A microseismic data processing method under low signal-to-noise ratio condition

Abstract

The present application relates to a kind of low signal-to-noise ratio conditions under microseismic data processing method, belong to microseismic data processing technical field, solve the initial arrival picking method in prior art Poor adaptability to industrial microseismic data, cannot utilize multi-channel two-dimensional spatial structure information, cost is high, the problem of unstable prediction accuracy. Obtain multi-channel microseismic data, and the multi-channel microseismic data is pretreated;Based on the multi-channel microseismic data after pretreatment, target seismic data matrix is constructed;Microseismic denoising data is obtained by using denoising network model to the target seismic data matrix processing;Event window is intercepted to the microseismic denoising data, and event level waveform matrix is obtained;Binary mask image is obtained based on the event level waveform matrix and pre-training semantic segmentation network;The initial time corresponding to P wave and S wave is calculated based on the binary mask image.

Description

Technical Field

[0001] This invention relates to the field of microseismic data processing technology, and in particular to a method for processing microseismic data under low signal-to-noise ratio conditions. Background Technology

[0002] Microseismic monitoring is a crucial technology for monitoring the safety of shale gas, tight oil and gas, geothermal development, mining, carbon dioxide seismic storage, and underground engineering. By deploying geophone arrays in wells, tunnels, on the surface, or other monitoring areas, seismic wave signals generated by underground micro-fracture events can be continuously received. This data is then used for event detection, phase picking, event location, focal mechanism inversion, and reservoir stimulation effect evaluation. In the microseismic data processing workflow, picking the first arrival times of P-waves and S-waves is a fundamental and critical step. The accuracy of the first arrival times directly affects subsequent event location, velocity model correction, focal mechanism calculation, and fracture propagation interpretation. Automatic and accurate first arrival picking is essential for passive seismic monitoring because it directly impacts event location and focal mechanism estimation results.

[0003] In the existing technology, the first arrival picking methods for industrial microseismic data mainly include three types: (1) traditional first arrival picking methods, which mainly include STA / LTA (short-time average / long-time average ratio method), energy feature method, AIC (Akaike information content criterion), cross-correlation or matched filtering methods, etc. However, these methods have obvious shortcomings in adaptability to industrial microseismic data and are prone to problems such as mispicking, missed picking or unstable picking; (2) single-channel deep learning methods, but these methods usually process seismic records independently channel by channel, ignoring the common problems in industrial microseismic monitoring arrays. The fundamental problem with cross-channel spatial continuity and the overall shape of the first arrival boundary is that the first arrival picking of industrial microseismic events is simplified to a one-dimensional time series problem, without making full use of its multi-channel two-dimensional spatial structure; (3) Multi-channel or array-type deep learning, but such methods usually do not explicitly represent the absolute coordinates of the detector position, channel number position, time sampling position or two-dimensional image pixels. Spatial information is only implicit in the arrangement of input data. The model needs to learn geometric relationships by itself through complex structures, resulting in an increase in the number of model parameters, training costs and deployment costs, and cannot guarantee a stable improvement in picking accuracy.

[0004] Therefore, there is a need for a highly adaptable, cost-effective, and accurate method for picking first arrival times that utilizes multichannel two-dimensional spatial structure information. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a microseismic data processing method under low signal-to-noise ratio conditions, in order to solve the problems of poor adaptability of existing first-arrival picking methods to industrial microseismic data, inability to utilize multi-channel two-dimensional spatial structure information, high cost, and unstable prediction accuracy.

[0006] This invention provides a method for processing microseismic data under low signal-to-noise ratio conditions, the method comprising: Acquire multichannel microseismic data and preprocess the multichannel microseismic data; Construct a target seismic data matrix based on preprocessed multichannel microseismic data; The target seismic data matrix is ​​processed using a denoising network model to obtain denoised microseismic data; The event-level waveform matrix is ​​obtained by performing event windowing on the microseismic denoising data; A binary mask image is obtained based on the event-level waveform matrix and the pre-trained semantic segmentation network; The first arrival times of the P-wave and S-wave are calculated based on the binary mask image.

[0007] Based on a further improvement of the above method, the denoising network model includes an input layer, an encoder module, a denoising module, and a decoder module connected in sequence. The input layer is used to receive the target seismic data matrix; the encoder module is used to extract features and downsample the target seismic data matrix; the denoising module is used to denoise the output of the encoder module; and the decoder module is used to upsample the output of the denoising module and output the upsampled result.

[0008] Based on a further improvement of the above method, the encoder module includes a first convolution module, a first downsampling layer, a second convolution module, a second downsampling layer, and a third convolution module connected in sequence. The decoder module includes a first upsampling layer and a second upsampling layer connected in sequence; The denoising module includes a strong denoising module, a weak denoising module, and a gated fusion module; wherein, the strong denoising module processes the output of the encoder module to obtain a first denoising result, the weak denoising module processes the first denoising result to obtain a second denoising result; and the gated fusion module calculates a denoised feature map based on the first denoising result and the second denoising result.

[0009] Based on a further improvement of the above method, the strong denoising module includes a first fully connected layer, a LeakyReLU activation function layer, a second fully connected layer, a LeakyReLU activation function layer, and a first self-attention layer connected in sequence. The weak denoising module includes a third fully connected layer, a LeakyReLU activation function layer, a fourth fully connected layer, a LeakyReLU activation function layer, and a second self-attention layer connected in sequence.

[0010] As a further improvement on the above method, the processing the target seismic data matrix by using a denoising network model to obtain microseismic denoised data comprises: performing sliding window processing on the target seismic data matrix, dividing the matrix into a plurality of different sliding window data blocks, and inputting each sliding window data block into the denoising network model to obtain a denoised waveform; implementing fusion of the denoised waveforms of all the sliding window data blocks based on an overlap-add method, so as to obtain the microseismic denoised data corresponding to the target seismic data matrix.

[0011] As a further improvement on the above method, the input dimension of the input layer is × × ; is the number of sampling points; is the number of channels; is the number of input features; wherein the input features at least comprise a channel weight.

[0012] As a further improvement on the above method, the channel weight is determined in the following manner: , , wherein, is the channel weight corresponding to the i-th trace of microseismic data; is the minimum continuous weight; is the anomaly threshold; , , , , , , are weight coefficients; is the signal-to-noise ratio of the i-th trace of microseismic data; is the peak amplitude of the i-th channel; is the short-term energy of the i-th trace of microseismic data; is the energy proportion of the main frequency band of the i-th trace of microseismic data; is the inter-trace correlation of the i-th trace of microseismic data; is the geometric reliability of the i-th trace of microseismic data; is a normalized penalty term for the deviation degree; for , when x≤0, output 0; when 0<x<1, output x; when x≥1, output 1; i=1,2,3,…,N, where N is the total number of channels.

[0013] As a further improvement on the above method, the loss function of the denoising network model is: , , , , , , , in, , , , These are the weighting coefficients. For mask reconstruction loss, For inter-channel coherence loss, To ensure the accuracy of the initial candidate area, This is the residual constraint loss; This is the mask matrix corresponding to the b-th sliding window data block. The nonlinear mapping function representing the denoising network, This indicates element-wise multiplication. This is the b-th sliding window data block after being masked. To introduce channel weights, the b-th sliding window data block contains the original noisy data. To prevent small constants with a denominator of zero, This is the mask replacement value matrix corresponding to the b-th sliding window data block; For the set of neighborhood channel pairs, Let i be the total number of channel pairs within the set, and let i and j represent adjacent channel pairs, where i ≠ j. The inter-channel coherence coefficient between the i-th and j-th channels after denoising. The inter-channel coherence coefficients between the i-th and j-th channels of the original noisy data; Let p be the initial arrival pick mask for the b-th sliding window data block, where p represents the pick. This refers to the b-th denoised sliding window data block output by the denoising network model. This refers to the noise residual block corresponding to the b-th sliding window data block output by the denoising network model. , These are the weighting coefficients for gradient regularization in the time direction. These are the weighting coefficients for inter-channel directional gradient regularization. , These are the difference operators for the vertical and horizontal directions, respectively.

[0014] Based on a further improvement to the above method, the event window is obtained in the following way: Obtain the theoretical arrival time of the P-wave and the theoretical arrival time of the S-wave for each channel; The minimum theoretical arrival time of the P-wave in all channels is taken as the target value of the P-wave, and the maximum theoretical arrival time of the S-wave in all channels is taken as the target value of the S-wave. The start time of the event window is determined based on the preset pre-buffer and the target value of the P-wave, and the end time of the event window is determined based on the target value of the S-wave and the preset post-buffer.

[0015] Based on the further improvement of the above method, the horizontal axis of the binary mask image is the channel number, the vertical axis is the sampling time, and the matrix element value is 0 or 1, where 0 indicates that the point belongs to the PS wave interval and 1 indicates that the point does not belong to the PS wave interval. The calculation of the first arrival times of the P-wave and S-wave based on the binary mask image includes: For each channel, perform the following operation: Calculate the effective length of each column under the channel, and take the maximum effective length as the target longest chain of the column. If the length of the target longest chain is not 0, then take it as an element in the set of effective columns corresponding to the channel; wherein, the starting row number of the target longest chain corresponds to the first arrival time of the P wave, and the ending row number corresponds to the first arrival time of the S wave. The first arrival sampling points corresponding to the P-wave and S-wave are calculated based on the effective column set. The P-wave first arrival time and S-wave first arrival time corresponding to this channel are calculated based on the first arrival sampling point and the event window.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention provides a microseismic data processing method under low signal-to-noise ratio (SNR) conditions. Based on a denoising network, it denoises microseismic data, effectively suppressing background random noise and mechanical interference, preserving waveform morphology, polarity, and amplitude characteristics to the greatest extent, and significantly improving the signal identifiability under low SNR conditions, thereby further reducing the missed detection rate of weak microseismic events. Furthermore, the denoising network used in this invention employs a cascaded structure combining strong and weak denoisers. The strong denoiser suppresses random noise, local anomalous amplitudes, and incoherent interference, while the weak denoiser recovers weak first arrivals, phase abrupt changes, and relative amplitudes based on the strong denoising results. This two-stage structure allows the model to suppress strong noise while avoiding the complete removal of low-amplitude first arrivals as noise, thus further improving the identifiability of microseismic signals.

[0017] 2. This invention provides a microseismic data processing method under low signal-to-noise ratio conditions. It transforms first arrival picking into a semantic segmentation task of the effective region of PS waves. By utilizing the modeling capability of deep learning networks for two-dimensional features between time and traces, it fully captures the inter-trace continuity and time-domain morphological features of the microseismic phase axis. Compared with traditional first arrival picking methods, the method proposed in this invention can more accurately distinguish between effective signals and discrete noise, significantly reduce the probability of false picking under strong interference environments, and improve the overall accuracy of first arrival identification.

[0018] 3. This invention provides a microseismic data processing method under low signal-to-noise ratio conditions. To address the defects of segmentation results that are prone to holes, fractures, and isolated false positive areas under low signal-to-noise ratio conditions, this invention uses post-processing logic such as longest white chain identification, discontinuity fault tolerance, and multi-column statistical aggregation to automatically repair fractures in effective areas, filter isolated noise interference, avoid the impact of local segmentation errors on the first arrival boundary, and output P and S wave first arrival times with higher stability, thus reducing the cost of manual verification and correction.

[0019] 4. This invention provides a microseismic data processing method under low signal-to-noise ratio conditions. The entire process, from input of raw multichannel data to output of P and S wave arrival times, can be automated. After model training, the inference stage does not require manual annotation of each channel and parameter tuning of each event. It has high processing efficiency and can be adapted to batch processing scenarios of massive continuous data such as mine microseismic monitoring, which greatly reduces manual processing costs and improves data processing timeliness.

[0020] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0021] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is an example diagram illustrating a microseismic data processing method under low signal-to-noise ratio conditions, provided as an embodiment of the present invention. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0023] A specific embodiment of the present invention discloses a method for processing microseismic data under low signal-to-noise ratio conditions, such as... Figure 1 As shown, the method includes: Acquire multichannel microseismic data and preprocess the multichannel microseismic data; Construct a target seismic data matrix based on preprocessed multichannel microseismic data; The target seismic data matrix is ​​processed using a denoising network model to obtain denoised microseismic data; The event-level waveform matrix is ​​obtained by performing event windowing on the microseismic denoising data; A binary mask image is obtained based on the event-level waveform matrix and the pre-trained semantic segmentation network; The first arrival times of the P-wave and S-wave are calculated based on the binary mask image.

[0024] This invention provides a method for processing microseismic data under low signal-to-noise ratio (SNR) conditions. This method can obtain the first arrival times of P-waves and S-waves corresponding to multi-channel microseismic data. Specifically, the method first uses a custom network model to denoise the multi-channel microseismic data (the denoising process can be regarded as a filtering process). The denoised data is then used as input to a pre-trained semantic segmentation network model to obtain the corresponding binary mask image. Finally, the first arrival times of P-waves and S-waves are obtained through a preset post-processing method. It can be seen that the method proposed in this invention significantly improves the signal recognizability under low SNR conditions by suppressing background random noise and mechanical interference, preserving waveform morphology, polarity, and amplitude characteristics to the greatest extent, thereby further reducing the missed detection rate of weak microseismic events. Furthermore, the first arrival times of P-waves and S-waves output by the proposed post-processing method are more stable, reducing the cost of manual verification and correction.

[0025] The following content provides a detailed description of a microseismic data processing method under low signal-to-noise ratio conditions.

[0026] S1: Acquire multichannel microseismic data and preprocess the multichannel microseismic data.

[0027] The multichannel microseismic data can be acquired by single-component geophones, three-component geophones, distributed acoustic sensing fibers, or other seismic sensing devices; that is, multichannel microseismic data corresponds to multichannel waveform data. Specifically, each microseismic event corresponds to several waveform records, and each waveform record contains a time sampling sequence of a preset length. The method proposed in this invention is applicable to the acquisition of P-wave first arrival times and S-wave first arrival times in microseismic monitoring scenarios such as wells, tunnels, mines, the ground, or other industrial applications.

[0028] After acquiring multichannel microseismic data (i.e., multichannel waveform data), preprocessing is required. Specifically, preprocessing operations may include bandpass filtering, mean removal, trend removal, amplitude normalization, outlier removal, channel reordering, and waveform quality evaluation. Channel reordering is a spatial alignment operation in microseismic data preprocessing, which rearranges the acquired multichannel data from the hardware channel number order to the physical spatial position order of the detectors, ensuring that the horizontal dimension of the final two-dimensional waveform matrix corresponds one-to-one with the distribution order of the detectors in actual space. This invention does not limit the specific implementation method of the preprocessing operations; those skilled in the art can freely choose the preprocessing operation method and execution order based on experience, as long as it improves the prediction accuracy of the P-wave and S-wave first arrival times.

[0029] S2: Construct the target seismic data matrix based on the preprocessed multichannel microseismic data.

[0030] This step involves discretizing the preprocessed multichannel microseismic data, specifically constructing a matrix based on the correspondence between each time sampling point, channel, and amplitude. Where T is the number of time sampling points and M represents the number of channels. If three-component data is used, M can represent the product of the number of detectors and the number of components, or different components can be processed as different input channels.

[0031] S3: Process the target seismic data matrix using a denoising network model to obtain denoised microseismic data, including: S31: Perform sliding window processing on the target seismic data matrix, divide the matrix into multiple different sliding window data blocks, and input each sliding window data block into a denoising network model to obtain a denoised waveform.

[0032] In fact, the size of complete target seismic data is usually very large. If it is fed into the network for inference all at once, it may be impossible to calculate due to hardware reasons such as insufficient video memory. Therefore, this invention cuts the large matrix into many small pieces (Patch), feeds them into the network for denoising one piece at a time, and then splices the small pieces back into the complete matrix to obtain complete microseismic denoised data. This method can improve the computational efficiency to a certain extent.

[0033] Specifically, the target seismic data matrix is ​​subjected to sliding window processing to divide the matrix into multiple different sliding window data blocks, including: sliding the target seismic data matrix based on a preset sliding window to obtain multiple different sliding window data blocks.

[0034] Each sliding window data block can be represented as: , P is the b-th sliding window data block containing the original noisy data; P() is the sliding window function, which performs sliding processing on the preprocessed multichannel microseismic data X to obtain multiple sliding window data blocks. This represents the number of sampling points; C represents the number of channels; C represents the number of input features, which must include at least one of the following: channel weights, original waveform amplitude, normalized time coordinates, normalized channel coordinates, initial arrival candidate masks, or noise estimation maps. The preset sliding window includes window size, horizontal sliding step size, and vertical sliding step size. The sliding operation uses the time-channel two-dimensional matrix as the processing object, which can be denoted as the time-direction sliding step size. and the sliding step length in the inter-lane direction For example, setting the sampling interval. With a single-event time window of 880 sampling points and 12 three-component data points in the well (corresponding to 36 channels), in this case, the following settings can be configured: or Simultaneously set the time direction step size Alternatively, 256 is acceptable; in this case, there is no slippage in the channel direction. When the number of channels is large, this can be adjusted. Set the sliding step size to For DAS or large-scale terrestrial arrays, the following settings can be configured: or Set the sliding step size to or .

[0035] In this design, adjacent data blocks overlap in both the time and channel directions. When the boundary is insufficient to form a complete block, zero padding, mirroring, or boundary duplication are used. Channels with entire columns of zeros or clearly invalid channels can be temporarily removed before network processing and restored using the original channel number after reconstruction to ensure the output matrix remains consistent with the original geometric relationship. For example, the sliding step size is denoted as... The overlap ratios are respectively and When the boundary is not a complete block, mirroring or boundary copying can be used. For scenarios that need to maintain zero-value bad sector indicators, zero filling can be used while simultaneously inputting a valid channel mask.

[0036] This invention employs rectangular blocks with a longer time direction and a relatively shorter channel direction, instead of the square blocks commonly used in traditional image processing, for the following reasons: The microseismic time-channel matrix is ​​not a typical natural image; the physical meaning, sampling density, correlation scale, and event continuity of the time and channel directions differ. The first-arrival phase axis typically requires a long time context before and after the first arrival, while the channel direction mainly exhibits finite neighborhood continuity. Using square blocks would treat time sampling and spatial channels as equal-scale dimensions, easily disrupting the first-arrival slope, channel delay, and local coherence structure. Rectangular blocks, on the other hand, can cover a sufficiently long time context before and after the P-wave first motion, while limiting the channel direction range and reducing interference from distant, uncorrelated channels on local denoising. For borehole geophones and small-channel arrays, rectangular blocks can preserve depth-direction continuity without expanding spurious channels. For sparse arrays in the ground or mines, rectangular blocks can form local channel groups based on nearest-neighbor stations. For long-channel DAS arrays, rectangular blocks can slide along the fiber direction while maintaining local coherence. This design helps to preserve the initial polarity, relative amplitude, and transchannel continuity of the P-wave.

[0037] S32: Based on the overlapping addition method, the denoised waveforms of all sliding window data blocks are fused to obtain the microseismic denoised data corresponding to the target seismic data matrix.

[0038] Among them, the overlapping addition method allows the overlapping areas between adjacent blocks to be merged through weighted smoothing, thereby eliminating boundary seams while ensuring that the waveform amplitude and coordinates are completely accurate.

[0039] Specifically, the seams are eliminated by overlapping and merging the prediction results of multiple blocks, that is: , Where D is the complete denoising matrix; This represents the position of the matrix corresponding to the b-th sliding window data block, which is also the position where the b-th sliding window data block is placed back into the complete matrix; The fusion weights can be average weights, Hann window weights, or channel weights. Or the product of the above weights; For the first Denoising results corresponding to each sliding window data block; For small positive numbers, usually take For example, the average weighting is... Suitable for simple implementations. The Hann window weight is... This can reduce block boundary stitching artifacts.

[0040] The channel weights are determined based on the following method: , , wherein, is the channel weight corresponding to the i-th trace of microseismic data; is the minimum continuous weight; is the anomaly threshold; , , , , , , are weight coefficients; is the signal-to-noise ratio of the i-th trace of microseismic data; is the peak amplitude of the i-th channel; is the short-term energy of the i-th trace of microseismic data; is the dominant frequency band energy proportion of the i-th trace of microseismic data; is the inter-trace correlation of the i-th trace of microseismic data; is the geometric reliability of the i-th trace of microseismic data; is the normalized penalty term for the deviation degree; for , when x ≤ 0, the output is 0; when 0 < x < 1, the output is x; when x ≥ 1, the output is 1; i=1,2,3,…,N, where N is the total number of channels.

[0041] wherein, the denoising network model includes an input layer, an encoder module, a denoising module, and a decoder module connected in sequence. The input layer is configured to receive the target seismic data matrix, the encoder module is configured to perform feature extraction and downsampling processing on the target seismic data matrix, the denoising module is configured to perform denoising processing on the output result of the encoder module, and the decoder module is configured to perform upsampling processing on the output result of the denoising module and output the upsampling result.

[0042] the input dimension of the input layer is × × ; is the number of sampling points; is the number of channels; is the number of input features; wherein, the input features at least include the channel weight.

[0043] the encoder module includes a first convolution module, a first downsampling layer, a second convolution module, a second downsampling layer, and a third convolution module connected in sequence. The first convolution module is used to extract local first-break mutations and noise textures, and may adopt 3 3. Further, after convolution, batch normalized (BN) layers, ReLU activation function layers, or GELU activation function layers can be connected. The first downsampling layer is used to expand the temporal-channel receptive field, and can be a convolution with a stride of 2 to reduce the feature map size to half its original size. The second convolutional module is used to extract continuous in-phase axes across channels, and can use 3. 3. Convolution. The second downsampling layer is used to obtain event-level context and can be a convolution with a stride of 2 to reduce the feature map size to 1 / 4 of its original size. The third convolution module is used to form deep coherent features and can use 3. 3. Convolution.

[0044] The decoder module consists of a first upsampling layer, a second upsampling layer, and an output layer connected in sequence. The first and second upsampling layers can be implemented using deconvolution or a combination of bilinear upsampling and convolutional modules. These two upsampling layers correspond to the two downsampling modules in the encoder module. The output layer can use a 1×1 convolution, ultimately outputting denoised microseismic data (i.e., denoised sliding window data blocks). Essentially, the decoder restores the fused feature map to the size of the input features.

[0045] The denoising module includes a strong denoising module, a weak denoising module, and a gated fusion module. The strong denoising module comprises, in sequence, a first fully connected layer, a LeakyReLU activation function layer, a second fully connected layer, a LeakyReLU activation function layer, and a first self-attention layer. The weak denoising module comprises, in sequence, a third fully connected layer, a LeakyReLU activation function layer, a fourth fully connected layer, a LeakyReLU activation function layer, and a second self-attention layer. The strong denoising module has a large receptive field, strong regularization, and a high mask ratio to suppress random noise, mechanical noise, and abnormal impulses; its mask ratio can be set to 0.4-0.7. The weak denoising module has a small receptive field and a low mask ratio to preserve weak first arrivals, polarity, and relative amplitude; its mask ratio can be set to 0.1-0.3. The strong denoising module processes the output of the encoder module to obtain a first denoising result, and the weak denoising module processes the first denoising result to obtain a second denoising result. The gating fusion module calculates a denoised feature map based on the first and second denoising results. This includes concatenating the output of the strong denoising module (i.e., the first denoising result), the output of the weak denoising module (i.e., the second denoising result), and the original input features (i.e., the features output by the third convolutional module). This concatenation is then processed by a convolutional layer, and the output is constrained to between 0 and 1 using the sigmoid function. Finally, the gating weight A at each position is obtained, i.e.: The fusion output result is obtained based on the gating weights, i.e.: F = A· +(1-A)· This method, which adaptively fuses local signal-to-noise ratio and signal correlation, balances denoising effectiveness with signal fidelity.

[0046] For example, the strong denoising output and the weak denoising output are denoted as zk and dk, respectively:

[0047] in This indicates a strong noise reduction module. This indicates a weak noise reduction module. and These are network parameters. Weak denoising output is preferred during inference. As a final candidate block. In blocks with particularly strong random noise, they can also be evaluated based on quality in the strong denoising output. With weak noise reduction output Weighted combinations are performed between them.

[0048] The denoising module can employ self-supervised training to train the network. Self-supervised training does not require paired clean labels. Actual noisy blocks and their perturbations, masking, or adjacent overlapping blocks can be used as inputs and consistency constraints, enabling the network to learn effective signals from the recurring multichannel coherent structures within the data.

[0049] In self-supervised training, the mask It is not the noise itself, but a binary or continuous weight matrix used to mark which locations are occluded, disturbed, or used as a reconstruction supervision target. Indicates the first The sampling point, the first The location of the road was affected by the reconstruction losses. This indicates that the location is not included in this loss or has a low weight. Noise is data containing noise. Random perturbations, mechanical interference, abnormal impulses, or coherent noise in the training process. The mask is only used to construct sampling markers for the self-supervised training task. If a continuous mask is used, then... This indicates the weight at which this position participates in the loss.

[0050] mask It can be generated in the following three ways: (1) Random point mask, in block According to proportion Randomly select elements: , in, It can be set to 0.05 to 0.20. For low signal-to-noise ratio and weak event scenarios, it can be set to 0.05 to 0.10. For strong random noise, it can be set to 0.10 to 0.20.

[0051] (2) Random short strip mask: Short strips are selected for masking along the time direction or channel direction to simulate local abnormal noise, short-term missing data, or local bad channels. Time direction strip length. The recommended sampling points are 4 to 16, with a strip width in the interchannel direction. It is recommended to have 1 to 3 dishes.

[0052] , (3) Noise candidate region mask: Noise candidate regions are identified based on local energy, residual intensity, low correlation, or P-wavefront noise window. , in, For The local energy centered on the point can be taken as The mean square amplitude within the small window. It represents the 90th percentile of the local energy of this block. For the first Correlation between different paths. A value of 0.3 to 0.5 is recommended.

[0053] Noisy blocks after disturbance or masking From the original noisy block Mask and replacement value matrix constitute: , in, This represents element-wise multiplication. For any two matrices of the same size... and ,have . Zero-value replacement, neighborhood mean replacement, Gaussian perturbation replacement, or adjacent channel interpolation replacement can be used: , In low signal-to-noise ratio microseismic studies, neighborhood mean replacement combined with random short strip masks is used to avoid introducing excessively strong artificial boundaries through zero-value replacement. When the goal is to enhance the model's robustness to anomalous impulse noise, Gaussian perturbation or local outlier replacement can be used.

[0054] The loss function of the denoising network model is: , , , , , , , in, , , , These are the weighting coefficients. For mask reconstruction loss, For inter-channel coherence loss, To ensure the accuracy of the initial candidate area, This is the residual constraint loss; This is the mask matrix corresponding to the b-th sliding window data block. The nonlinear mapping function representing the denoising network, This indicates element-wise multiplication. This is the b-th sliding window data block after being masked. To introduce channel weights, the b-th sliding window data block contains the original noisy data. To prevent small constants with a denominator of zero, This is the mask replacement value matrix corresponding to the b-th sliding window data block; For the set of neighborhood channel pairs, Let i be the total number of channel pairs within the set, and let i and j represent adjacent channel pairs, where i ≠ j. The inter-channel coherence coefficient between the i-th and j-th channels after denoising. The inter-channel coherence coefficients between the i-th and j-th channels of the original noisy data; Let p be the initial arrival pick mask for the b-th sliding window data block, where p represents the pick. This refers to the b-th denoised sliding window data block output by the denoising network model. This refers to the noise residual block corresponding to the b-th sliding window data block output by the denoising network model. , These are the weighting coefficients for gradient regularization in the time direction. These are the weighting coefficients for inter-channel directional gradient regularization. , These are the vertical and horizontal difference operators, respectively. For example, , The value ranges from 0.05 to 0.2. , , , The value can be 1.0, 0.2-0.5, 0.5-1, or 0.05-0.2, increasing when the initial value is weak and the damage is extremely easy. Improvement when random noise is strong and Multi-channel, high-continuity in-well arrays or DAS can improve .

[0055] S4: Perform event windowing on the microseismic denoising data to obtain the event-level waveform matrix.

[0056] The event-level waveform matrix refers to the matrix data obtained after truncating the microseismic denoising data through an event window. This event-level waveform matrix includes at least the first arrival times of the P-wave and S-wave, representing a portion of the microseismic denoising data. By truncating, irrelevant waveform data can be further filtered out, thereby improving the picking accuracy and efficiency.

[0057] For an event window, the trigger times of its left and right endpoints can be obtained through an event catalog, manually labeled results, STA / LTA triggers, energy detection, or existing event detection models. For example, an event window can be obtained in the following ways: M4-1: Obtain the theoretical arrival time of the P-wave and the theoretical arrival time of the S-wave for each channel, including: For each channel, first calculate the distance from the seismic source to the receiving point: , in, The coordinates of the receiving point for each channel, Here are the coarse coordinates for locating the earthquake source 's'. Then, based on these distances, the theoretical arrival time of the P-wave is calculated: , The P-wave velocity; the theoretical arrival time of the S-wave is calculated based on the above distance: , This represents the S-wave velocity.

[0058] M4-2: The minimum theoretical arrival time of the P-wave in all channels is taken as the target value of the P-wave, and the maximum theoretical arrival time of the S-wave in all channels is taken as the target value of the S-wave.

[0059] After obtaining the theoretical arrival times of the P-wave for all channels, the minimum value among them is taken as the target value of the P-wave (i.e., the P-wave trigger time). After obtaining the theoretical arrival times of the S-wave for all channels, the maximum value among them is taken as the target value of the S-wave (i.e., the trigger time of the S-wave), that is: .

[0060] M4-3: The start time of the event window is determined based on the preset pre-buffer and the target value of the P wave, and the end time of the event window is determined based on the target value of the S wave and the preset post-buffer.

[0061] The preset pre-buffer and preset post-buffer can be preset based on experience, and the preset pre-buffer is used to ensure weak initialization integrity. The event window start time is... , Preset pre-buffer. Event window end time: , This is a preset back buffer. For example, and The optimal parameters can be determined based on the sampling rate, PS wave time difference, target event duration, noise level, and computational resources. For training samples, it is preferable that both the first arrival of the P wave and the first arrival of the S wave are within the intercept window.

[0062] S5: Obtain a binary mask image based on the event-level waveform matrix and the pre-trained semantic segmentation network.

[0063] In the binary mask image, the horizontal axis represents the channel number, the vertical axis represents the sampling time, and the matrix element value is 0 or 1. 0 indicates that the point belongs to the PS wave interval, and 1 indicates that the point does not belong to the PS wave interval.

[0064] The semantic segmentation network can be implemented based on the U-Net network architecture, which includes an encoder and a decoder. Skip connections can be set between the encoder and decoder. The encoder downsamples step-by-step to extract multi-scale features, and the decoder upsamples step-by-step to restore spatial resolution. Skip connections are used to fuse shallow localization information and deep semantic information. For example, this invention makes further improvements to the traditional U-Net network: (1) Adding a coordinate feature fusion branch, inputting the waveform amplitude channel into the waveform encoding branch, and... , The coordinate channel input is a lightweight coordinate encoding branch. After the features are extracted by the convolutional layer, they are spliced ​​or fused at different scales of the encoder. (2) Coordinate gating attention is added. Spatial weight map is generated using coordinate features. The skip connection features of the encoder or the upsampled features of the decoder are weighted to enhance the network response to the reasonable first arrival time range and cross-channel continuous trend region, and suppress the noise region far away from the first arrival boundary. The semantic segmentation network can also include residual connections. Residual connections are used to improve gradient transfer during the training process of deep networks and reduce feature degradation.

[0065] The original event-level waveform matrix size is ,in The number of time sampling points, For the number of channels, when Smaller than the expected width of the semantic segmentation network At this time, linear interpolation can be performed on adjacent channels in the channel direction to extend the width to... It is also possible to replicate each waveform along the channel direction, so that multiple adjacent matrix columns correspond to the same original detector channel. If... Highly in line with network expectations If there is an inconsistency, resampling can be performed in the time direction to adjust the height. The temporal sampling order and channel spatial arrangement order remain unchanged during the reconstruction process.

[0066] Copying, interpolation, resampling, or padding should meet the following constraints: First, maintain the sequential order of time sampling directions. Second, maintain the sequential order of detector channel numbers, well depths, or array spatial orientations. Third, if scaling or resampling is performed on the time direction, the tag mask and prediction results should be transformed using the same scale and mapped back to the original time coordinates at the initial arrival time of the output according to the sampling interval or scaling ratio. Fourth, if copying or interpolating the channel direction, the extended channel is only used to adjust the size ratio of the two-dimensional matrix and is not used as a new physical detector channel.

[0067] For example, symmetrical padding or boundary duplication can be used to position the original valid waveform region in the center of the reconstruction matrix to reduce boundary effects. For instance, for well monitoring data, when the number of time sampling points is typically greater than the number of channels, the waveform matrix can be reconstructed into a square matrix by duplicating channel data. For surface monitoring data, a square matrix can also be constructed using resampling.

[0068] After data augmentation, amplitude normalization of the time-channel seismic waveform matrix is ​​required. The normalization range can be [-1,1], [0,1], or other preset ranges. Normalizing the waveform amplitude to [-1,1] reduces amplitude differences between different events and channels, improving model training stability.

[0069] To train the semantic segmentation network, a binary mask with labels needs to be constructed for the training samples. For each training event, the first arrival times of the P-wave and S-wave are obtained for each trace based on manual or high-confidence picking results. Subsequently, in a mask matrix with the same size as the time-trace seismic waveform matrix, the region between the first arrival times of the P-wave and S-wave is labeled as the target region, and the remaining regions are labeled as background regions. The label mask is a binary mask.

[0070] in, Indicates the first in the label mask The sampling location at time, the first The matrix element values ​​at each channel number position. Indicates the first The index of the sampling point corresponding to the first arrival time of the P-wave of the channel. Indicates the first The sampling point index corresponding to the first arrival time of the S-wave. The target region can be represented by matrix elements with a target value of 0, and the background region can be represented by matrix elements with a background value of 1. The target region refers to the area between the first arrival times of the P-wave and the first arrival times of the S-wave.

[0071] After model training is complete, the same data preprocessing procedure as in the training phase is performed on the microseismic events to be processed, and an event-level waveform matrix is ​​obtained based on the preprocessed data. This event-level waveform matrix is ​​then input into the trained semantic segmentation network, which outputs a prediction probability map. Each matrix element in the prediction probability map represents the probability that the location belongs to the target region. The prediction probability map is then thresholded to obtain a binary prediction mask. The threshold can be set to 0.5 or adaptively determined based on the validation set results.

[0072] The black-and-white image or binary mask output by the semantic segmentation model has a clear physical meaning. Let the output binary mask be B∈{0,1}^{H×W}, where H is the image height, corresponding to the temporal sampling direction; W is the image width, corresponding to the channel direction. B(r,c)=0 indicates that the pixel belongs to the target region between the first arrival of the P-wave and the first arrival of the S-wave, i.e., a white pixel; B(r,c)=1 indicates the background region, noise region, or non-target region, i.e., a black pixel. r is the image row number, corresponding to the temporal sampling point, and c is the image column number, corresponding to the inter-channel position. The upper boundary of the white region corresponds to the first arrival of the P-wave, and the lower boundary of the white region corresponds to the first arrival of the S-wave. Therefore, the model output image is not the final picking time, but an intermediate physical mask used to recover the boundaries of the first arrival of each P / S wave.

[0073] If the model output is a probability graph P(r,c)∈[0,1], then it first passes through a threshold. Binarization:

[0074] in, The target region probability threshold is preferably set to 0.5. If the model output is a black and white image with a grayscale range of 0 to 255, then a grayscale threshold of 127 can be used, meaning that pixels with a grayscale value greater than or equal to 127 are considered white target pixels. This threshold can also be adjusted based on the P / S arrival time error, the success rate, and the anomaly picking ratio of the validation set.

[0075] S6: Calculate the first arrival times of the P-wave and S-wave based on the binary mask image, including: For each channel, perform the following operation: S61: Calculate the effective length of each column under the channel, and take the maximum effective length as the target longest chain of the column. If the length of the target longest chain is not 0, then take it as an element in the set of effective columns corresponding to the channel; wherein, the starting row number of the target longest chain corresponds to the arrival time of the P wave, and the ending row number corresponds to the arrival time of the S wave.

[0076] This involves calculating the effective length of each column under the channel, using the maximum effective length as the target longest chain for that column, including: For the i-th sub-image Column-by-column scanning. For the c-th column in the sub-image, search for continuous segments of white pixels from top to bottom. To avoid the PS white target region output by the model being cut off by a few holes, short breaks, or isolated black dots under low signal-to-noise ratio conditions, this invention allows no more than [number missing] white pixels within the same white segment. A black pixel is interrupted. The maximum allowed number of black dots is a post-semantic segmentation fault tolerance parameter, not a seismic physics parameter. For example, for the i-th sub-image in the c-th column, the q-th candidate white segment is denoted as: , in, Let q be the starting line number of the q-th candidate white link segment. To terminate the line number, This represents the number of white pixels within that segment. This represents the number of black discontinuous pixels within this segment. The value can be determined based on the predicted image height H and the time interval corresponding to the image row number. Maximum allowable fracture time for repair And the determination of the validation set picking error. For example, one possible value is: , Where α is the allowable fracture ratio, recommended to be between 0.02 and 0.08. Another way to determine its value is:

[0077] in, The maximum allowable fracture length for repair. This is used to predict the time interval between two adjacent rows of an image. For example, for a predicted image with H=880, A value of 18 to 70 is acceptable; in embodiments with low signal-to-noise ratios and where the white area is prone to breakage, a value of [missing value] is acceptable. =60.

[0078] The effective length of a candidate segment is defined as: , Select the candidate white chain segment with the largest effective length as the effective PS region for this column: , in, The row number of the P-wave boundary identified in the c-th column of the i-th channel. This is the row number of the S-wave boundary identified in the c-th column of the i-th channel. The upper boundary of the longest white chain corresponds to the first arrival position of the P-wave, and the lower boundary corresponds to the first arrival position of the S-wave.

[0079] Using the longest white chain instead of simply taking the first white point can avoid interference from isolated noisy white points or local missegmentation on P / S first arrival extraction.

[0080] If the image height is the same as the original number of sampling points, then It is equal to the sampling interval Δt. For example, when Δt = 0.5 ms and a short break of about 30 ms is allowed to be repaired, =60. If the value is too small, the real white area of ​​PS will be cut off by short black dots, resulting in the longest white chain becoming shorter and the P / S boundary becoming unstable; When the value is too large, it may connect independent erroneous white areas into the same white chain, causing the P wave boundary to be earlier and the S wave boundary to be later. Therefore, the search can be performed in {10,20,30,40,50,60,80,100}, and the average absolute error of P / S arrival time, the excellent rate, and the anomaly picking ratio can be used as selection criteria.

[0081] After obtaining the target longest chain for each column, if the length of the target longest chain is not 0, then it is used as an element in the set of valid columns corresponding to that channel.

[0082] S62: Calculate the first arrival sampling points corresponding to the P wave and S wave based on the set of effective columns.

[0083] We statistically analyze the effective white chain boundaries of all columns in the i-th sub-image, and let the set of effective columns be: , The first arrival sampling points of the i-th P-wave and S-wave can be determined using the mean or median:

[0084]

[0085]

[0086] .

[0087] The mean is suitable for cases with smooth prediction boundaries and small inter-column fluctuations; the median is suitable for cases with isolated outliers or local segmentation errors. For example, first calculate... and The median and median absolute deviation (MAD) are used to calculate the average of the remaining columns. Outlier columns deviating more than 2 × MAD from the median are removed. Columns without valid white links are not included in the mean calculation but are removed first. If the proportion of valid columns is high... Below the threshold This channel is marked as a low-confidence channel, that is: , in, Let be the width of the i-th sub-image. It can be set to 0.3 to 0.5. This processing is more robust than using invalid columns as 0 in the mean, and can prevent the P / S first arrival error from being pulled to near the start of the event window when the model fails to predict valid white chains.

[0088] S63: Calculate the P-wave arrival time and S-wave arrival time corresponding to the channel based on the first arrival sampling point and the event window.

[0089] If the event window start time is Given a sampling interval of Δt and a predicted image height H that is the same as the number of sampling points T in the original event window, the physical arrival time corresponding to the image row number is: , .

[0090] If the predicted image height H is inconsistent with the original event window sampling point number T, then first map the image row number back to the original sampling point number: , , , in, The physical arrival time of the i-th P-wave is... The physical arrival time of the i-th S-wave; and These are the row numbers obtained from image post-processing. and Number the original sampling points after mapping.

[0091] For example, the calculation method for the fusion of three components is as follows: For the same event and the same detector, if the E, N, and Z components obtain the P / S initial arrival sampling points respectively, then average fusion can be used: , ; Confidence-weighted fusion can also be used: , Where 'a' represents a component. The post-processing confidence score for the i-th channel and the a-th component can be determined by the proportion of effective columns, the average length of the longest white chain, the boundary variance, the model probability mean, and the denoising quality. Determined jointly; ε is a small constant to prevent the denominator from being zero, and it is recommended to take . When a component has no valid white chains or the proportion of valid columns is lower than [a certain percentage]... The components are not included in the average or are given a smaller weight.

[0092] Compared with existing technologies, this embodiment provides a microseismic data processing method under low signal-to-noise ratio (SNR) conditions. Based on a denoising network, it effectively suppresses background random noise and mechanical interference, while preserving waveform morphology, polarity, and amplitude characteristics to the greatest extent possible. This significantly improves the signal identifiability under low SNR conditions, thereby further reducing the missed detection rate of weak microseismic events. Furthermore, the denoising network used in this invention employs a cascaded structure combining strong and weak denoisers. The strong denoiser suppresses random noise, local anomalous amplitudes, and incoherent interference, while the weak denoiser recovers weak first arrivals, phase abrupt changes, and relative amplitudes based on the strong denoising results. This two-stage structure allows the model to suppress strong noise while avoiding the complete removal of low-amplitude first arrivals as noise, further enhancing the identifiability of microseismic signals. This invention transforms first-arrival picking into a semantic segmentation task of the effective region of PS waves. Utilizing the modeling capabilities of deep learning networks for two-dimensional time-channel features, it fully captures the inter-channel continuity and time-domain morphological characteristics of microseismic phase axes. Compared to traditional first-arrival picking methods, the proposed method can more accurately distinguish between effective signals and discrete noise, significantly reducing the probability of false picking under strong interference environments and improving the overall accuracy of first-arrival identification. Addressing the shortcomings of segmentation results prone to holes, breaks, and isolated false positive regions under low signal-to-noise ratio conditions, post-processing logic such as longest white chain identification, discontinuity tolerance, and multi-column statistical aggregation can automatically repair breaks in the effective region, filter isolated noise interference, and avoid the impact of local segmentation errors on the first-arrival boundary. The output P and S wave first-arrival times are more stable, reducing the cost of manual verification and correction. From the input of raw multichannel data to the output of P and S wave arrival times, the entire process can be automated. After the model training is completed, the inference stage does not require manual annotation of each channel and parameter tuning of each event. It has high processing efficiency and can be adapted to batch processing scenarios of massive continuous data such as mine microseismic monitoring, which greatly reduces the cost of manual processing and improves the timeliness of data processing.

[0093] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing microseismic data under low signal-to-noise ratio conditions, characterized in that, The method includes: Acquire multichannel microseismic data and preprocess the multichannel microseismic data; Construct a target seismic data matrix based on preprocessed multichannel microseismic data; The target seismic data matrix is ​​processed using a denoising network model to obtain denoised microseismic data; The event-level waveform matrix is ​​obtained by performing event windowing on the microseismic denoising data; A binary mask image is obtained based on the event-level waveform matrix and the pre-trained semantic segmentation network; The first arrival times of the P-wave and S-wave are calculated based on the binary mask image.

2. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 1, characterized in that, The denoising network model includes an input layer, an encoder module, a denoising module, and a decoder module connected in sequence. The input layer is used to receive the target seismic data matrix; the encoder module is used to extract features and downsample the target seismic data matrix; the denoising module is used to denoise the output of the encoder module; and the decoder module is used to upsample the output of the denoising module and output the upsampled result.

3. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 2, characterized in that, The encoder module includes a first convolution module, a first downsampling layer, a second convolution module, a second downsampling layer, and a third convolution module connected in sequence. The decoder module includes a first upsampling layer and a second upsampling layer connected in sequence; The denoising module includes a strong denoising module, a weak denoising module, and a gated fusion module; wherein, the strong denoising module processes the output of the encoder module to obtain a first denoising result, the weak denoising module processes the first denoising result to obtain a second denoising result; and the gated fusion module calculates a denoised feature map based on the first denoising result and the second denoising result.

4. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 3, characterized in that, The strong denoising module includes a first fully connected layer, a LeakyReLU activation function layer, a second fully connected layer, a LeakyReLU activation function layer, and a first self-attention layer connected in sequence. The weak denoising module includes a third fully connected layer, a LeakyReLU activation function layer, a fourth fully connected layer, a LeakyReLU activation function layer, and a second self-attention layer connected in sequence.

5. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 4, characterized in that, The process of processing the target seismic data matrix using a denoising network model to obtain denoised microseismic data includes: The target seismic data matrix is ​​subjected to sliding window processing, which divides the matrix into multiple different sliding window data blocks. Each sliding window data block is input into a denoising network model to obtain a denoised waveform. The overlapping addition method is used to fuse the denoised waveforms of all sliding window data blocks, thereby obtaining the microseismic denoised data corresponding to the target seismic data matrix.

6. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 5, characterized in that, The input dimension of the input layer is × × ; This represents the number of sampling points; Number of channels; The number of input features; wherein, the input features include at least channel weights.

7. The microseismic data processing method under low signal-to-noise ratio conditions according to claim 6, characterized in that, The channel weights are determined based on the following method: ， ， where, is the channel weight corresponding to the i-th trace of microseismic data; is the minimum continuous weight; is the anomaly threshold; , , , , , , are weight coefficients; is the signal-to-noise ratio of the i-th trace of microseismic data; is the peak amplitude of the i-th channel; is the short-term energy of the i-th trace of microseismic data; is the dominant frequency band energy ratio of the i-th trace of microseismic data; is the inter-trace correlation of the i-th trace of microseismic data; is the geometric reliability of the i-th trace of microseismic data; is the normalized penalty term for the deviation degree; for , when x ≤ 0, the output is 0; when 0 < x < 1, the output is x; when x ≥ 1, the output is 1; i = 1, 2, 3, ..., N, where N is the total number of channels.

8. A method for processing microseismic data under low signal-to-noise ratio conditions according to claim 7, characterized in that, The loss function of the denoising network model is: ， ， ， ， ， ， ， in, , , , These are the weighting coefficients. For mask reconstruction loss, For inter-channel coherence loss, To ensure the accuracy of the initial candidate area, This is the residual constraint loss; This is the mask matrix corresponding to the b-th sliding window data block. The nonlinear mapping function representing the denoising network, This indicates element-wise multiplication. This is the b-th sliding window data block after being masked. To introduce channel weights, the b-th sliding window data block contains the original noisy data. To prevent small constants with a denominator of zero, This is the mask replacement value matrix corresponding to the b-th sliding window data block; For the set of neighborhood channel pairs, Let i be the total number of channel pairs within the set, and let i and j represent adjacent channel pairs, where i ≠ j. The inter-channel coherence coefficient between the i-th and j-th channels after denoising. The inter-channel coherence coefficients between the i-th and j-th channels of the original noisy data; Let p be the initial arrival pick mask for the b-th sliding window data block, where p represents the pick. This refers to the b-th sliding window data block output by the denoising network model; This refers to the noise residual block corresponding to the b-th sliding window data block output by the denoising network model. , These are the weighting coefficients for gradient regularization in the time direction. These are the weighting coefficients for inter-channel directional gradient regularization. , These are the difference operators for the vertical and horizontal directions, respectively.

9. A method for processing microseismic data under low signal-to-noise ratio conditions according to claim 1, characterized in that, The event window is obtained in the following way: Obtain the theoretical arrival time of the P-wave and the theoretical arrival time of the S-wave for each channel; The minimum theoretical arrival time of the P-wave in all channels is taken as the target value of the P-wave, and the maximum theoretical arrival time of the S-wave in all channels is taken as the target value of the S-wave. The start time of the event window is determined based on the preset pre-buffer and the target value of the P-wave, and the end time of the event window is determined based on the target value of the S-wave and the preset post-buffer.

10. A method for processing microseismic data under low signal-to-noise ratio conditions according to claim 9, characterized in that, In the binary mask image, the horizontal axis represents the channel number, the vertical axis represents the sampling time, and the matrix element value is 0 or 1. 0 indicates that the point belongs to the PS wave interval, and 1 indicates that the point does not belong to the PS wave interval. The calculation of the first arrival times of the P-wave and S-wave based on the binary mask image includes: For each channel, perform the following operation: Calculate the effective length of each column under the channel, and take the maximum effective length as the target longest chain of the column. If the length of the target longest chain is not 0, then take it as an element in the set of effective columns corresponding to the channel; wherein, the starting row number of the target longest chain corresponds to the first arrival time of the P wave, and the ending row number corresponds to the first arrival time of the S wave. The first arrival sampling points corresponding to the P-wave and S-wave are calculated based on the effective column set. The P-wave first arrival time and S-wave first arrival time corresponding to this channel are calculated based on the first arrival sampling point and the event window.

Images