Microseismic signal noise reduction reconstruction method and system based on multi-scale decomposition

Abstract

The invention relates to the technical field of signal noise reduction and reconstruction, in particular to a microseismic signal noise reduction and reconstruction method and system based on multi-scale decomposition, and the method comprises the steps: S1, carrying out the multi-scale Shaplet decomposition of a noise-containing microseismic signal based on different sequence lengths, measuring the matching degree between each signal block and all candidate Shaplets through Euclidean distance, and obtaining the matching degree between each signal block and each candidate Shaplet; constructing a multi-scale feature matrix M; s2, based on the multi-scale feature matrix M, multi-scale weighted importance measurement is carried out through a time convolution network and an attention mechanism so as to evaluate the importance of each signal block in the noise reduction process; and S3, constructing a U-Net microseismic signal noise reduction model, designing a loss function in combination with the reconstruction error and dynamic time warping, and carrying out multiple iterative training on the U-Net microseismic signal noise reduction model until the error meets a preset requirement. According to the method, noise can be effectively identified and removed, and especially in data containing different noise, the method is beneficial to high-quality noise reduction of micro-seismic monitoring data.

Description

Technical Field

[0001] The present invention relates to the technical field of signal noise reduction and reconstruction, and particularly to a microseismic signal noise reduction and reconstruction method and system based on multi-scale decomposition. Background Art

[0002] The stability of open-pit mine slopes is crucial for mining safety. Real-time monitoring of microseismic signals can effectively warn of potential instability risks of slopes. However, in practical applications, the collected microseismic signals are often disturbed by complex environmental noises and mechanical vibrations, such as environmental noises and mechanical vibrations, resulting in a decline in signal quality.

[0003] Traditional signal noise reduction methods, such as low-pass filtering, short-time Fourier transform, and wavelet transform, all have certain limitations, such as the need for manual parameter setting, poor noise reduction effect, difficult parameter setting, and difficulty in obtaining a training set. Especially when facing multi-scale noises, obvious limitations are shown.

[0004] Currently, existing research has attempted to combine deep learning and signal decomposition techniques to optimize signal noise reduction effects. For example, unsupervised deep learning denoising: The seismic signal is transformed into the time-frequency domain through the short-time Fourier transform (STFT), the real and imaginary parts are separated and processed in blocks, an unsupervised deep learning network is used to predict the signal mask, a binary mask is generated through adaptive threshold processing to suppress the noise components in the time-frequency domain, and finally the denoised seismic signal is reconstructed through the inverse STFT, which can effectively improve the signal-to-noise ratio without clean data labels.

[0005] A new wavelet threshold denoising method based on discrete wavelet transform: Based on the adaptive threshold calculation method of wavelet coefficients at different decomposition levels, a continuously differentiable threshold function is constructed, and a more reasonable threshold function is determined by adjusting the preset shape adjustment parameter to achieve denoising processing for first arrival wave enhancement.

[0006] However, these methods often cannot effectively process multi-scale signal features and lack the targeted decomposition ability for noises at different time scales.

[0007] Based on the problems in the existing technology, the present invention provides a microseismic signal noise reduction and reconstruction method and system based on multi-scale decomposition. Summary of the Invention

[0008] The object of the present invention is to provide a microseismic signal noise reduction and reconstruction method and system based on multi-scale decomposition to solve the technical problem in the existing technology that when facing noisy microseismic signals, the processing level is single and it is difficult to perform comprehensive noise reduction processing from different time scales.

[0009] The technical solution of the present invention is: a microseismic signal denoising and reconstruction method based on multi-scale decomposition. This method extracts local features of the signal through Shapelet, combines a temporal convolutional network to score the importance of the signal, constructs a microseismic signal denoising model based on the U-Net network to remove noise and reconstruct the real microseismic signal; this method includes:

[0010] S1. Based on different sequence lengths, perform multi-scale Shapelet decomposition on the noisy microseismic signal, measure the matching degree between each signal block and all candidate Shapelets through the Euclidean distance, and construct a multi-scale feature matrix M;

[0011] S2. Based on the multi-scale feature matrix M, perform multi-scale weighted importance measurement through a temporal convolutional network and an attention mechanism to evaluate the importance of each signal block in the denoising process;

[0012] S3. Construct a U-Net microseismic signal denoising model, design a loss function by combining the reconstruction error and dynamic time warping, and perform multiple iterative trainings on the U-Net microseismic signal denoising model until the error meets the preset requirements.

[0013] Preferably, in S1, the method for performing multi-scale Shapelet decomposition on the noisy microseismic signal and constructing candidate Shapelets is as follows:

[0014] The input signal is S i ={x1,x2,...,x N},

[0015] where N is the sequence length, and the set of Shapelet lengths selected is L={L1,L2,...,L k}, for each length L j ∈L, use a sliding window with a fixed length L to slide point by point on the signal S, that is, the step size is 1;

[0016] For each length L, extract the subsequence: c m =S i [p:p+L j -1];

[0017] where p is the starting position of the window, satisfying 1≤p≤N-L j +1; c m represents a candidate Shapelet.

[0018] Preferably, based on the candidate Shapelet, the process of constructing the multi-scale feature matrix M is as follows:

[0019] All candidate subsequences extracted based on the candidate Shapelets constitute a candidate Shapelet set, and the formula is:

[0020] For the multi-scale case, merge the candidate Shaplets of all lengths:

[0021] Based on each candidate Shapelet For signal S i For each signal block in it, calculate its minimum matching distance:

[0022] where ||·|| is usually the Euclidean distance;

[0023] Organize all distance values into a feature matrix M, and the formula is:

[0024]

[0025] where the dimension of matrix M is N blocks ×N Shapelets , N blocks is the number of segments of signal S i ; N Shapelets is the total number of candidate Shapelets.

[0026] Preferably, in S2, score the importance of each signal block in the multi-scale feature matrix through a temporal convolutional network, and obtain the query vector Q = M·W Q , key vector K = M·W K and value vector V = M·W V ,

[0027] W Q , W K , W V are three trainable parameter matrices, where, d v is the dimension of the value vector, d k is the dimension of the key vector, and M is the multi-scale feature matrix;

[0028] Based on the query vector Q and the key vector K, calculate the dot product similarity, and the dot product similarity matrix is expressed as:

[0029]

[0030] where Q[i,:] and K[j,:] represent the vectors of the i-th and j-th signal blocks; d k is the dimension of the key vector and is used as a scaling factor;

[0031] The dot product similarity is transformed into an attention weight matrix A through Softmax normalization, and the weights represent the important relationships between signal blocks:

[0032]

[0033] Among them, A[i, j] represents the attention weight of the i-th signal block to the j-th signal block, and Score ij is the dot product similarity matrix;

[0034] The value vector V is weighted by the attention weight matrix A to obtain the weighted output feature O, and the expression is: O = A·V;

[0035] Among them, Each row O[i, :] represents the important features of the i-th signal block after weighting.

[0036] Preferably, the microseismic signal is denoised and reconstructed through a U-Net network, and the steps include:

[0037] The encoder extracts multi-scale high-level features by gradually increasing the number of convolutional channels layer by layer, that is, the number of channels in each layer is twice that of the previous layer, and gradually reduces the spatial resolution through pooling operations;

[0038] After each layer of convolution, the high-resolution features are saved. The skip connection directly transmits the output features of each layer of the encoder to the symmetric position of the decoder, and fuses them with the output features of the decoder during the decoding stage;

[0039] The decoder gradually restores the signal resolution through upsampling and convolution operations, generates a denoised signal, and restores it to the original feature dimension of the signal;

[0040] Among them, during model training, the output signal of the U-Net is compared with the real signal, and a combination of loss functions is used to balance the reconstruction error and the time alignment error, and optimize the reconstruction performance.

[0041] Preferably, the encoder gradually extracts high-level features, and the hierarchical structure of the encoder is as follows:

[0042] B1. Select a matrix with a 3×3 convolution kernel through convolution operations to slide on the input data, and perform element-wise multiplication and summation at each position to generate a feature map;

[0043] Among them, the number of convolution channels is C1; the activation function uses ReLU; the output dimension:

[0044] B2. Perform downsampling operations. Each operation can reduce the spatial size of the feature map and increase the depth of the features at the same time;

[0045] Among them, the window size is 2 and the step size is 2; the output dimension:

[0046] B3. Repeat the convolution and downsampling operations;

[0047] After each layer of convolution, the number of feature channels is increased to twice: C2, C3,..., C d ; the spatial resolution is gradually reduced to N blocks / 4, N blocks / 8,...

[0048] Preferably, the decoder gradually restores the signal resolution, generates a denoised signal through upsampling and convolution operations, and the hierarchical structure of the decoder is as follows:

[0049] J1. Perform upsampling, using transposed convolution or bilinear interpolation; each time, expand the resolution of the feature map by 2 times;

[0050] J2. Convolve the features after skip connection, and the convolution formula is as follows:

[0051]

[0052] J3. Output layer: Use a 1×1 convolution kernel to restore to the original feature dimension of the signal:

[0053] Among them,

[0054] Preferably, during the training process of the U-Net network, the following combination of loss functions is adopted to balance the reconstruction error and the time alignment error, and the total loss is:

[0055]

[0056] Among them, L reconstruction is the reconstruction error, L alignment is the time alignment error (DTW), and α, β are weight hyperparameters;

[0057] The optimizer adopts the Adam optimizer, and the learning rate is 10 -3 .

[0058] Preferably, the hyperparameters in the U-Net design are selected as follows: the initial number of channels in the encoder is selected as C1 = 32, and the number of channels in subsequent layers is gradually increased to 64, 128, 256;

[0059] The number of channels in the decoder is symmetric to that of the encoder, and the number of channels is gradually reduced to 256, 128, 64, 32;

[0060] Pooling uses 2×2 max pooling; upsampling selects bilinear interpolation or transposed convolution.

[0061] A microseismic signal denoising and reconstruction system based on multi-scale decomposition, comprising:

[0062] At least one processor; and,

[0063] A memory communicatively connected to at least one of the processors; wherein,

[0064] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the microseismic signal denoising and reconstruction method based on multi-scale decomposition.

[0065] Compared with the prior art, the advantages of the present invention are:

[0066] (1) The present invention applies multi-scale Shapelet decomposition to microseismic signal processing, provides a more refined signal decomposition method, and can accurately extract the local features of the signal at different time scales compared with traditional filtering or wavelet transform methods.

[0067] The present invention uses Euclidean distance to calculate the matching degree between Shapelet and the signal, and constructs a multi-scale feature matrix. The feature matrix can effectively capture the local change features of the signal in the time scale, so as to realize the efficient isolation and removal of noise. At the same time, this method performs excellently in retaining the integrity of the original signal, ensuring that key feature information will not be weakened during the denoising process, and laying a solid foundation for the accurate analysis of microseismic signals.

[0068] (2) The present invention introduces a temporal convolutional network to preliminarily process the input feature matrix, captures the temporal dependencies between signal segments, and evaluates the correlation of each signal part. Subsequently, an attention mechanism is introduced to dynamically score the importance of the signal. The attention mechanism further strengthens the weight allocation of key signal segments, ensures that key features are preferentially retained, and significantly improves the retention effect of important microseismic features compared with the traditional uniform processing method, thereby improving the accuracy of mine slope stability assessment.

[0069] (3) The present invention uses a U-Net architecture for signal denoising and reconstruction. The encoder-decoder structure combined with skip connections can effectively retain the high-frequency details of the signal and maintain the temporal integrity of the microseismic signal.

[0070] The encoder extracts multi-scale features layer by layer, and at the same time passes high-resolution features to the symmetric positions of the decoder through skip connections to ensure that important details are not lost. The decoder restores the temporal resolution of the signal through progressive upsampling and feature fusion, and realizes the high-precision reconstruction of real microseismic signals. This architecture can accurately restore the key features of microseismic signals while maintaining the integrity of their temporal and frequency domain information, and finally generate a more accurate and reliable denoised signal output, providing high-quality data support for microseismic monitoring.

[0071] (4) The present invention designs a loss function. By combining two loss metrics, namely the reconstruction error and the dynamic time warping error, it comprehensively measures the restoration quality and time alignment degree of the signal. In particular, the dynamic time warping error can finely adjust the small deviations in the signal time series, thereby further improving the authenticity and accuracy of signal reconstruction. This comprehensive loss function design enhances the robustness and adaptability of the model for microseismic signal denoising, and significantly improves the overall performance of signal processing. Brief Description of the Drawings

[0072] The present invention will be further described below in conjunction with the drawings and embodiments:

[0073] Figure 1 It is a method block diagram for microseismic signal denoising and reconstruction based on multi-scale decomposition according to the present invention;

[0074] Figure 2 It is a flow block diagram for microseismic signal denoising and reconstruction based on multi-scale decomposition according to the present invention;

[0075] Figure 3 It is a waveform diagram and signal-to-noise ratio comparison diagram of other methods for microseismic signal denoising and reconstruction on a synthetic dataset in Experiment 1 of the present invention;

[0076] Figure 4 It is a waveform diagram and signal-to-noise ratio schematic diagram of the present invention for microseismic signal denoising and reconstruction on a synthetic dataset in Experiment 1 of the present invention;

[0077] Figure 5 It is a waveform diagram and signal-to-noise ratio schematic diagram of the present invention for microseismic signal denoising and reconstruction on a real-recorded Stanford earthquake dataset in Experiment 2 of the present invention;

[0078] Figure 6 It is a waveform diagram and signal-to-noise ratio comparison diagram of other methods for microseismic signal denoising and reconstruction on a real-recorded Stanford earthquake dataset in Experiment 2 of the present invention. Detailed Embodiments

[0079] The content of the present invention will be further described in detail below in conjunction with specific embodiments:

[0080] As Figure 1As shown in the figure, a microseismic signal denoising and reconstruction method based on multi-scale decomposition includes:

[0081] Step 1: Obtain the original microseismic signal and preprocess the data set.

[0082] Collect and input the complete microseismic signal with noise, and divide the complete microseismic signal with noise into a training set and a validation set. In this embodiment, the training set is 70% and the validation set is 30%;

[0083] Denoise and reconstruct the complete microseismic signal with noise in the training set and the validation set using the same method.

[0084] Taking the 70% training set as an example below, denoise and reconstruct the complete microseismic signal with noise in the data set.

[0085] Step 2: Decompose the microseismic signal and perform feature extraction and feature representation.

[0086] Traditional denoising methods such as low-pass filtering, short-time Fourier transform, wavelet transform, etc. can often only process signal features at a single scale, are difficult to handle noises at different time scales simultaneously, and cannot comprehensively capture the changes of the signal at different time scales, resulting in unsatisfactory denoising effects.

[0087] In terms of multi-scale analysis ability: In this embodiment, Shapelet decomposition is adopted. For a single signal, sliding windows with different lengths (i.e., different sequence lengths) are used, which can analyze the features of the signal at different time scales simultaneously, providing a more refined signal decomposition method, especially suitable for processing complex noises;

[0088] In terms of local feature extraction: Shapelet decomposition can extract the local features of the signal, especially suitable for signals such as microseismic signals with obvious local change features. Through Shapelet decomposition, the local patterns of the signal at different time scales can be captured, so as to more accurately separate the noise and the true signal.

[0089] Perform multi-scale Shapelet decomposition on the microseismic signal with noise in the training set to construct a multi-scale feature matrix M; the method is as follows:

[0090] S21: The input signal is S i ={x1, x2,..., x N}, where N is the sequence length, and the set of Shapelet lengths is selected as L = {L1, L2,..., L k}; for each length L j ∈L, use a sliding window with a fixed length L to slide on the signal S, and the step size is 1, that is, slide point by point.

[0091] For each L, extract the subsequence: c m = S i [p:p + L j - 1];

[0092] where p is the starting position of the window, satisfying 1 ≤ p ≤ N - L j + 1; c m represents a candidate Shapelet.

[0093] S22. Based on the candidate Shapelets, form a candidate Shapelet set with all the candidate subsequences extracted by the above formula, and merge the candidate Shapelets of all lengths. The formula is:

[0094]

[0095] S23. Based on the candidate Shapelets of all lengths, construct a multi-scale feature matrix M to represent the matching degree between each signal block and all candidate Shapelets.

[0096] Among them, a signal block refers to: a smaller continuous subsequence extracted from the original noisy microseismic signal by sliding a window. Each signal block represents the local features of the signal at a certain moment. These features reflect the change trend of the signal during this period and represent the local patterns of the signal at different time resolutions. By calculating the matching degree between each signal block and all candidate Shapelets, the effective features of the signal can be extracted and the noise part can be removed.

[0097] For each signal block in S i and each candidate Shapelet, satisfy Extract the subsequence by sliding the window, calculate its Euclidean distance, calculate the matching degree between the Shapelet and the signal through the Euclidean distance, can effectively capture the local change features of the signal, and take the minimum matching distance (Euclidean distance) as the measurement standard, fill it into the matrix, and construct the multi-scale feature matrix M, so as to realize the efficient isolation and removal of noise. Through this measurement, the similarity and temporal correlation between the local features of the signal can be quantified, clearly identify which signal blocks match the noise, simplify the model structure, avoid repeated calculation of distances inside the model, reduce the computational burden during training, and facilitate the subsequent attention mechanism to weight and suppress noise. Calculate the matching distance between each signal block and all candidate Shapelets (c m ):

[0098]

[0099] where ||·|| is usually the Euclidean distance.

[0100] Organize the obtained distance values into a multi-scale feature matrix M, and the multi-scale feature matrix M is expressed as:

[0101]

[0102] Among them, the dimension of matrix M is N blocks ×N Shapelets , N blocks is the number of segments of signal S i ; N Shapelets is the total number of candidate Shapelets.

[0103] Noise and effective signals in microseismic signals often have specific waveform patterns (such as mechanical vibration noise being periodic and real microseismic signals being sudden). By extracting these patterns through multi-scale Shapelets, noise and signals can be separated.

[0104] Perform multi-scale decomposition and distance metric standardization before the noise reduction model. The multi-scale feature matrix obtained after preprocessing can clearly identify which signal blocks match the noise Shapelets, facilitating subsequent weighted suppression of noise by the attention mechanism. Generating a fixed feature matrix as input in the preprocessing stage can simplify the model structure, reduce complexity, and the computing power overhead of the model.

[0105] Step 3: Based on the multi-scale feature matrix M, perform importance measurement with multi-scale weighting through a temporal convolutional network (TCN) and an attention mechanism to evaluate the importance of each signal block in the noise reduction process. The method is as follows:

[0106] S31. Use a temporal convolutional network (TCN) to capture the temporal dependence of the multi-scale feature matrix and dynamically evaluate the relevance of each signal part.

[0107] Obtain the query vector Q = M·W Q , key vector K = M·W K and value vector V = M·W V ,

[0108] W Q , W K , W V , are three trainable parameter matrices, where d v is the dimension of the value vector, d k is the dimension of the key vector, and M is the multi-scale feature matrix.

[0109] Based on the query vector Q and the key vector K, calculate the dot product similarity, and the dot product similarity matrix is expressed as:

[0110]

[0111] where Q[i, :] and K[j, :] represent the vectors of the i-th and j-th signal blocks; d k is the dimension of the key vector, used as a scaling factor to prevent the dot product result from being too large, which may lead to the vanishing gradient problem.

[0112] The dot product similarity is transformed into an attention weight matrix A through Softmax normalization, and the weights represent the important relationships between signal blocks:

[0113]

[0114] where, A[i, j] represents the attention weight of the i-th signal block to the j-th signal block, and Score ij is the dot product similarity matrix.

[0115] S22. Based on the attention weight matrix A, the value vector V is weighted to obtain the weighted output feature O, and the expression is: O = A · V;

[0116] where, Each row O[i, :] represents the important features of the i-th signal block after weighting.

[0117] Signal blocks with higher importance will be strengthened through attention weights, while signal blocks with stronger noise components will be relatively weakened. The feature O is provided as input to the constructed noise reduction model to optimize the noise reduction effect.

[0118] Step 4. Denoise and reconstruct the microseismic signal through the U-Net architecture, and the denoising and reconstruction method is as follows:

[0119] This architecture includes an encoder, skip connections, and a decoder.

[0120] The encoder is composed of three stacked convolutional modules. The convolutional kernel size of each convolutional layer is 3×3, the sliding step of the convolutional kernel is 1, and the activation function uses ReLU. The decoder is composed of three stacked transposed convolutional modules.

[0121] Its various components and their specific implementations are as follows:

[0122] The first part: The encoder gradually extracts more abstract and general high-level features to capture the global patterns or long-term dependencies in the input feature O.

[0123] The hierarchical structure of the encoder is as follows:

[0124] B1. Through convolutional operations, a 3×3 convolutional kernel matrix slides on the input data, and element-wise multiplication and summation are performed at each position to generate a feature map.

[0125] Among them, the initial number of convolutional channels is 32; the activation function uses ReLU; output dimension:

[0126] B2. Perform max pooling. Through downsampling operations with a window size of 2 and a stride of 2, capture a larger receptive field, reduce the spatial size of the feature map, and increase the depth of the features simultaneously. In this way, higher-level features can be extracted while improving the network robustness. Among them, the output dimension:

[0127] B3. Repeat the convolution and downsampling operations to reduce the spatial resolution and increase the number of channels. After each layer of convolution, perform two repeated "convolution + downsampling" operations to gradually abstract the multi-scale global features of the signal, capture the global patterns and temporal dependencies. The number of feature channels increases to twice that of the previous layer: C2, C3,..., C d ; the spatial resolution gradually decreases to half of the previous layer, denoted as N blocks / 4, N blocks / 8,...

[0128] The second part: The skip connection is a key mechanism in the U-Net architecture and acts as a bridge in U-Net. After each layer of convolution, the skip connection directly transfers the output features of each layer of the encoder to the symmetric layer of the decoder, and fuses the output features of each layer of the encoder with the output features of the decoder through a concatenation operation. By concatenating channels to fuse high- and low-resolution features, avoid the loss of high-frequency details caused by downsampling, and improve the signal reconstruction accuracy.

[0129] To a certain extent, the skip connection mechanism solves the problem of possible loss of feature information during the upsampling process and enhances the network's reconstruction ability.

[0130] The third part: The decoder gradually restores the signal resolution and generates a denoised signal through upsampling and convolution operations.

[0131] Extract the global features of the signal by gradually reducing the resolution, and then restore the detailed parts during the process of restoring the resolution.

[0132] The hierarchical structure of the decoder is as follows:

[0133] Then concatenate the encoder features passed by the skip connection with the decoder features and perform convolution operations to gradually reduce the number of channels. After each layer of transposed convolution, perform two repeated "upsampling + convolution" operations, and finally use a 1×1 convolution kernel to restore the number of channels to the original signal dimension to generate a denoised signal.

[0134] J1. Perform upsampling. In each layer of the module, use transposed convolution or bilinear interpolation to restore the spatial resolution of the feature map from a lower dimension to a higher dimension, gradually restoring the detailed part of the signal. Each time, double the resolution of the feature map.

[0135] J2. Through skip connections, concatenate the output features of each layer of the encoder with the corresponding layer features of the decoder, and perform a convolution operation after concatenation to gradually reduce the number of channels. The convolution formula is as follows:

[0136]

[0137] J3. After each layer of transposed convolution, perform two repeated "upsampling + convolution" operations, and use a 1×1 convolution kernel to restore to the original feature dimension of the signal, generating the denoised signal. The 1×1 convolution kernel formula is expressed as:

[0138] Among them, L is the length of the signal block.

[0139] In model training, this embodiment uses a hybrid loss function to optimize the reconstruction performance of the model. The total loss function consists of a reconstruction error and a time alignment error:

[0140]

[0141] Among them, L reconstruction is the reconstruction error, which measures the difference between the output signal and the true signal; L alignment is the time alignment error (DTW), which evaluates the alignment degree of the signal in the time series. α and β are weight hyperparameters.

[0142] Perform denoising reconstruction on the noisy microseismic signals in 30% of the validation set; the method of performing denoising reconstruction on the noisy microseismic signals in the validation set is the same as that of the noisy microseismic signals in the training set, and will not be elaborated here.

[0143] Refer to Appendix Figure 3 , in Experiment 1 provided by the present invention, on the synthetic microseismic data set, the comparative experiment uses unsupervised deep learning denoising and a new wavelet threshold denoising method based on discrete wavelet transform, which have achieved remarkable results in the field of microseismic signal denoising, to calculate the signal-to-noise ratio (the larger the signal-to-noise ratio, the better) result schematic diagram of the output signal and the true signal; combined with Appendix Figure 4 , shows the signal-to-noise ratio schematic diagram of the output signal and the true signal calculated by using the denoising method provided by the present invention for denoising.

[0144] From the comparison results of Experiment 1, it can be seen that compared with the short-time Fourier transform and wavelet transform noise reduction methods, the waveform trend of the present invention is highly consistent with the real signal, and the steep rising and falling edges of the microseismic events are completely retained; the amplitudes of the wave peaks and wave valleys have an error of less than 5% compared with the real signal, the wave peak amplitude is restored to 0.98 (error 2%), and the wave valley is restored to -0.28 (error 4%); the amplitude of the high-frequency spike is restored to 0.58 (error 3%), and the shape is consistent with the real signal. The background noise amplitude is reduced from 0.2 to 0.02 (a decrease of 90%), and the pseudo-peaks are completely eliminated; the signal-to-noise ratios are increased by 6.34 dB and 7.00 dB respectively, and the noise reduction effect is better.

[0145] Referring to the attached Figure 5 , in Experiment 2 provided by the present invention, the real recorded Stanford earthquake dataset is selected, and the comparative experiment uses the unsupervised deep learning denoising and the new wavelet threshold denoising method based on discrete wavelet transform that have achieved remarkable results in the field of microseismic signal noise reduction. Combining with the attached Figure 6 As shown, the noise reduction method provided by the present invention is used for noise reduction, and the signal-to-noise ratio of the output signal and the real signal is calculated (the larger the signal-to-noise ratio, the better). From the comparison results of Experiment 2, it can be seen that compared with the short-time Fourier transform and wavelet transform noise reduction methods, the waveform trend of the present invention is highly consistent with the real signal, the main shock phase single peak is sharp, which is consistent with the sudden characteristics of the real microseismic event; the amplitudes of the wave peaks and wave valleys have an error of less than 5% compared with the real signal, the wave peak amplitude error is only 2.16%, and the wave valley amplitude error is 3.28; the high-frequency spike amplitude is restored to 98%, and the shape is consistent with the real signal. The background noise is reduced by 97%, and the pseudo-peaks are completely eliminated; the signal-to-noise ratios are increased by 11.67 dB and 12.02 dB respectively.

[0146] Based on the comparative examples of the synthetic microseismic dataset and the real data records respectively, the superiority of the present invention in the noise reduction and reconstruction of microseismic signals under complex noise combinations is verified compared with other methods.

[0147] The present invention also provides a microseismic signal noise reduction and reconstruction system based on multi-scale decomposition, including:

[0148] At least one processor; and,

[0149] A memory communicatively connected to the at least one processor; wherein,

[0150] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the microseismic signal noise reduction and reconstruction method based on multi-scale decomposition.

[0151] The present invention applies multi-scale Shapelet decomposition to microseismic signal processing, which can accurately extract the local features of the signal at different time scales. This is particularly crucial for the processing of complex microseismic signals contaminated by various noises. Traditional filtering or wavelet transform methods have obvious limitations in dealing with noises at different scales, while Shapelet decomposition provides a more refined signal decomposition method through its unique multi-scale analysis ability.

[0152] The Euclidean distance is used to calculate the matching degree between the Shapelet and the signal, and a multi-scale feature matrix is constructed. The feature matrix can effectively capture the local change features of the signal in the time scale, so as to realize the efficient isolation and removal of noises. At the same time, this method performs excellently in preserving the integrity of the original signal, ensuring that the key feature information will not be weakened during the noise reduction process, laying a solid foundation for the accurate analysis of microseismic signals.

[0153] The present invention introduces a temporal convolutional network (TCN) to preliminarily process the input feature matrix, capture the temporal dependencies between signal segments, and evaluate the relevance of each signal part. Subsequently, an attention mechanism is introduced to dynamically score the importance of the signal. The attention mechanism further strengthens the weight assignment of key signal segments, ensuring that key features are preferentially retained. Compared with the traditional uniform processing method, this method significantly improves the retention effect of important microseismic features, thus improving the accuracy of mine slope stability assessment.

[0154] The present invention uses a U-Net architecture for signal denoising and reconstruction. The encoder-decoder structure combined with skip connections can effectively retain the high-frequency details of the signal, which is crucial for maintaining the temporal integrity of microseismic signals. The encoder extracts multi-scale features layer by layer, and at the same time, through skip connections, high-resolution features are transmitted to the symmetric positions of the decoder to ensure that important details are not lost. The decoder restores the temporal resolution of the signal through gradual upsampling and feature fusion, realizing the high-precision reconstruction of real microseismic signals. This architecture can accurately restore the key features of microseismic signals while maintaining the integrity of their temporal and frequency domain information, and finally generate a more accurate and reliable denoised signal output, providing high-quality data support for microseismic monitoring.

[0155] The present invention designs a loss function. By combining two loss metrics, namely the reconstruction error and the dynamic time warping error, it comprehensively measures the restoration quality and temporal alignment degree of the signal. In particular, the dynamic time warping error can finely adjust the small deviations in the signal time series, thereby further improving the authenticity and accuracy of signal reconstruction. This comprehensive loss function design enhances the robustness and adaptability of the model for microseismic signal denoising, and significantly improves the overall performance of signal processing.

[0156] The above embodiments are only used to illustrate the technical concept and features of the present invention. The purpose is to enable those who are familiar with this technology to understand the content of the present invention and implement it accordingly, and it cannot be used to limit the protection scope of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic features of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to embrace all changes that fall within the meaning and scope of the equivalent elements of the claims in the present invention.

Claims

1. A method for denoising and reconstructing microseismic signals based on multi-scale decomposition, characterized in that: This method extracts local features of the signal through Shapelet, scores the importance of the signal in combination with the temporal convolutional network, builds a microseismic signal denoising model based on the U-Net network, removes noise and reconstructs the real microseismic signal; the method includes: S1. Based on different sequence lengths, the noisy microseismic signal is decomposed into multi-scale Shapelet, the matching degree between each signal block and all candidate Shapelets is measured by Euclidean distance, and a multi-scale feature matrix M is constructed; S2. Based on the multi-scale feature matrix M, a multi-scale weighted importance measurement is performed through a temporal convolutional network and an attention mechanism to evaluate the importance of each signal block in the denoising process; S3. Construct a U-Net microseismic signal denoising model, design a loss function by combining reconstruction error and dynamic time warping, and perform multiple iterative training on the U-Net microseismic signal denoising model until the error meets the preset requirements.

2. According to the method of claim 1, the method is characterized by: In S1, the noisy microseismic signal is decomposed into multi-scale Shapelets, and the method of constructing candidate Shapelets is as follows: The input signal is S i ={x1,x2,...,x N }, Where N is the sequence length, and the length set of Shapelet is selected as L = {L1, L2, ..., L k }, for each length L j ∈L, a sliding window of fixed length L is used to slide point by point on the signal S, that is, the step size is 1; For each length L, extract the subsequence: c m =S i [p:p+L j -1]; Where p is the starting position of the window, satisfying 1≤p≤NL j +1;c m Represents a candidate Shapelet.

3. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 2 is characterized in that: Based on the candidate Shapelet, the candidate constructs a multi-scale feature matrix M, and the process is as follows: Based on the candidate Shapelet, all extracted candidate subsequences constitute a candidate Shapelet set, and the formula is: For the multi-scale case, candidate Shaplet of all lengths are merged: Based on each candidate Shape et satisfies For the signal S i For each signal block in , calculate its minimum matching distance: where ||·|| is usually the Euclidean distance; All distance values ​​are organized into a feature matrix M, the formula is: Among them, the dimension of matrix M is N blocks ×N Shapelets , N blocks The signal S i The number of segments; N Shapelets is the total number of candidate Shapelets.

4. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 3 is characterized in that: In S2, The importance of each signal block in the multi-scale feature matrix is ​​scored through a temporal convolutional network, and the query vector Q = M W is obtained through linear transformation. Q , key vector K = M·W K Sum vector V = M·W V , Among them, W Q , W K , W V , are three trainable parameter matrices, where d v is the dimension of the value vector, d k is the dimension of the key vector, M is the multi-scale feature matrix; Based on the query vector Q and the key vector K, the dot product similarity is calculated, and the dot product similarity matrix is ​​expressed as: Where Q[i,:] and K[j,:] represent the vectors of the i-th and j-th signal blocks; d k is the dimension of the key vector, used as a scaling factor; The dot product similarity is converted into the attention weight matrix A through Softmax normalization, and the weight represents the importance relationship between signal blocks: in, A[i,j] represents the attention weight of the i-th signal block to the j-th signal block, Score ij is the dot product similarity matrix; The value vector V is weighted by the attention weight matrix A to obtain the weighted output feature O, expressed as: O = A·V; in, Each row O[i,:] represents the important features of the i-th signal block after weighting.

5. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 2 is characterized in that: The microseismic signal is denoised and reconstructed through the U-Net network. The steps include: The encoder increases the number of convolution channels layer by layer, that is, the number of channels in each layer is twice that of the previous layer, and gradually reduces the spatial resolution through pooling operations, thereby extracting multi-scale high-level features; After each convolution layer, high-resolution features are saved, and the jump connection directly transfers the output features of each layer of the encoder to the symmetrical position of the decoder, and fuses them with the output features of the decoder in the decoding stage; The decoder gradually restores the signal resolution through upsampling and convolution operations, generates a noise-reduced signal, and restores the original feature dimension of the signal; In the model training, the U-Net output signal is compared with the real signal, and a loss function combination is used to balance the reconstruction error and time alignment error to optimize the reconstruction performance.

6. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 5 is characterized in that: Features: The encoder gradually extracts high-level features. The encoder's hierarchical structure is as follows: B1. Select a 3×3 convolution kernel matrix through the convolution operation to slide on the input data, and perform element-level multiplication and summation at each position to generate a feature map; Among them, the number of convolution channels is C1; the activation function uses ReLU; the output dimension is: B2. Perform downsampling operations. Each operation can reduce the spatial size of the feature map while increasing the depth of the feature. Among them, the window size is 2, the step size is 2; the output dimension is: B3, repeat convolution and downsampling operations; After each convolution layer, the number of feature channels increases by two times: C2, C3, ..., C d ; The spatial resolution is gradually reduced to N blocks / 4,N blocks / 8,....

7. According to the method for denoising and reconstructing microseismic signals based on multi-scale decomposition of claim 5, the decoder gradually restores the signal resolution and generates a denoised signal through upsampling and convolution operations. The hierarchical structure of the decoder is as follows: J1. Upsample, using deconvolution or bilinear interpolation; each time the resolution of the feature map is doubled; J2. Convolve the features after the jump connection. The convolution formula is as follows: J3, output layer: Use a 1×1 convolution kernel to restore the original feature dimension of the signal: in, 8. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 6, characterized in that: During the U-Net network training process, the following loss function combination is used to balance the reconstruction error and time alignment error. The total loss is: Among them, L reconstruction is the reconstruction error, L alignment is the time alignment error, α and β are weight hyperparameters; The optimizer uses Adam optimizer with a learning rate of 10 -3 .

9. The method for denoising and reconstructing microseismic signals based on multi-scale decomposition according to claim 6, characterized in that: The hyperparameters selected in the U-Net design are: the initial number of channels of the encoder is selected as C1=32, and the number of channels in subsequent layers is gradually increased to 64, 128, and 256. The number of decoder channels is symmetrical to that of the encoder, and the number of channels is gradually reduced to 256, 128, 64, and 32; The pooling uses 2×2 maximum pooling; the upsampling uses bilinear interpolation or deconvolution.

10. A microseismic signal denoising and reconstruction system based on multi-scale decomposition, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the microseismic signal denoising and reconstruction method based on multi-scale decomposition according to any one of claims 1 to 9.

Images