] Therefore, there is an urgent need for a deep neural network architecture that combines spatial feature extraction, temporal dependency modeling, and frequency band adaptive adjustment to achieve high-fidelity noise reduction of mine microseismic signals under low signal-to-noise ratio conditions. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies in denoising low signal-to-noise ratio (SNR) microseismic signals in mines, this invention provides a method for denoising low SNR microseismic signals in mines. Based on multi-scale feature extraction and temporal dependency modeling, this method introduces a channel attention mechanism and a dual-decoder architecture to construct a hybrid neural network architecture of "U-Net + BiLSTM + channel attention + dual decoder," abbreviated as U-BiLSTM-CA-DDNet model. Through multi-layer feature fusion, adaptive frequency band adjustment, and explicit signal separation mechanisms, it achieves explicit separation and joint optimization of signal and noise under low SNR conditions, thereby ensuring both denoising strength and waveform fidelity in complex non-stationary noise environments.

[0010] To achieve the above objectives, this application provides the following solution:

[0011] A method for denoising low signal-to-noise ratio microseismic signals in mines includes the following steps:

[0012] S1 collects raw microseismic signals from the mine and synthesizes Gaussian noise to construct a labeled mixed signal sample set.

[0013] S2, construct the U-BiLSTM-CA-DDNet model.

[0014] S3. Construct a joint loss function and iterate the U-BiLSTM-CA-DDNet model multiple times until the error meets the preset requirements.

[0015] S4. The microseismic signal to be processed is truncated and standardized by a fixed time window signal and then input into the U-BiLSTM-CA-DDNet model to obtain the predicted clean signal and the predicted noise signal.

[0016] According to some embodiments of this application, a method for denoising low signal-to-noise ratio microseismic signals in mines is provided. Specifically, S1 involves: acquiring raw microseismic signals from the mine; truncating each signal to a uniform length using a fixed time window; obtaining real noise signals and clean signals through time-frequency analysis and manual verification; simultaneously synthesizing Gaussian noise, which is used together with the real noise signals as noise signals; linearly superimposing the noise signals and clean signals according to a set signal-to-noise ratio to obtain a mixed signal; and performing Z-score normalization on the mixed signal to obtain a labeled mixed signal sample set.

[0017] According to some embodiments of this application, a method for denoising low signal-to-noise ratio microseismic signals in mines is provided. In step S2, the U-BiLSTM-CA-DDNet model is based on an improvement to the U-Net structure, including: embedding a bidirectional long short-term memory network between the encoder module and the decoder module to capture temporal dependence features and energy decay trends; setting the decoder module as two parallel branches, respectively used to predict clean signals and noisy signals; and introducing a channel attention mechanism in the noise branch decoder module to achieve adaptive adjustment of channel feature weights.

[0018] Furthermore, the U-BiLSTM-CA-DDNet model includes an encoder module, a temporal modeling module, a bridging layer, a channel attention module, and a dual decoder module.

[0019] The encoder module consists of multiple encoders connected in series. Each encoder layer contains two convolutional layers and one max pooling layer. After the mixed signal samples are input into the encoder, multi-scale feature extraction is achieved by increasing the number of channels and reducing the temporal dimension layer by layer.

[0020] The temporal modeling module consists of two layers of bidirectional long short-term memory network and one layer of one-dimensional convolutional layer; the encoder module output is processed by the bidirectional long short-term memory network to extract temporal dependent features and energy decay trends, and then processed by the convolutional layer to achieve feature mapping and dimensionality transformation.

[0021] The bridging layer is used to establish a fusion connection between the temporal modeling features and the encoder features. It includes a feature concatenation layer between two convolutional layers. It is used to concatenate the features output by the temporal modeling module after convolutional mapping with the features of the last layer of the encoder module and then perform convolutional fusion to provide high-dimensional feature input for the decoder.

[0022] The dual decoder module includes a clean signal branch and a noise signal branch set in parallel. Each branch includes multiple cascaded decoders, which are symmetrical to the encoder module. Each decoder layer contains an upsampling layer, a splicing layer, a convolutional layer, and an output layer. Features from the corresponding encoder layer are fused with the decoder layer through skip connections. By combining upsampling, cropping, and padding operations, the signal resolution and length are gradually restored. The clean signal branch outputs the predicted clean signal, and the noise signal branch outputs the predicted noise signal.

[0023] The channel attention module is embedded after the convolutional layers of each decoder layer in the noise signal branch. It generates channel weights through global average pooling and a two-layer fully connected network, enabling adaptive adjustment of the importance of feature channels and thus enhancing the noise feature separation capability. Specifically:

[0024] The input to the channel attention module is the output feature F of the convolutional layer. First, global average pooling is used to compress the input feature F in the time dimension to obtain the channel statistics:

[0025] (1);

[0026] Then, channel statistics Dimensionality reduction and dimensionality expansion are performed sequentially through two layers of fully connected networks:

[0027] (2);

[0028] Where W is the generated channel weight vector, and These are the learnable weight matrices for two fully connected network layers, respectively. This represents the Sigmoid activation function. This represents the modified linear unit function.

[0029] Finally, the generated channel weight vector W is fused with the input features F through channel-by-channel multiplication:

[0030] (3);

[0031] Here, ⊙ represents element-wise multiplication. This is the output of the channel attention module.

[0032] According to some embodiments of this application, a method for denoising low signal-to-noise ratio microseismic signals in mines is provided. Specifically, S3 involves: using a multi-output joint loss function to jointly optimize the prediction errors of clean and noisy signals in the U-BiLSTM-CA-DDNet model; training the U-BiLSTM-CA-DDNet model using a labeled mixed signal sample set; updating parameters through multiple iterations during model training until the overall error meets the preset convergence condition.

[0033] Furthermore, the multi-output joint loss function is as follows:

[0034] (4);

[0035] (5);

[0036] (6);

[0037] Where: t represents the time sampling point index, and N represents the total number of sampling points for a single signal sample. Let c(t) represent the clean signal predicted by the model, and c(t) represent the true clean signal. Let n(t) represent the noise signal predicted by the model, and n(t) represent the actual noise signal. This is the total loss function, used as the optimization objective for model training; The clean signal loss measures the deviation between the predicted clean signal and the actual clean signal. The noise signal loss measures the deviation between the predicted noise and the actual noise; MSE(*,*) represents the mean square error; α and β are weight hyperparameters.

[0038] The optimizer uses the Adam optimizer with a learning rate of 1×10⁻⁶. -3 Momentum parameters β1=0.9, β2=0.999; batch size 32; training epochs 100; early stopping patience value 20; convolutional layers use L2 regularization and batch normalization (BatchNorm) to prevent overfitting; activation function is ReLU.

[0039] The beneficial effects of this invention are:

[0040] 1. This invention introduces an explicit dual-decoder separation mechanism into the network structure. By using a parallel signal decoder and a noise decoder to learn the features of the effective phase and noise components respectively, it achieves independent modeling and adversarial optimization of clean signals and noise in the feature space. This structure constrains the outputs of the two branches through a joint loss function, ensuring that the model suppresses noise without destroying the signal structure. This explicit separation strategy breaks the limitations of the traditional single-channel network's "implicit residual learning," giving the model stronger interpretability and robustness. Experiments demonstrate that this network structure can maintain stable signal reconstruction accuracy even in low SNR environments.

[0041] 2. Convolutional neural networks excel at capturing local waveform details of microseismic signals, but struggle to characterize the continuous evolution of seismic phases in the time domain; recurrent neural networks possess the ability to model long-term dependencies, but easily overlook short-term, sudden seismic phase characteristics. To address this complementary nature, this invention proposes a deep architecture that fuses U-Net and a bidirectional long short-term memory network. This "spatial-temporal joint representation" structure can achieve dynamic modeling of complex source signals while maintaining the synchronization of seismic phase energy and phase. Experimental results show that this fusion architecture can improve the signal correlation coefficient to 0.97 and reduce the root mean square error by approximately 40%, effectively improving temporal consistency and seismic coherence under low signal-to-noise ratio conditions.

[0042] 3. Traditional multi-layer downsampling structures are prone to causing phase energy attenuation and phase drift, while fixed threshold filtering strategies cannot adaptively adjust according to noise type. Therefore, this invention embeds a channel attention adaptive control mechanism into the noise decoder branch. This mechanism dynamically adjusts the response intensity of each frequency band according to the spectral energy distribution of the input signal, thereby highlighting the dominant phase frequency band, suppressing non-stationary noise frequency bands, and achieving adaptive balance of frequency band energy. Compared to traditional fixed threshold filtering, the channel attention module effectively prevents energy loss and waveform distortion caused by over-filtering, maintaining the stability of phase and amplitude while ensuring noise reduction intensity, thus significantly improving waveform fidelity.

[0043] 4. To address the complex and diverse sources of noise in mining sites, this invention employs a strategy of constructing mixed signal samples using both Gaussian noise and real noise. This dual-source noise strategy enables the model to simultaneously learn the structural characteristics of real noise and the statistical characteristics of Gaussian noise during the training phase, thereby improving the adaptability and generalization ability of deep learning models to multi-noise environments.

[0044] In summary, this invention effectively overcomes the problems of spectral co-band interference, insufficient feature coupling, and energy distortion in traditional mine signal noise reduction methods. It still achieves high-precision phase reconstruction and waveform fidelity under low signal-to-noise ratio (SNR<10dB) conditions, and has significant engineering practical value and promotion potential. Attached Figure Description

[0045] Figure 1 This is a diagram of the U-BiLSTM-CA-DDNet model structure.

[0046] Figure 2 The training and validation loss curves are shown, where (a) is the total loss curve, (b) is the loss curve for predicting clean signals, and (c) is the loss curve for predicting noisy signals.

[0047] Figure 3 The following are comparison diagrams of time-domain and frequency-domain analysis, where (a) is the original waveform of the noisy signal; (b) is a comparison diagram of time-domain reconstruction in this embodiment; (c) is a comparison diagram of time-domain reconstruction using U-Net; (d) is a comparison diagram of time-domain reconstruction using the bandpass filtering method; (e) is a comparison diagram of time-domain reconstruction using the empirical mode decomposition method; (f) is a comparison diagram of frequency-domain analysis in this embodiment; (g) is a comparison diagram of frequency-domain analysis using U-Net; (h) is a comparison diagram of frequency-domain analysis using the bandpass filtering method; and (i) is a comparison diagram of frequency-domain analysis using the empirical mode decomposition method.

[0048] Figure 4The following are generalization results of the U-BiLSTM-CA-DDNet model on the cross-mine dataset in Example 2: (a) is the time-domain waveform of the noisy signal; (b) is the time-spectrum of the noisy signal; (c) is the time-domain waveform of the clean signal predicted by the model; (d) is the time-spectrum of the clean signal predicted by the model; (e) is the time-domain waveform of the noisy signal predicted by the model; and (f) is the time-spectrum of the noisy signal predicted by the model. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] This invention proposes a method for denoising low signal-to-noise ratio microseismic signals in mines. Based on the channel attention mechanism of U-Net and bidirectional long short-term memory network, it achieves accurate denoising and waveform feature preservation of microseismic signals under complex noise environment by extracting multi-scale features of noisy signals, modeling time-dependent features, and adaptively adjusting channels.

[0051] Example 1

[0052] This embodiment uses microseismic data collected from the Xiadian Gold Mine in Shandong Province as an example to further illustrate the method of the present invention, and conducts relevant experimental comparisons and performance evaluations. A method for denoising low signal-to-noise ratio microseismic signals in mines includes:

[0053] Step 1: Collect raw microseismic signals from the mine and synthesize Gaussian noise to construct a labeled mixed signal sample set.

[0054] Raw microseismic signals were collected from the mine microseismic monitoring system at a sampling rate of 6000 Hz and truncated to a length of 4500 sampling points using a fixed time window. For any part of the length less than 4500 sampling points, zeros were used to fill the remaining part to 4500 sampling points.

[0055] Based on actual mine operations, time-frequency analysis was performed on the extracted raw microseismic signals, and manual verification was conducted to select 1930 sets of real noise, including mechanical vibration, electromagnetic interference, and background noise from ventilation equipment, as the real noise sample set. 2205 sets of high SNR rock mass fracturing microseismic events were selected as the clean signal sample set. Simultaneously, to supplement the dataset diversity, white noise was used as a base, superimposed with multi-frequency oscillations ranging from 0.05 to 20 Hz (large amplitude at low frequencies and attenuation at high frequencies), and 20 to 50 narrow pulses smoothed by a Hanning window were added to synthesize 275 sets of composite Gaussian noise closely resembling the noise characteristics of the mine site, serving as the Gaussian synthetic noise sample set. The real noise, clean signals, and Gaussian synthetic noise sample sets together constitute the complete database required for model training.

[0056] Clean signal samples and noisy signal samples are linearly superimposed according to a set signal-to-noise ratio (0–20dB) to obtain a mixed signal sample set. Specifically, in this embodiment, clean signal samples are matched with real noise samples and Gaussian noise samples respectively to construct noisy signals, generating mixed signal samples containing real noise signals and mixed signal samples containing Gaussian noise signals, which together constitute a labeled mixed signal sample set. This embodiment, by establishing a scenario covering real mine noise and synthetic Gaussian noise, can effectively improve the model's generalization ability to mixed noise environments.

[0057] To improve data consistency, Z-score standardization was performed on all samples to make the signal mean 0 and the standard deviation 1, and the samples were divided into training set, validation set and test set in a ratio of 8:1:1.

[0058] Step 2: Construct the U-BiLSTM-CA-DDNet model.

[0059] Based on the U-Net network, improvements are made by introducing a bidirectional long short-term memory network and a channel attention mechanism to construct the U-BiLSTM-CA-DDNet model. The improvements include: embedding a bidirectional long short-term memory network between the encoder and decoder modules to capture temporal dependency features and energy decay trends; setting the decoder module as two parallel branches, one for predicting clean signals and the other for predicting noisy signals; and introducing a channel attention mechanism in the noisy branch decoder module to achieve adaptive adjustment of channel feature weights. Specifically:

[0060] like Figure 1 As shown, the U-BiLSTM-CA-DDNet model consists of an encoder module, a temporal modeling module, a bridging layer, a dual decoder module, and a channel attention module. It is used for separate denoising and waveform reconstruction of low signal-to-noise ratio micro-seismic signals from mines. The model input is a one-dimensional mixed signal sample of length 4500 obtained in step 1. After feature extraction, temporal modeling, and dual-branch decoding, it outputs a clean prediction signal and a noise prediction signal, respectively.

[0061] The encoder module is a three-layer cascaded coding structure. Each encoder layer consists of two one-dimensional convolutional layers (Conv1D) and one-dimensional max-pooling layers (MaxPooling1D). The kernel size of each convolutional layer is 15, and the activation function is ReLU. The number of output channels is 32, 64, and 128 respectively. The encoder module extracts temporal features and seismic morphology at different scales through local convolution operations, and performs feature downsampling through max-pooling to enhance the network's ability to express local waveforms and multi-scale features. In addition, the output features of each encoder layer are fused with the decoder features of the symmetric layer in the subsequent decoding stage through the skip connection mechanism of the U-Net structure to preserve the original detailed information.

[0062] The temporal modeling module, located between the encoder and decoder, comprises two layers of bidirectional long short-term memory (BiLSTM) networks and one one-dimensional convolutional layer (kernel size 1, activation function ReLU). Each BiLSTM layer contains 128 hidden units. The BiLSTM reshapes the encoder module output into a time-series input, simultaneously capturing the forward and backward temporal dependencies of the signal using a bidirectional scanning mechanism, enabling dynamic modeling of P-wave initiation, energy decay, and phase propagation patterns. The BiLSTM output undergoes feature transformation via a one-dimensional convolutional layer to enhance the feature representation dimension and facilitate fusion with the encoder output in the bridging layer.

[0063] The bridging layer consists of two one-dimensional convolutional structures and a feature concatenation layer, located between the two one-dimensional convolutional structures. Both one-dimensional convolutional structures have a kernel size of 15 and 256 channels. After the output of the temporal modeling module is mapped by a 1×1 convolutional layer, its output features are concatenated with the final output of the encoder through the feature concatenation layer, achieving the fusion of local spatial features and global temporal features. The fused features are then enhanced with a one-dimensional convolutional operation to strengthen the high-level abstract representation, providing high-resolution feature input for the subsequent dual decoder.

[0064] The dual-decoder module includes parallel clean signal and noise signal branches. Each branch adopts a three-layer cascaded decoding structure symmetrical to the encoder module. Each decoder layer consists of a one-dimensional upsampling layer, a concatenation layer, two one-dimensional convolutional layers, and a linear output layer. The one-dimensional upsampling layer is used to progressively recover the feature length. The concatenation layer fuses the output features of the corresponding encoder layer with the features of the current decoder layer through skip connections. The convolutional layer is used for local feature reconstruction. The output of the clean signal branch generates "clean_output," i.e., the predicted clean signal, through a linear activation function, while the output of the noise branch serves as "noise_output," i.e., the predicted noise signal, thus achieving explicit separation between the clean and noise signals.

[0065] The Channel Attention (CA) module is embedded after the output of two one-dimensional convolutional layers in each decoder layer of the noise signal branch. It dynamically adjusts the importance of different channel features to achieve adaptive weighting of frequency band energy and noise enhancement. This module takes the current feature map as input, extracts channel statistics through Global Average Pooling 1D, and performs dimensionality reduction and enhancement operations through two fully connected layers. The first fully connected layer compresses the input channels using a reduction ratio of 8 and an activation function of ReLU. The second fully connected layer restores the original channel dimension, uses a Sigmoid activation function, and generates a channel weight vector. This weight is multiplied channel-by-channel by the input features (multiply operation) to form a weighted feature output, thereby strengthening the expression of the main noise frequency bands and suppressing redundant components. The introduction of the channel attention mechanism enables the U-BiLSTM-CA-DDNet model to adaptively adjust the noise reduction amplitude according to the spectral characteristics of the input signal, avoiding energy attenuation and waveform distortion caused by fixed threshold filtering, and achieving higher waveform fidelity and noise reduction accuracy. Specifically:

[0066] The input to the channel attention module is the output features of the convolutional layer. , where R is the real number field, T is the time step, and C is the number of channels. First, global average pooling (GAP) is used to compress the input feature F in the time dimension to obtain the channel statistics, as shown in equation (1).

[0067] Then, channel statistics Dimensionality reduction and expansion are performed sequentially through two fully connected network layers. The first layer reduces the dimensionality by 1 / 8, that is, the number of channels is compressed from C to C / 8, and the activation function is ReLU. The second layer restores the original channel dimension C, and the activation function is Sigmoid. The mathematical expression of this process is shown in Equation (2).

[0068] Finally, the generated channel weight vector W is fused with the input feature F through a channel-by-channel multiplication operation, as shown in Equation (3).

[0069] The resulting output feature 𝐹′ maintains the original phase structure while optimizing energy distribution and preserving signal spectrum fidelity through adaptive enhancement of dominant frequency band features and dynamic suppression of noise frequency bands.

[0070] The U-BiLSTM-CA-DDNet model described above organically combines spatial feature extraction, temporal dependency modeling, and dynamic channel control. Through multi-scale feature fusion of the U-Net structure, bidirectional temporal modeling of BiLSTM, and adaptive weighting mechanism of the CA module, the model can achieve high-precision phase recovery and noise separation in low signal-to-noise ratio mining environments, significantly improving noise reduction performance and signal restoration quality.

[0071] Step 3: Construct a joint loss function and iterate the U-BiLSTM-CA-DDNet model multiple times until the error meets the preset requirements.

[0072] The joint loss function shown in equations (4) to (6) is designed to jointly optimize the prediction error of clean and noisy signals. The U-BiLSTM-CA-DDNet model is trained using the training and validation sets of the labeled mixed signal sample set obtained in step 1. During the model training process, the parameters are updated through multiple iterations until the overall error meets the preset convergence condition. Specifically, in this embodiment, N is 4500 in equations (4) and (5), and the weight hyperparameter α is 1 and β is 3 in equation (6) to improve the training priority of the noise branch and enhance the separation accuracy of non-stationary noise.

[0073] The model training process uses the Adam optimizer with a learning rate of 1×10⁻⁶. -3 Momentum parameters β1=0.9, β2=0.999; training parameters: batch size 32, number of training epochs 100; early stopping strategy: terminate early if the validation set loss does not decrease after 15 epochs; adaptive learning rate decay factor: 0.5; regularization: L2 regularization and Batch Normalization; activation function: ReLU. The optimal model weights on the validation set are automatically saved during training.

[0074] Specifically, in this embodiment, the U-BiLSTM-CA-DDNet model converges rapidly after 40 rounds of training, with the validation set loss being essentially the same as the training set loss, showing no overfitting. Figure 2 As shown, the total loss, clean signal loss, and noisy signal loss curves for both the training and validation sets all exhibit the typical characteristic of "rapid descent – ​​smooth convergence." Figure 2 As shown in (a), the overall loss stabilizes after approximately 50 training epochs. The validation set loss is slightly higher than the training set loss but does not show significant divergence, indicating that the model maintains a good convergence trend during training. Figure 2 As shown in (b), the loss on both the training and validation sets of the clean signal branch decreases synchronously and converges at similar speeds, indicating that the model can effectively extract the main signal features through the synergistic effect of multi-scale convolution and temporal modeling. In contrast, the loss on the noisy signal branch decreases more rapidly (e.g., ...). Figure 2 As shown in (c), the final stable value is approximately (5 × 10). -5 The loss curve is on the order of magnitude lower than that of the clean signal branch, which aligns with the optimal design of a 1:3 loss weight ratio. This result reflects that the model can maintain a balance between signal reconstruction and noise separation under dual-task constraints, and the smooth convergence of the loss curve also indicates that the overall training process is stable and there are no signs of overfitting.

[0075] Step 4: After the microseismic signal to be processed is truncated and standardized by a fixed time window, it is input into the U-BiLSTM-CA-DDNet model to obtain the predicted clean signal and the predicted noise signal.

[0076] Performance evaluation and comparative experiments:

[0077] To evaluate the performance of this embodiment in denoising mine microseismic signals, traditional methods and existing deep learning models were selected as controls. Based on the test set of the mixed signal sample set obtained in step 1, comparative experiments were conducted with signal-to-noise ratio improvement (ΔSNR), signal correlation (R), and mean square error (MSE) as the main evaluation indicators. The traditional methods included bandpass filtering (fixed frequency threshold) and empirical mode decomposition (EMD, local feature extraction based on adaptive signal decomposition), while the existing deep learning model was U-Net. The comparative experimental results are shown in Table 1.

[0078] Table 1 shows the results of the comparative experiment.

[0079]

[0080] The results show that the U-BiLSTM-CA-DDNet model proposed in this embodiment exhibits excellent overall noise reduction performance. In terms of signal-to-noise ratio (SNR) improvement, the average SNR improvement of the U-BiLSTM-CA-DDNet model is 11.40 dB, significantly higher than bandpass filtering (3.77 dB), U-Net (8.21 dB), and empirical mode decomposition (4.63 dB). Compared with the other three methods, the SNR improvement of the method in this embodiment reaches 202.4%, 38.9%, and 146.2%, respectively, fully reflecting the stronger noise suppression capability of the U-BiLSTM-CA-DDNet model in low SNR environments. Traditional methods, relying on fixed frequency thresholds or global smoothing strategies, struggle to effectively address the time-varying noise characteristics in mining environments, thus exhibiting significant shortcomings in overall noise reduction. Regarding feature preservation, the correlation coefficient between the output signal of this embodiment and the real clean signal is 0.97, which is higher than that of bandpass filtering (0.90), U-Net (0.93), and Empirical Mode Decomposition (0.91), indicating that the model in this embodiment can better maintain the waveform consistency and temporal structure of the original signal while reducing noise. In contrast, the fixed frequency band limitation of bandpass filtering easily causes some phase energy to weaken, while U-Net and EMD methods have insufficient recovery accuracy in some frequency bands, resulting in a slight decrease in the phase and amplitude matching degree of the output signal. In terms of recovery error index, the mean square error (MSE) of the model in this embodiment is 1.21 × 10⁻⁶. -4 This is far lower than traditional methods—bandpass filtering (1.82×10⁻⁶). -3 ) and empirical mode decomposition (1.07×10 -3 Compared to U-Net, the MSE of this embodiment model is still reduced by about 84.5%, indicating that the model in this embodiment has smaller errors, clearer phase boundaries, and significantly improved overall reconstruction accuracy during signal reconstruction.

[0081] In summary, compared with methods such as bandpass filtering, U-Net, and EMD, the method in this embodiment achieves the best results in terms of signal-to-noise ratio improvement, signal correlation, and mean square error. These results fully verify the high robustness and waveform fidelity of the model in this embodiment for denoising low signal-to-noise ratio mine microseismic signals, demonstrating a significant improvement in overall denoising performance and strong engineering application value.

[0082] like Figure 3The temporal reconstruction results of different denoising methods in (a) to (e) show that the model in this embodiment maintains a high degree of consistency with the real clean signal in the overall waveform reconstruction, and can better preserve key temporal features such as P-wave initiation and S-wave attenuation. The waveform output by the model in this embodiment almost overlaps with the real signal in the main phase, with smooth energy changes and no obvious amplitude changes, indicating that it still has good temporal fidelity and structural stability under low signal-to-noise ratio conditions. In contrast, the other three denoising methods all have certain differences in waveform details. U-Net, due to its multi-scale convolutional structure, can better recover the main outline of the waveform, but under strong noise interference, its ability to capture long-term energy attenuation is limited, resulting in a slightly smooth transition in the P-wave initiation segment and a slight blurring of the waveform boundaries; in addition, due to the lack of explicit temporal dependency modeling, U-Net's ability to recover local high-frequency details is insufficient, manifested as low local peak amplitude and slight energy diffusion. Bandpass filtering can preserve the main phase structure to some extent, but due to the fixed frequency threshold, some high-frequency detail signals are suppressed, resulting in weakened peaks and flattened local waveforms. Simultaneously, some noise components within the threshold range are not completely filtered out, leading to weak perturbations in the time-domain waveform and limiting the overall filtering effect. Empirical mode decomposition (EMD) can recover the main contour of the signal in the time domain, but slight residual fluctuations are visible in strongly noisy samples, and some local peak positions show slight offsets from the original signal, indicating that its mode division has certain instability under low SNR conditions.

[0083] To analyze the spectral preservation performance of each method, a Fourier transform was performed on the noise reduction results. For example... Figure 3As shown in (f) to (i), the model in this embodiment exhibits significant noise suppression within the 20–100 Hz frequency band dominated by mine blasting and mechanical coupling noise, while completely preserving the main frequency energy structure of the microseismic signal in the 0–200 Hz range. The spectral distribution shows that the output of the model in this embodiment almost completely overlaps with the clean signal, the energy in the main frequency region is continuous and smooth, high-frequency (>250 Hz) noise is thoroughly suppressed, and no pseudo-frequency or energy gaps appear, demonstrating excellent spectral fidelity. In contrast, the spectral consistency of the other three methods is significantly worse: although U-Net can better maintain the mid-to-low frequency structure than traditional filtering methods, there is still a slight energy shift in the range below 50 Hz, and weak noise residue is still visible at the high-frequency end (>200 Hz), mainly because it is difficult to completely distinguish high-frequency random noise from detailed feature signals within a pure convolution framework; in addition, feature fusion at multiple scales may still introduce slight amplitude-frequency smoothing, weakening some high-frequency detail energy. Due to the fixed frequency threshold setting, bandpass filtering exhibits local energy loss in the 80-120Hz range, resulting in discontinuities at the spectral edges. Although empirical mode decomposition can adaptively decompose signal modes, it produces slight spurious peaks and energy oscillations in the 100-250Hz range, reflecting its mode aliasing and energy leakage problems in strong noise environments. That is, some effective spectral energy is incorrectly divided into noise modes, thereby destroying the frequency band continuity of the signal.

[0084] Therefore, the model in this embodiment can not only effectively eliminate the energy of the main frequency band of blasting and mechanical noise, but also maintain an amplitude-frequency response almost identical to that of a clean signal across multiple frequency bands. The channel attention mechanism enhances the dominant frequency band features through adaptive weighting, enabling the model to achieve high-precision signal recovery and spectral consistency in complex non-stationary noise environments, thereby achieving the best balance between noise reduction amplitude and energy fidelity.

[0085] Example 2

[0086] To further evaluate the generalization ability of the U-BiLSTM-CA-DDNet model in cross-mine scenarios, this embodiment tests the U-BiLSTM-CA-DDNet model trained in Example 1 on the Hongtoushan copper-zinc mine dataset. Figure 4The denoising effect of typical microseismic signals collected from the mining area is demonstrated, showing the original noisy input signal, the predicted signal output by the model, the separated predicted noise, and the corresponding frequency domain energy distribution. The results show that compared with the original noisy signal, the denoised waveform significantly removes periodic mechanical interference and random blasting noise, without exhibiting the amplitude attenuation or phase delay common in traditional filtering methods. Noise components in the frequency domain are significantly suppressed, and the effective signal frequency band structure remains intact. It is evident that the U-BiLSTM-CA-DDNet model not only performs excellently in the mining area where the training set is located, but also maintains stable denoising performance and key feature preservation capabilities when facing other mine data with significant differences in geological structure, noise characteristics, and mining processes, demonstrating good engineering applicability and potential for widespread application.

[0087] The above description represents a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for denoising low signal-to-noise ratio microseismic signals in mines, characterized in that, Includes the following steps: S1 collects raw microseismic signals from the mine and synthesizes Gaussian noise to jointly construct a labeled mixed signal sample set; S2, construct the U-BiLSTM-CA-DDNet model; The U-BiLSTM-CA-DDNet model is based on improvements to the U-Net architecture, including: embedding a bidirectional long short-term memory network between the encoder and decoder modules to capture temporal dependency features and energy decay trends; setting the decoder module as two parallel branches, one for predicting clean signals and the other for predicting noisy signals; and introducing a channel attention mechanism in the noisy branch decoder module to achieve adaptive adjustment of channel feature weights; specifically: The U-BiLSTM-CA-DDNet model includes an encoder module, a temporal modeling module, a bridging layer, a channel attention module, and a dual decoder module. The encoder module consists of multiple encoders connected in series. Each encoder layer contains two convolutional layers and one max pooling layer. After the mixed signal samples are input into the encoder module, multi-scale feature extraction is achieved by increasing the number of channels and reducing the temporal dimension layer by layer. The temporal modeling module consists of two layers of bidirectional long short-term memory network and one convolutional layer; the output of the encoder module is processed by the bidirectional long short-term memory network to extract temporal dependency features and energy decay trends, and then processed by the convolutional layer to achieve feature mapping and dimensionality transformation. The bridging layer includes two convolutional layers and a feature concatenation layer between the two convolutional layers. It is used to convolve the features output by the temporal modeling module after convolutional mapping with the features of the last layer of the encoder module and then perform convolutional fusion. The dual decoder module includes a clean signal branch and a noise signal branch set in parallel. Each branch includes multiple decoders in series, and they are symmetrical to the encoder module. Each decoder layer contains an upsampling layer, a splicing layer, a convolutional layer, and an output layer. The features of the corresponding encoder layer and the decoder layer are fused through skip connections. The signal resolution and length are gradually restored by combining upsampling, cropping, and padding operations. The clean signal branch outputs the predicted clean signal, and the noise signal branch outputs the predicted noise signal. The channel attention module is embedded in the convolutional layers of each decoder layer in the noise signal branch. It generates channel weights through global average pooling and a two-layer fully connected network to achieve adaptive adjustment of the importance of feature channels. S3. Construct a multi-output joint loss function and iterate the U-BiLSTM-CA-DDNet model multiple times until the error meets the preset requirements. S4. Input the microseismic signal to be processed into the U-BiLSTM-CA-DDNet model to obtain the predicted clean signal and the predicted noise signal.

2. The method for denoising low signal-to-noise ratio microseismic signals in mines according to claim 1, characterized in that, S1 specifically involves: acquiring raw microseismic signals from the mine; using a fixed time window to truncate each signal to a uniform length; filtering to obtain real noise signals and clean signals, and simultaneously synthesizing Gaussian noise, which is then used together with the real noise signals as noise signals; linearly superimposing the noise signals and clean signals according to a set signal-to-noise ratio to obtain a mixed signal; and performing Z-score normalization on the mixed signal to obtain a labeled mixed signal sample set.

3. The method for denoising low signal-to-noise ratio microseismic signals in mines according to claim 1, characterized in that, The channel attention module is specifically as follows: The input to the channel attention module is the output features of the convolutional layer. F First, global average pooling is used to process the input features. F Compression along the time dimension yields the channel statistics: (1)； Then, channel statistics Dimensionality reduction and dimensionality expansion are performed sequentially through two layers of fully connected networks: (2)； in, W For the generated channel weight vector, and These are the learnable weight matrices for two fully connected network layers, respectively. This represents the Sigmoid activation function. This represents the modified linear unit function; Finally, the generated channel weight vector W Input features F Fusion is achieved through channel-by-channel multiplication: (3)； Here, ⊙ represents element-wise multiplication. This is the output of the channel attention module.

4. The method for denoising low signal-to-noise ratio microseismic signals in mines according to claim 1, characterized in that, In S3, the multi-output joint loss function is as follows: (4)； (5)； (6)； in: t Indicates the index of the time sampling point. N This represents the total number of sampling points for a single signal sample. This represents the clean signal predicted by the model. c ( t This indicates a true, clean signal. This represents the noise signal predicted by the model. n ( t () represents the actual noise signal. L This is the total loss function; For clean signal loss; This represents the noise signal loss; MSE(*,*) represents the mean square error. α , β For weight hyperparameters.

Images