Microseismic event first arrival picking method based on context-aware multi-task network

Abstract

The application discloses a microseismic event first arrival picking method based on a context-aware multi-task network, which comprises the following steps: step 1, loading and preprocessing of original microseismic data; step 2, bad channel identification and repair of microseismic events; step 3, construction of a context-aware multi-task network model; step 4, design of an adaptive multi-task loss function; step 5, training of the context-aware multi-task network model; step 6, inference on newly collected microseismic waveforms, and output of first arrival time prediction values and corresponding uncertainties of each sensor; step 7, source positioning based on joint inversion, wherein the inverse of the uncertainty is used as the weight, and a differential evolution algorithm is used to jointly invert the three-dimensional coordinates of the source and the time of the earthquake occurrence. The application realizes the intelligentization of the whole process from data repair to reliable picking, and significantly improves the picking accuracy, the automation level and the result reliability.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and signal processing technology, specifically relating to an intelligent processing method for microseismic monitoring data, and in particular, a method for picking the first arrival of microseismic events based on a context-aware multi-task network. Background Technology

[0002] Microseismic monitoring technology plays an irreplaceable role in mine safety as a key means of early warning of dynamic disasters and evaluation of rock mass stability. With increasing mining depth and more complex geological conditions, the data collected by microseismic monitoring systems exhibits a trend towards massive data volume and low signal-to-noise ratio, posing a severe challenge to traditional processing methods. Currently, the first-arrival time acquisition technology for microseismic events mainly faces the following technical bottlenecks:

[0003] (1) In traditional methods, the threshold method is easily affected by subjective factors, and the false alarm rate and missed detection rate are high in low signal-to-noise ratio environments, and the ability to identify weak signals is insufficient; the time-frequency analysis method is computationally complex, has poor adaptability to non-stationary noise, and has limited real-time processing capability.

[0004] (2) Deep learning methods provide a new approach to first arrival picking, but existing models based on convolutional neural networks or recurrent neural networks mostly adopt a serial processing flow of "detecting events first and then picking first arrivals", which leads to loss of feature information and error accumulation effect. At the same time, these methods usually only process single-channel waveforms and fail to fully utilize the spatiotemporal context information between multiple channels (such as detector spatial layout and wave field propagation law), which is not robust enough to bad channels and noise, and has limited generalization ability under complex geological conditions.

[0005] In summary, the existing methods have the following problems: (1) In low signal-to-noise ratio environments, the initial arrival time of weak signals is not accurately determined and the stability is poor; (2) The serial processing flow causes error transmission and accumulation, affecting the final positioning accuracy; (3) The spatiotemporal evolution law and physical constraints of the wave field of microseismic events are not fully utilized, and the fusion perception and joint optimization of multi-dimensional contextual information are not achieved.

[0006] Therefore, developing an intelligent acquisition technology that can achieve end-to-end collaborative signal sensing and first arrival acquisition, deeply integrate temporal and spatial context, and embed geophysical prior knowledge has become a key breakthrough direction for improving the automation and accuracy of microseismic monitoring. Summary of the Invention

[0007] The purpose of this invention is to propose a microseismic event first arrival picking method based on a context-aware multi-task network, so as to solve the problem of low accuracy of microseismic event first arrival picking in low signal-to-noise ratio environments in existing mine microseismic monitoring technologies.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A method for picking the first arrival of microseismic events based on a context-aware multi-task network includes the following steps: Step 1. Loading and preprocessing of raw microseismic data; Step 2. Based on the data obtained in Step 1, identify and repair bad tracks caused by microseismic events; Step 3. Construction of a context-aware multi-task network model, including the following sub-steps: Step 31, Constructing a waveform encoder: Using a multi-scale one-dimensional convolution and bidirectional LSTM structure, extract the multi-time-scale temporal features of the waveform to construct a waveform encoder; Step 32, construct the detector context encoder: the detector spatial coordinates are mapped to the detector context embedding vector through a multilayer perceptron; Step 33, Constructing the context fusion module: The waveform features and sensor context are embedded to construct a gated fusion module to obtain the fused output features; Step 34, Construct multi-task output heads: Based on the fused output features, construct the first arrival picking head, the source location head, and the uncertainty estimation head respectively; Step 4. Design of adaptive multi-task loss function; Step 5. Model training and validation: Using the labels obtained in Step 2 as training samples, and employing the loss function designed in Step 4, train the context-aware multi-task network model constructed in Step 3 to obtain the trained context-aware multi-task network model. Step 6. First arrival time prediction and uncertainty output: Infer the newly acquired microseismic waveforms and output the first arrival time prediction values ​​and corresponding uncertainties for each sensor; use the context-aware multi-task network model trained in Step 5 to pick up the first arrival events of the newly acquired microseismic data. During the processing, the original normalized waveform of the microseismic event and the geophone coordinates are input. After processing, the first arrival time prediction values ​​of each sensor, the first arrival uncertainty of each sensor, and the preliminary prediction of the source location are obtained; after the first arrival prediction, a valid picking and filtering is added. If the filtering criteria are met, the predicted first arrival is considered valid. Step 7. Based on the joint inversion of the seismic source location, using the reciprocal of the uncertainty obtained in Step 6 as the weight, the differential evolution algorithm is used to jointly invert the three-dimensional coordinates of the seismic source and the time of occurrence.

[0009] Compared with the prior art, the present invention has the following technical effects: (1) This invention systematically solves the fundamental defects of error accumulation in traditional methods and lack of context in deep learning models through an innovative end-to-end multi-task fusion perception architecture. Traditional sequential processes (detection before picking) and single-feature deep learning models suffer from insufficient accuracy and robustness in complex scenarios due to stage error propagation and neglect of spatial and physical context information, respectively. The context-aware multi-task network pioneered in this invention jointly optimizes and shares features of event perception, first arrival picking, and source localization, and deeply integrates multi-dimensional context such as temporal, spatial, and physical laws, thereby avoiding error accumulation at the source and significantly improving adaptability to low signal-to-noise ratio data, bad channels, and complex wavefield environments. This technical effect reflects the advantages of the model design in step 3 of the method of this invention.

[0010] (2) This invention constructs a complete technical closed loop of intelligent enhancement throughout the entire process, fundamentally improving the automation level and reliability of microseismic monitoring results. Existing technology chains have gaps or deficiencies in data quality assurance, processing procedures, and result verification, such as crude handling of bad channels, reliance on manual processes, and a lack of physical interpretability in the results. This invention integrates a complete technical chain from intelligent perception and repair of bad channels and fully automated intelligent acquisition to the fusion of physical model verification and output uncertainty assessment. This not only achieves a high degree of automation and real-time performance in the processing flow but also, by introducing physical constraints and quantifying uncertainty, endows the processing results with physical interpretability and higher reliability, forming an enhanced closed loop from data input to reliable result output. This technical effect demonstrates the advantages of multi-task joint learning in step 4 of the method of this invention.

[0011] In summary, this invention achieves intelligent processing of the entire process from data repair to reliable picking by integrating multi-dimensional context and physical constraints through an end-to-end multi-task network, significantly improving picking accuracy, automation level and result reliability. Attached Figure Description

[0012] Figure 1 This is a flowchart of the microseismic event first arrival picking method based on a context-aware multi-task network according to the present invention. Figure 2 The raw microseismic event signals acquired by microseismic monitoring; Figure 3 A visualization of the preprocessing results of the original microseismic signals; Figure 4 The results are from model training and validation using measured microseismic data; Figure 5 Verification results for first arrival pickup of high signal-to-noise ratio microseismic events; Figure 6 The results are for the initial arrival pickup verification of low signal-to-noise ratio microseismic events. Figure 7Comparison results of source location for 5 verification events.

[0013] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0014] like Figure 1 As shown, the microseismic event first arrival picking method based on context-aware multi-task networks provided in this invention includes the following steps: Step 1. Loading and preprocessing of raw microseismic data. The main goal of this step is to load multi-channel microseismic waveform data, detector coordinates, and real event labels, and to standardize and unify the format to provide high-quality input for subsequent analysis.

[0015] Step 1 includes the following sub-steps: Step 11: Read the SGY format microseismic event waveform data file collected by the microseismic monitoring system, load the source coordinate file and detector coordinate file corresponding to the microseismic event; parse the manually calibrated first arrival time of the microseismic event from the file header of the SGY format microseismic event waveform data file, and match it with the SGY format microseismic event waveform data file through the microseismic event ID, and use it as a label and verification reference for subsequent model training.

[0016] Step 12, perform data preprocessing on the data obtained in Step 11, including: (1) Linear fitting is used to correct the baseline drift of each data point of the microseismic event. The core idea is to assume that the baseline drift is approximately a straight line (first-order trend) within the event window. By fitting this straight line and subtracting it from the original data, the "zero mean" baseline can be recovered. This step can be achieved by a known formula. (2) Based on the frequency band range of the effective signal of the microseismic event (e.g., 10-80 Hz), a bandpass filter is designed and applied to process the result of (1) in order to suppress high-frequency noise and low-frequency interference and highlight the first arrival signal of the microseismic event. (3) The results of (2) are processed by the surface uniform amplitude equalization method to compensate for the energy difference between channels caused by factors such as detector coupling difference and near-surface attenuation, so as to normalize the energy level of each channel signal. (4) Apply automatic gain control based on sliding window root mean square energy calculation to the result of (3) to make the amplitude of the shallow, middle and deep parts of the waveform relatively balanced, and further highlight the effective signal.

[0017] Step 2. Based on the data obtained in Step 1, perform microseismic event bad channel identification and intelligent repair. The main goal of this step is to automatically identify invalid or low-quality data channels (bad channels) and repair them using spatial context information, maximizing the preservation of valid information and improving data integrity and utilization.

[0018] The specific method for bad channel identification is as follows: For each microseismic event obtained in step 1, a Boolean mask is generated to identify the location of the bad channel based on the preset threshold (such as the arrival time being close to zero, energy anomalies (significantly lower than (less than one-tenth) or higher than (greater than 10 times) the average energy of all channels) and the event-by-event and channel-by-channel detection is performed based on the preset threshold (such as the arrival time being close to zero, energy anomalies (significantly lower than (less than one-tenth) or higher than (greater than 10 times) the average energy of all channels)).

[0019] Adaptive repair method for waveform and first arrival in bad channels: For the identified bad channels, the waveform and first arrival label of the bad channels are reconstructed by interpolation point by point using the inverse distance weighted average algorithm based on the spatial distance between the three-dimensional coordinates and the coordinates of the surrounding valid channels.

[0020] The formula is:

[0021] Among them, weight The formula for the inverse distance weighted average algorithm is as follows:

[0022] In the above formula: The value to be repaired for bad sectors (waveform amplitude at the initial arrival or a certain sampling time). The number of valid channels, For the first Observations of one effective channel, For the first The weights of each effective path are determined by their proximity; the closer the paths are, the greater the weight. The sum of all weights is 1. The smoothing constant is a very small positive number (e.g., 1e-6) used to prevent the smoothing effect when... = 0 will result in a division by zero error or an infinite weight. For spatial distance, the detector position of the bad channel is relative to the first... The Euclidean distance between the detector positions of each effective channel The calculation formula is:

[0023] in, The three-dimensional coordinates of the detector corresponding to the bad channel that needs to be repaired. For the first The three-dimensional coordinates of the detector corresponding to each effective channel.

[0024] First arrival tag repair: The first arrival time tag of bad channels is also filled by inverse distance weighted interpolation based on its spatial distance with the surrounding valid channels. The formula principle is the same as that of waveform repair, ensuring the spatial consistency of the tags.

[0025] Step 2 yields the microseismic event waveforms and corresponding tags for adaptive repair of bad paths and first arrival tags.

[0026] Step 3. Context-Aware Multi-Task Network Model Construction. The main goal of this step is to construct a deep neural network that includes a temporal feature extractor, a spatial feature extractor, a channel attention mechanism, a bidirectional LSTM temporal modeling unit, and a multi-task learning head.

[0027] Step 3 includes the following sub-steps: Step 31, Construct the waveform encoder A multi-scale one-dimensional convolution and bidirectional LSTM structure are used to extract multi-time-scale temporal features of the waveform to construct a waveform encoder. Specifically, the operation involves processing the input waveform... Parallel use K Different convolution kernel widths Perform convolution:

[0028] And the above multi-scale feature maps are stitched together along the channel dimension:

[0029] The multi-scale features obtained from the above concatenation are input into a bidirectional LSTM to capture forward and reverse temporal dependencies:

[0030]

[0031] in, The model has all trainable parameters, including the parameters of the forward LSTM and the parameters of the backward LSTM, which are learned through training. Let be the hidden layer dimension of the LSTM. Self-attention is applied in the temporal dimension to enhance the features of nearby keyframes:

[0032]

[0033] In the above formula, is the attention-weighted feature matrix, and is the output value; The temporal feature matrix output by the LSTM network, i.e., the input to self-attention, is a feature representation of the forward and backward temporal dependencies at each time step; Q The query matrix represents the search vector for actively querying whether relevant waveform features exist at other times in the current time step; K Labels representing the waveform characteristics at each time step; It is a normalization coefficient that scales the attention score; This means transforming the scaled score matrix into a probability distribution; The waveform feature content actually carried at each time step is the information carrier matrix of the self-attention mechanism; The learnable parameter matrix that linearly projects the input features H onto the value space; The feature dimension of the bidirectional LSTM hidden layer output is typically taken as... =256, For matrix The dimension of the feature vector at each time step, typically taking a value of [value missing]. .

[0034] Step 32, Construct the detector context encoder Detector spatial coordinates Mapped to detector context embedding vectors via a multilayer perceptron (MLP):

[0035]

[0036] in, for MLP The set of learnable parameters includes the weights / biases of fully connected layers and the scaling / offset parameters of batch normalized layers. γ,β ), etc., are parameters optimized through backpropagation during training. i For detector index. Contextual embedding dimension. By using a sensor encoder, waveform features become aware of their spatial location, thereby establishing implicit geometric constraints.

[0037] Step 33, Build the context fusion module waveform characteristics (After global pooling of temporal features) and sensor context embedding Build the gating fusion module:

[0038]

[0039] Gating mechanisms allow the network to adaptively decide whether to rely more on waveform features or spatial context in different samples, as shown in the above equation. This is the gate weight matrix; This is the gated bias vector; Waveform characteristics Gating weighting; Spatial features Gating weighting; To fuse output features, the gated fusion composite feature vector integrates waveform timing information and sensor spatial context information.

[0040] Step 34, Construct the multi-task output header Based on the fusion output features in step 33, construct the initial arrival pickup heads respectively:

[0041] In the above formula, The normalized first arrival offset of the output is multiplied by the time window length to restore it to the absolute first arrival time; To initially pick up fully connected neural network layers; For the first i The context-aware fusion feature vector corresponding to each sensor.

[0042] Constructing a seismic source location head:

[0043] in, A global aggregation (attention-weighted average) of features fused from all sensors. σ for Sigmoid Activation function To monitor the spatial range vector, The fully connected layer for earthquake source localization functions to perform linear transformation and nonlinear activation on the input features.

[0044] Constructing an uncertainty estimation head:

[0045] in, Softplus for Softplus Activation function This is the network branch for "uncertainty estimation". The above formula uses... Softplus Activation ensures that the uncertainty output is positive, reflecting the confidence level of the initial arrival pick-up.

[0046] The context-aware multi-task network model is constructed through the specific process described in step 3.

[0047] Step 4. Adaptive multi-task loss function design: Introduce learnable task weight parameters to adaptively weight and optimize the first arrival picking loss, source location loss, and uncertainty regularization loss.

[0048] The initial pickup loss is:

[0049] in, ∈{0,1} is the effective detector mask, which only accumulates the loss for the effective detector, and the error of the invalid channel is directly ignored; The number of detectors; For the first i The true first arrival time offset of each detector; For the first i The model predicts the first arrival time offset for each detector; The model predicts the first i The initial arrival uncertainty of each detector.

[0050] The loss due to focal location is:

[0051]

[0052] in, The spatial coordinates of the earthquake source; The epicenter was at the 1st d Dimensional coordinates; To predict the earthquake source in the model at the 1st d Dimensional coordinates; = To predict the deviation between the coordinates and the actual coordinates; for Huber Loss function, for residuals r Robustness measure; for Huber The threshold for loss.

[0053] Uncertainty canonical loss:

[0054] in, The model predicts the first i Initial arrival uncertainty of each detector, This is the upper limit threshold for uncertainty; This means that only the portion exceeding the upper limit threshold is penalized, and the value is 0 if it does not exceed it; this loss prevents the uncertainty estimate from increasing indefinitely.

[0055] Based on the above loss terms, and drawing on the concept of multi-task learning for homoscedastic uncertainty, learnable task weight parameters are introduced. Construct adaptive multi-task loss weights:

[0056] in, As model parameters, they are automatically optimized during training, achieving adaptive balancing of loss weights across multiple tasks without the need for manual parameter tuning.

[0057] Step 5. Model training and validation: Using the labels obtained in Step 2 as training samples, and employing the loss function designed in Step 4, train the context-aware multi-task network model constructed in Step 3 to obtain the trained context-aware multi-task network model.

[0058] The data obtained in step 2 was divided proportionally into a training set (80%) and a validation set (20%), and stratified sampling was used to ensure a balanced spatial distribution of seismic sources. During training, the hyperparameters were set as follows: batch size = 32; AdamW optimizer was selected; initial learning rate = 1. 10 -3 The number of LSTM layers is 2-4; the hidden layer dimension is... =128; Maximum number of training rounds: 500; Early stop patience: 20.

[0059] Step 6. First arrival time prediction and uncertainty output: Perform inference on the newly acquired microseismic waveforms and output the first arrival time prediction values ​​and corresponding uncertainties for each sensor.

[0060] After training, the trained context-aware multi-task network model is used to pick up the first arrival of newly acquired microseismic data. During the processing, the normalized waveform of the original microseismic event and the coordinates of the detector are input. After processing, the predicted first arrival time of each sensor, the first arrival uncertainty of each sensor, and the preliminary prediction of the source location are obtained.

[0061] After the initial arrival prediction, an effective picking filter is added, with the following selection criteria:

[0062] in, The model predicts the first i Initial arrival uncertainty of each detector, This is the threshold for the initial uncertainty of the model. For an effective detector mask to satisfy the above criteria, the predicted first arrival is considered valid.

[0063] Step 7. Based on the joint inversion of the seismic source location, using the reciprocal of the uncertainty obtained in Step 6 as the weight, the differential evolution algorithm is used to jointly invert the three-dimensional coordinates of the seismic source and the time of occurrence.

[0064] Using the coordinates of the earthquake source With the time of the earthquake Define the weighted residual objective function for the joint inversion parameters:

[0065] in, For an effective set of detectors; = Using the reciprocal of the pickup uncertainty as a weight, detectors with higher pickup accuracy contribute more to positioning. The estimated wave velocity.

[0066] The differential evolution algorithm is used to obtain the global optimization value of the objective function, and the microseismic event location result is obtained with the reciprocal of the initial picking uncertainty as the weight.

[0067] As can be seen from the above, the key inventive points of this invention are as follows: 1. This paper proposes an integrated multi-task network structure that simultaneously performs microseismic first arrival pickup, source localization, and uncertainty estimation. It adaptively fuses waveform temporal features with the sensor's 3D coordinate context through a gating mechanism, achieving high-precision pickup and localization with context awareness. Compared to existing methods, this invention is the first to collaboratively optimize sensor spatial geometric information, waveform temporal context, and multi-task outputs within a unified network structure through gating, significantly improving first arrival pickup accuracy and localization stability under low signal-to-noise ratio conditions.

[0068] 2. An adaptive multi-task optimization mechanism driven by prediction uncertainty is constructed: uncertainty is used to construct the regularized loss for initial arrival picking, and a learnable homoscedasticity uncertainty parameter is used to automatically adjust the weights of each task. In the source location stage, prediction uncertainty is used as a weight for weighted joint inversion, forming a closed-loop uncertainty utilization link from training to inference. Unlike traditional methods that simply add hard weights and multiple tasks, this invention uses uncertainty as a unified metric, running through both the loss design and location inversion stages, achieving adaptive adjustment of task weights and quantifiable evaluation of result quality.

[0069] The technical effects achieved by this invention are significant, specifically reflected in the following aspects: This invention achieves significant picking accuracy on actual coal mine microseismic monitoring data, effectively overcoming the limitation of traditional methods in significantly decreasing picking accuracy for low signal-to-noise ratio events; (2) The bad path adaptive processing proposed in this invention effectively improves the availability of actual engineering data and reduces the workload of manual screening of bad paths; (3) This invention incorporates first arrival picking and source location into a unified multi-task learning framework, and introduces a joint objective function based on prediction uncertainty weighting in the inversion stage. The results show that this invention can effectively utilize the first arrival picking results and their uncertainty information to achieve high-precision source location; (4) This invention introduces learnable task weight parameters in the multi-task loss design, enabling the network to automatically learn the optimal balance ratio between first arrival picking loss, source location loss and uncertainty regularization loss during training, without the need for manual repeated experimentation to adjust the loss weight coefficients of each task. This design significantly reduces the hyperparameter tuning cost of the model when migrating and deploying it under different mining areas and different observation system conditions, and improves the engineering practicality and promotion value of the method.

[0070] To verify the feasibility and effectiveness of the method of this invention, the following describes the modeling and processing of the initial arrival picking and localization process of microseismic events from microseismic monitoring data of a coal mine. A total of 120 tagged events were collected, each with 12 events; 100 of these events were selected for model training. Figure 2The original microseismic signal collected by microseismic monitoring has had its first arrival manually picked out and stored in the header. It can be used as a label for model training. The red line in the figure indicates the manually picked first arrival position. Figure 3 This is a schematic diagram showing the results of preprocessing and adaptive processing of bad channels in the original microseismic records. Figure 4 The results of model training and validation were obtained using measured microseismic data. The validation results show that 44.9% of the traces had an event picking accuracy of less than 10ms, and 93.9% of the traces had a picking accuracy of less than 50ms. This indicates that the model has high accuracy after training with 100 events. Figure 5 , Figure 6 The results of first arrival picking using the model for microseismic events with different signal-to-noise ratios are shown. It can be seen that the method also has high first arrival picking accuracy for microseismic events with low signal-to-noise ratios. Figure 7 The results of source location analysis were obtained by using the method described in this patent on 5 events out of 120 labeled data. The comparison shows that 2 microseismic events have high location results, and the location results of 2 microseismic events have a location error within 30 meters. It can be seen that 80% of the microseismic event location results are within 30 meters of error.

Claims

1. A method for picking up the first arrival of microseismic events based on a context-aware multi-task network, characterized in that, Includes the following steps: Step 1. Loading and preprocessing of raw microseismic data; Step 2. Based on the data obtained in Step 1, identify and repair bad tracks caused by microseismic events; Step 3. Construction of a context-aware multi-task network model, including the following sub-steps: Step 31, Constructing a waveform encoder: Using a multi-scale one-dimensional convolution and bidirectional LSTM structure, extract the multi-time-scale temporal features of the waveform to construct a waveform encoder; Step 32, construct the detector context encoder: the detector spatial coordinates are mapped to the detector context embedding vector through a multilayer perceptron; Step 33, Constructing the context fusion module: The waveform features and sensor context are embedded to construct a gated fusion module to obtain the fused output features; Step 34, Construct multi-task output heads: Based on the fused output features, construct the first arrival picking head, the source location head, and the uncertainty estimation head respectively; Step 4. Design of adaptive multi-task loss function; Step 5. Model training and validation: Using the labels obtained in Step 2 as training samples, and employing the loss function designed in Step 4, train the context-aware multi-task network model constructed in Step 3 to obtain the trained context-aware multi-task network model. Step 6. First arrival time prediction and uncertainty output: Infer the newly acquired microseismic waveforms and output the first arrival time prediction values ​​and corresponding uncertainties for each sensor; Use the context-aware multi-task network model trained in Step 5 to pick up the first arrival events of the newly acquired microseismic data. During the processing, input the original normalized waveform of the microseismic event and the geophone coordinates. After processing, obtain the first arrival time prediction values ​​of each sensor, the first arrival uncertainty of each sensor, and the preliminary prediction of the source location. After the initial arrival prediction, an effective picking filter is added. If the filter criteria are met, the initial arrival prediction is considered effective. Step 7. Based on the joint inversion of the seismic source location, using the reciprocal of the uncertainty obtained in Step 6 as the weight, the differential evolution algorithm is used to jointly invert the three-dimensional coordinates of the seismic source and the time of occurrence.

2. The method for microseismic event first arrival picking based on context-aware multi-task network of claim 1, wherein, Step 1 includes the following sub-steps: Step 11: Read the SGY format microseismic event waveform data file collected by the microseismic monitoring system, and load the source coordinate file and detector coordinate file corresponding to the microseismic event; The manually calibrated first arrival times of microseismic events are parsed from the header of the SGY format microseismic event waveform data file, and then matched with the SGY format microseismic event waveform data file using the microseismic event ID; Step 12, perform data preprocessing on the data obtained in step 11, including: (1) using linear fitting to perform baseline drift correction on each data of the microseismic event; (2) Based on the frequency band range of the effective signal of the microseismic event, a bandpass filter is designed and applied to process the result of (1); (3) The surface uniform amplitude equalization method is used to process the result of (2); (4) Automatic gain control based on sliding window root mean square energy calculation is applied to the result of (3).

3. The microseismic event first arrival picking method based on context-aware multi-task networks as described in claim 1, characterized in that, In step 2, the specific method for bad sector identification is as follows: for each micro-seismic event, detection is performed event by event and channel by channel based on a preset threshold, and a Boolean mask is generated to identify the location of the bad sector; Bad sector repair method: For identified bad sectors, based on the spatial distance between their three-dimensional coordinates and the coordinates of surrounding valid sectors, an inverse distance weighted average algorithm is used to interpolate and reconstruct the bad sector waveform and first-arrival label point by point; the formula is: Among them, weight The formula for the inverse distance weighted average algorithm is as follows: In the above formula: The value represents the bad sectors that need to be repaired. The number of valid channels, For the first Observations of one effective channel, For the first The weights of each effective path are determined by their proximity; the closer the paths are, the greater the weight. The sum of all weights is 1. The smoothing constant is a very small positive number. For spatial distance, the detector position of the bad channel is relative to the first... The Euclidean distance between the detector positions of each effective channel The calculation formula is: in, The three-dimensional coordinates of the detector corresponding to the bad channel that needs to be repaired. For the first The three-dimensional coordinates of the detector corresponding to each effective channel; First arrival label repair: The first arrival time label of bad channels is filled by inverse distance weighted interpolation based on its spatial distance with the surrounding valid channels.

4. The microseismic event first arrival picking method based on context-aware multi-task networks as described in claim 3, characterized in that, Step 3 includes the following sub-steps: Step 31, Construct the waveform encoder A multi-scale one-dimensional convolution and bidirectional LSTM structure are used to extract multi-time-scale temporal features of the waveform to construct a waveform encoder. Specifically, the operation involves processing the input waveform... Parallel use K Different convolution kernel widths Perform convolution: And the above multi-scale feature maps are stitched together along the channel dimension: The multi-scale features obtained from the above concatenation are input into a bidirectional LSTM to capture forward and reverse temporal dependencies: in, The model has all trainable parameters, including the parameters of the forward LSTM and the parameters of the backward LSTM, which are learned through training. The hidden layer dimension is LSTM; self-attention is applied in the time dimension to enhance the features of keyframes near the initial arrival point: In the above formula, is the attention-weighted feature matrix, and is the output value; The temporal feature matrix output by the LSTM network, i.e., the input to self-attention, is a feature representation of the forward and backward temporal dependencies at each time step; Q The query matrix represents the search vector for actively querying whether relevant waveform features exist at other times in the current time step; K Labels representing the waveform characteristics at each time step; It is a normalization coefficient that scales the attention score; This means transforming the scaled score matrix into a probability distribution; The waveform feature content actually carried at each time step is the information carrier matrix of the self-attention mechanism; The learnable parameter matrix that linearly projects the input features H onto the value space; Let the feature dimension of the bidirectional LSTM hidden layer output be taken as... =256, For matrix The dimension of the feature vector at each time step is taken as... ; Step 32, Construct the detector context encoder Detector spatial coordinates This is mapped to a detector context embedding vector via a multilayer perceptron: in, for MLP The set of learnable parameters, including the weights / biases of fully connected layers and the scaling / offset parameters of batch normalized layers, are parameters optimized through backpropagation during training. i For detector index; Embed dimensions for context; Step 33, Build the context fusion module waveform characteristics With sensor context embedding Build the gating fusion module: In the formula, This is the gate weight matrix; This is the gated bias vector; Waveform characteristics Gating weighting; Spatial features Gating weighting; To fuse output features; Step 34, Construct the multi-task output header Based on the fusion output features in step 33, construct the initial arrival pickup heads respectively: In the above formula, The normalized first arrival offset of the output is multiplied by the time window length to restore it to the absolute first arrival time; To initially pick up fully connected neural network layers; For the first i Context-aware fusion feature vectors corresponding to each sensor; Constructing a seismic source location head: in, A global aggregation of features fused from all sensors. σ for Sigmoid Activation function To monitor the spatial range vector, A fully connected layer for locating the earthquake source; Constructing an uncertainty estimation head: in, Softplus for Softplus Activation function This is the network branch for uncertainty estimation.

5. The microseismic event first arrival picking method based on context-aware multi-task networks as described in claim 4, characterized in that, In step 4, construct the adaptive multi-task loss weights: in: These parameters are automatically optimized during training. These are learnable task weight parameters; Initial pickup loss: in, ∈{0,1} is the effective detector mask, which only accumulates the loss for the effective detector, and the error of the invalid channel is directly ignored; The number of detectors; For the first i The true first arrival time offset of each detector; For the first i The model predicts the first arrival time offset for each detector; The model predicts the first i Initial arrival pickup uncertainty of each detector; Loss due to earthquake focal location: in, The spatial coordinates of the earthquake source; The epicenter was at the 1st d Dimensional coordinates; To predict the earthquake source in the model at the 1st d Dimensional coordinates; = To predict the deviation between the coordinates and the actual coordinates; for Huber Loss function, for residuals r Robustness measure; for Huber The threshold for loss; Uncertainty canonical loss: in, The model predicts the first i Initial arrival uncertainty of each detector, This is the upper limit threshold for uncertainty; This means that only the portion exceeding the upper limit threshold will be penalized; if it does not exceed the limit, the value will be 0.

6. The microseismic event first arrival picking method based on context-aware multi-task networks as described in claim 5, characterized in that, In step 6, the screening criteria are: in, The model predicts the first i Initial arrival uncertainty of each detector, This is the threshold for the initial uncertainty of the model. For an effective detector mask to satisfy the above criteria, the predicted first arrival is considered valid.

7. The microseismic event first arrival picking method based on context-aware multi-task networks as described in claim 6, characterized in that, In step 7, the source coordinates are used. With the time of the earthquake Define the weighted residual objective function for the joint inversion parameters: in, For an effective set of detectors; = Using the reciprocal of the pickup uncertainty as a weight, detectors with higher pickup accuracy contribute more to positioning. The estimated wave velocity; The differential evolution algorithm is used to obtain the global optimization value of the objective function, and the microseismic event location result is obtained with the reciprocal of the initial picking uncertainty as the weight.

Images