Micro-seismic signal P-wave pickup method based on dual-channel conditional attention mechanism

Abstract

The invention relates to the technical field of mining engineering micro-seismic monitoring, and particularly discloses a micro-seismic signal P-wave pickup method based on a dual-channel conditional attention mechanism, and the method comprises the steps: firstly collecting a mine micro-seismic event, carrying out the multiple filtering processing of original waveform data, and manually marking a P-wave arrival time mark; then, the position of a waveform signal trigger point is detected through a short-time-window and long-time-window energy ratio algorithm, and data enhancement and standardization processing are carried out on a related data set; then, constructing a dual-channel conditional attention model, and inputting the standardized data set for training to obtain a P-wave pickup model; and finally, inputting a to-be-processed micro-seismic event waveform into the model to obtain a P-wave arrival time pickup result. According to the method, the waveform features and the trigger point position features are fused, the key feature extraction capability is enhanced in combination with a conditional attention mechanism, and the problems of low P-wave pickup precision and poor robustness in a complex noise environment are effectively solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of mining engineering microseismic monitoring, and particularly discloses a microseismic signal P-wave picking method based on a double-channel conditional attention mechanism. BACKGROUND

[0002] In the process of mining, microseismic signal processing plays a crucial role in microseismic positioning and rock burst warning. Microseismic signals are usually generated by the rupture of rock mass during mining, and contain key information about the internal damage evolution of the rock mass. The accurate picking of P-wave arrival time is the core basis for microseismic positioning and rock burst warning, because P-wave is the longitudinal wave in seismic waves, and its accuracy directly affects the positioning accuracy of microseismic events and the accuracy of rock burst warning.

[0003] At present, P-wave picking mainly relies on traditional methods and deep learning-based methods. Traditional methods include short-time window / long-time window energy ratio (STA / LTA) and autoregressive (AR) model method, etc. The core of these methods is to identify the arrival of P-wave through artificially set threshold. However, this method has significant limitations: first, it relies on artificially set threshold, which is prone to misjudgment in complex noise environment; second, it has limited detection capability for weak signals, because the artificially set threshold is often difficult to adapt to the change of signal intensity in different environments; third, the use of signal time sequence correlation is incomplete, and the time characteristics of the signal are not fully utilized to enhance the accuracy of detection.

[0004] In recent years, single-channel models such as convolutional neural network (CNN) and long short-term memory network (LSTM) based on deep learning have been introduced into the field of microseismic signal processing. These methods achieve automatic extraction of P-wave arrival time by learning the time and spectral characteristics of the signal, showing strong potential. However, existing deep learning methods also have shortcomings: on the one hand, only the characteristics of the signal itself are used, without fully combining prior information such as STA / LTA trigger points, resulting in incomplete use of signal time sequence correlation; on the other hand, there is a lack of attention mechanism for the characteristics of microseismic signals, making it difficult to focus on effective P-wave features in strong noise background.

[0005] In summary, the existing microseismic signal processing technology has obvious shortcomings in the mining environment, especially in the accurate picking of P-wave arrival time in complex noise environment, and a new microseismic signal P-wave picking method based on double-channel conditional attention mechanism is needed to improve the accuracy and robustness of microseismic signal processing. SUMMARY

[0006] The present application is a microseismic signal P-wave picking method based on a double-channel conditional attention mechanism, which is proposed to solve the problems in the existing microseismic signal processing technology, such as easy misjudgment due to manual threshold setting, limited weak signal detection capability, lack of prior information fusion, difficulty in effective feature focusing under strong noise background, and inaccurate P-wave picking.

[0007] The present application provides a microseismic signal P-wave picking method based on a double-channel conditional attention mechanism, which comprises the following steps:

[0008] S1. Collecting microseismic events of a mine to obtain an original waveform data set;

[0009] S2. Multiple filtering processing is performed on the original waveforms in the original waveform data set to obtain a waveform data set composed of waveforms; P-wave arrival time picking is performed on the original waveforms in the original waveform data set by a manual method, and P-wave arrival time labeling is done to obtain a P-wave arrival time labeled data set;

[0010] S3. The trigger point positions of each waveform signal in the waveform data set are detected based on a short-time window and long-time window energy ratio algorithm to obtain a trigger point position data set;

[0011] S4. The waveform data set, trigger point position data set and P-wave arrival time labeled data set are respectively subjected to data enhancement and standardization processing to obtain a standardized waveform data set, standardized trigger point position data set and standardized P-wave arrival time labeled data set;

[0012] S5. A double-channel conditional attention model is constructed, which comprises an input layer, a double-channel feature module, a feature fusion and conditional attention module, an output layer and a training module;

[0013] S6. The standardized waveform data set, standardized trigger point position data set and standardized P-wave arrival time labeled data set are input into the double-channel conditional attention model, and the double-channel conditional attention model is trained, and after the training is completed, a P-wave picking model is obtained;

[0014] S7. The waveform of a microseismic event to be processed is input into the P-wave picking model to obtain the picking result of the P-wave arrival time.

[0015] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism according to some embodiments of the present application, in step S1, the first 6000 sampling points are taken from each microseismic event as an original waveform, the corresponding time length of the sampling points is 1s, the sampling rate is 6000Hz, and the part of the original waveform with a signal length less than 6000 sampling points is zero-filled to obtain an original waveform data set.

[0016] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in the step S2, the multiple filtering processing includes band-pass filtering, notch filtering and Savitzky-Golay filtering, the band-pass filtering adopts a fourth-order Butterworth filter, and a passband frequency range is set as 2Hz-200Hz, and the band-pass filtering is used to suppress low-frequency mechanical vibration noise and high-frequency electromagnetic interference in the original waveform data set; the notch filtering adopts an IIR notch filter with a center frequency of 50Hz and a quality factor of 35, and the notch filtering is used to eliminate residual power frequency interference after the band-pass filtering; and the Savitzky-Golay filtering adopts a smoothing filter with a window length of 101 and a polynomial order of 4, and the Savitzky-Golay filtering is used to retain high-frequency details of the signal after the notch filtering while reducing random noise.

[0017] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in the short-time window and long-time window energy ratio algorithm in the step S3, a short-time window length is 72 sampling points, a corresponding time length of each sampling point is 12ms, a sampling rate is 6000Hz, a long-time window length is 900 sampling points, a corresponding time length of each sampling point is 150ms, a sampling rate is 6000Hz, and an energy ratio threshold is 2; and a preset trigger condition of the trigger point position is that energy ratios of three continuous sampling points exceed the energy ratio threshold, and a maximum ratio point in the three continuous sampling points is located as the trigger point position corresponding to the current filtering.

[0018] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in the step S4, the data enhancement includes time offset enhancement on the waveform data set, the trigger point position data set and the P-wave arrival time label data set respectively, maintaining a relative relationship between the trigger point position and the P-wave arrival time label of the same waveform, expanding data quantity, and obtaining a waveform enhanced data set, a trigger point position enhanced data set and a P-wave arrival time label enhanced data set.

[0019] The standardization processing includes Z-score standardization processing on the waveform enhanced data set to obtain a standardized waveform data set, Gaussian distribution thermodynamic map conversion on the trigger point position enhanced data set to obtain a standardized trigger point position data set, and mapping of P-wave arrival time labels in the P-wave arrival time label enhanced data set to an interval of [-1, 1] to obtain a standardized P-wave arrival time label data set.

[0020] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in step S5, the double-channel feature module includes a main channel and an auxiliary channel, the main channel includes a first main convolutional layer, a first main pooling layer, a second main convolutional layer, a second main pooling layer and a main bidirectional long short-term memory layer connected in sequence, and the auxiliary channel includes a first auxiliary convolutional layer, a first auxiliary pooling layer, a second auxiliary convolutional layer, a second auxiliary pooling layer and an auxiliary bidirectional long short-term memory layer connected in sequence.

[0021] The feature fusion and conditional attention module includes a conditional attention layer, a bidirectional long short-term memory layer and a full connection layer connected in sequence.

[0022] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in step S5, the input layer includes a first channel and a second channel, the first channel is connected with the main channel, and the second channel is connected with the auxiliary channel.

[0023] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in step S6, the input of the main channel is the standardized waveform data set, the main channel is used for extracting the feature of the waveform, and the output is the waveform feature; the input of the auxiliary channel is the standardized trigger point position data set, the auxiliary channel is used for extracting the position feature of the trigger point, and the output is the trigger point position feature, and the output of the output layer is the predicted P-wave arrival time data.

[0024] According to the microseismic signal P-wave picking method based on the double-channel conditional attention mechanism, in step S6, the conditional attention layer fuses the waveform feature and the trigger point position feature to obtain a fusion feature, and the conditional attention layer dynamically weights the hidden state of the fusion feature through a normalized attention weight matrix to obtain a weighted fusion feature, as shown in formula (1):

[0025] (1)

[0026] wherein, q represents a query vector, , f represents a fusion feature matrix, , w represents a waveform feature, p represents a trigger point position feature, and the size of the fusion feature matrix is T rows x D in columns, q represents a query vector weight matrix, and the size of the query vector weight matrix is D inRow x D out Column, denotes a key vector, , denotes a key vector weight matrix, a query vector weight matrix of size D in Row x D out Column, denotes a value vector, , denotes a value vector weight matrix, a value vector weight matrix of size D in Row x D out Column, denotes a normalized attention weight matrix;

[0027] The normalized attention weight matrix is obtained by normalizing the attention weight matrix to obtain the normalized attention weight matrix , as shown in equation (2):

[0028] (2)

[0029] wherein, denotes normalization processing, denotes an attention weight matrix;

[0030] The attention weight matrix is obtained by performing conditional attention weight calculation on the fused features, jointly adjusting feature interaction by a spatial constraint mask, to obtain the attention weight matrix , as shown in equation (3):

[0031] (3)

[0032] wherein, denotes the number of rows, is used to adjust the dot product scale, denotes an adaptive adjustment factor, denotes a spatial constraint mask;

[0033] The spatial constraint mask is shown in equation (4):

[0034] (4)

[0035] wherein, denotes an arbitrary sampling point corresponding on the waveform, denotes a trigger point position, denotes an attention radius.

[0036] According to the P-wave picking method of microseismic signals based on the dual-channel conditional attention mechanism, the input of the training module in step S6 is the predicted P-wave arrival time data and the standardized P-wave arrival time label data set, and the training module is used to evaluate the performance of the dual-channel conditional attention model. When the performance of the dual-channel conditional attention model meets the set accuracy, the P-wave picking model is output.

[0037] The P-wave picking method of microseismic signals based on the dual-channel conditional attention mechanism proposed by the application makes breakthrough progress in the deep fusion of traditional detection methods and data characteristics in terms of the precision bottleneck problem of P-wave arrival time picking in deep mine microseismic monitoring. The application constructs an algorithm detection-data dual-driven heterogeneous dual-channel model, solves the problem of insufficient utilization of prior knowledge of traditional algorithm detection by the traditional deep learning method through cross-modal fusion of the main channel and the auxiliary channel. An adaptive attention mechanism is proposed, which is constrained and guided by a short-time window and a long-time window energy ratio algorithm detection. Through adjustable parameters, the dynamic balance of the spatial constraint mask and the weight distribution driven by the data characteristics is achieved. This mechanism significantly improves the feature focusing ability of the model in the trigger point position neighborhood, maintains the error stability within ±20 sampling points under noise interference, and has significant optimization compared with ordinary attention mechanisms. By fusing waveform features and trigger point position features, and combining the conditional attention mechanism to enhance the key feature extraction capability, the problem of low P-wave picking precision and poor robustness of existing methods in complex noise environment is solved. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 FIG. 1 is a flowchart of the P-wave picking method of microseismic signals based on the dual-channel conditional attention mechanism according to an embodiment of the application;

[0039] Figure 2 FIG. 4 is a conditional attention mechanism principle diagram of the conditional attention layer according to an embodiment of the application;

[0040] Figure 3 FIG. 6 is a distribution characteristic diagram of P-wave arrival time picking errors of different models according to an embodiment of the application. DETAILED DESCRIPTION

[0041] The embodiments of the application will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the application, but cannot be used to limit the scope of the application.

[0042] Embodiment 1, the embodiment provides a P-wave picking method of microseismic signals based on a dual-channel conditional attention mechanism, as shown in FIG. 1, comprising the following steps: Figure 1

[0043] S1. Collect the microseismic events of the mine to obtain the original waveform data set.​

[0044] S2. Multiple filtering is performed on the original waveforms in the original waveform dataset to obtain a waveform dataset composed of waveforms; P-wave arrival time picking is performed on the original waveforms in the original waveform dataset by an artificial method, and P-wave arrival time marking is performed to obtain a P-wave arrival time marking dataset.

[0045] S3. A short-time window and long-time window energy ratio algorithm is used to detect the trigger point positions of each waveform signal in the waveform dataset to obtain a trigger point position dataset.

[0046] S4. The waveform dataset, the trigger point position dataset, and the P-wave arrival time marking dataset are subjected to data enhancement and standardization processing respectively to obtain a standardized waveform dataset, a standardized trigger point position dataset, and a standardized P-wave arrival time marking dataset.

[0047] S5. A dual-channel conditional attention model is constructed, which includes an input layer, a dual-channel feature module, a feature fusion and conditional attention module, an output layer, and a training module.

[0048] S6. The standardized waveform dataset, the standardized trigger point position dataset, and the standardized P-wave arrival time marking dataset are input into the dual-channel conditional attention model, and the dual-channel conditional attention model is trained, and after the training is completed, a P-wave picking model is obtained, wherein the standardized waveform dataset and the standardized trigger point position dataset are training datasets, and the standardized P-wave arrival time marking dataset is a training label set.

[0049] S7. The waveforms of a microseismic event to be processed are input into the P-wave picking model to obtain a P-wave arrival time picking result.

[0050] Embodiment 2 provides a microseismic signal P-wave picking method based on a dual-channel conditional attention mechanism, which includes the following steps:

[0051] S1. Microseismic events of a mine are collected to obtain an original waveform dataset.

[0052] Specifically, in step S1, the first 6000 sampling points of each microseismic event are intercepted as original waveforms, the corresponding time length of the sampling points is 1 s, the sampling rate is 6000 Hz, and the part of the original waveforms with a signal length less than 6000 sampling points is zero-padded to obtain the original waveform dataset. More specifically, 1930 microseismic events are selected in this embodiment.

[0053] S2. Multiple filtering is performed on the original waveforms in the original waveform dataset to obtain a waveform dataset composed of waveforms; P-wave arrival time picking is performed on the original waveforms in the original waveform dataset by an artificial method, and P-wave arrival time marking is performed to obtain a P-wave arrival time marking dataset.

[0054] In step S2, the multiple filtering processing includes band-pass filtering, notch filtering and Savitzky-Golay filtering. The band-pass filtering adopts a fourth-order Butterworth filter, and the passband frequency range is set to 2 Hz-200 Hz. The band-pass filtering is used to suppress low-frequency mechanical vibration noise and high-frequency electromagnetic interference in the original waveform data set, retain the effective frequency band characteristics of the microseismic signal, and improve the signal-to-noise ratio. The notch filtering adopts an IIR notch filter with a center frequency of 50 Hz and a quality factor of 35. The notch filtering is used to eliminate the residual power frequency interference after the band-pass filtering. By adjusting the Q value, high-selectivity suppression of narrowband noise is achieved, and the interference of power grid noise on the P-wave initial motion characteristics is reduced. The Savitzky-Golay filtering adopts a smoothing filter with a window length of 101 and a polynomial order of 4. The Savitzky-Golay filtering is used to retain the high-frequency details of the signal after the notch filtering while reducing random noise. The Savitzky-Golay filtering realizes adaptive smoothing of the signal through sliding polynomial fitting, avoiding distortion of the P-wave front characteristics by traditional filtering.

[0055] S3. Detect the trigger point position of each waveform signal in the waveform data set based on the short-time window and long-time window energy ratio algorithm, and obtain a trigger point position data set.

[0056] In the short-time window and long-time window energy ratio algorithm in step S3, first, a sliding window is calculated, and a short-time window and a long-time window are selected. The energy ratio is dynamically calculated, wherein the short-time window length is 72 sampling points, each sampling point corresponds to a time length of 12 ms, the sampling rate is 6000 Hz, the long-time window length is 900 sampling points, each sampling point corresponds to a time length of 150 ms, the sampling rate is 6000 Hz, and the energy ratio threshold is 2. Then, the trigger point position is determined. The preset trigger condition of the trigger point position is that the energy ratio of three consecutive sampling points exceeds the energy ratio threshold for three consecutive sampling points, and the maximum ratio point in the three consecutive sampling points is located as the trigger point position corresponding to the current filtering.

[0057] S4. The waveform data set, the trigger point position data set and the P-wave arrival time label data set are respectively subjected to data enhancement and standardization processing, and the standardized waveform data set, the standardized trigger point position data set and the standardized P-wave arrival time label data set are obtained.

[0058] In step S4, the data enhancement includes time offset enhancement on the waveform dataset, the trigger point position dataset and the P-wave arrival time label dataset respectively, maintaining the relative relationship between the trigger point position and the P-wave arrival time label of the same waveform, expanding the data quantity, obtaining a waveform enhanced dataset, a trigger point position enhanced dataset and a P-wave arrival time label enhanced dataset; more specifically, in the time offset enhancement, the offset range is ±10 to ±40 sampling points, corresponding to ±1.67 ms to ±6.67 ms under a sampling rate of 6000 Hz, and the original 1930 microseismic event sample numbers are expanded to 5705 training numbers, and each training group is a "standardized waveform-standardized trigger point position-standardized P-wave arrival time label" sample pair.

[0059] The standardization processing includes Z-score standardization processing on the waveform enhanced dataset to obtain a standardized waveform dataset, Gaussian distribution heat map conversion on the trigger point position enhanced dataset to obtain a standardized trigger point position dataset, and mapping the P-wave arrival time label in the P-wave arrival time label enhanced dataset to the interval [-1, 1] to obtain a standardized P-wave arrival time label dataset.

[0060] S5. Constructing a dual-channel conditional attention model, the dual-channel conditional attention model including an input layer, a dual-channel feature module, a feature fusion and conditional attention module, an output layer and a training module.

[0061] In step S5, the dual-channel feature module includes a main channel and an auxiliary channel, the main channel including a first main convolutional layer, a first main pooling layer, a second main convolutional layer, a second main pooling layer and a main bidirectional long short-term memory layer connected in sequence, and the auxiliary channel including a first auxiliary convolutional layer, a first auxiliary pooling layer, a second auxiliary convolutional layer, a second auxiliary pooling layer and an auxiliary bidirectional long short-term memory layer connected in sequence; the main channel and the auxiliary channel are symmetrically designed and are both 1D convolution-LSTM processing chains. The feature fusion and conditional attention module includes a conditional attention layer, a bidirectional long short-term memory layer and a fully connected layer connected in sequence. The input layer includes a first channel and a second channel, the first channel being connected with the main channel, and the second channel being connected with the auxiliary channel.

[0062] More specifically, the input layer dimension is (6000, 2), the first channel input is the standardized waveform data set, and the second channel input is the standardized trigger point position data set; the convolution kernel size of the first main convolution layer and the first auxiliary convolution layer is 15, and the step is 4, which are used to capture long-time features; the convolution kernel size of the second main convolution layer and the second auxiliary convolution layer is 7, and the step is 2, which are used to capture short-time features; the maximum pooling factor of the first main pooling layer and the first auxiliary pooling layer is 4, and the maximum pooling factor of the second main pooling layer and the second auxiliary pooling layer is 2, which are used to gradually reduce the feature dimension and retain the key information; the number of units of the main bidirectional long short-term memory layer and the auxiliary bidirectional long short-term memory layer is 128, which are used to capture the bidirectional time sequence dependence of the input data. The width of the conditional attention layer is 40, that is, each time step pays attention to the context information of the previous and next 20 time steps, and the activation function is softmax, which ensures the normalization of the weight and focuses on the high-energy P wave feature area. After the conditional attention layer fuses the features of the two branches, a sequence self-attention mechanism is introduced to trigger the point position as the conditional weight.

[0063] S6. The standardized waveform data set, the standardized trigger point position data set and the standardized P wave to time label data set are input into the dual-channel conditional attention model, and the dual-channel conditional attention model is trained. After the training is completed, the P wave picking model is obtained. The standardized waveform data set and the standardized trigger point position data set are both training data sets, and the standardized P wave to time label data set is a training label set.

[0064] More specifically, before input, the standardized waveform data set, the standardized trigger point position data set and the standardized P wave to time label data set are divided into training set, validation set and test set respectively, and the training set (3993 groups), the validation set (856 groups) and the test set (856 groups) are divided in the proportion of 7:1.5:1.5, and the stratified random division strategy is adopted to ensure the consistency of data distribution.

[0065] In step S6, the input of the main channel is the standardized waveform data set, the main channel is used to extract the features of the waveform, and the output is the waveform feature; the input of the auxiliary channel is the standardized trigger point position data set, the auxiliary channel is used to extract the position feature of the trigger point, and the output is the trigger point position feature, and the output of the output layer is the predicted P wave to time data.

[0066] In step S6, the conditional attention layer fuses the waveform feature and the trigger point position feature to obtain the fused feature, and the trigger point position is embedded as a spatial constraint to calculate the attention. The conditional attention layer dynamically weights the hidden state of the fused feature by normalizing the attention weight matrix, so that the model focuses on the energy mutation feature in the trigger point position neighborhood, and obtains the weighted fused feature, as shown in formula (1):

[0067] (1)

[0068] wherein, denotes a query vector, , denotes a fusion feature matrix, , denotes a waveform feature, denotes a trigger point position feature, the fusion feature matrix has a size of T rows x D in columns, denotes a query vector weight matrix, the query vector weight matrix has a size of D in rows x D out columns, denotes a key vector, , denotes a key vector weight matrix, the query vector weight matrix has a size of D in rows x D out columns, denotes a value vector, , denotes a value vector weight matrix, the value vector weight matrix has a size of D in rows x D out columns, denotes a normalized attention weight matrix;

[0069] the normalized attention weight matrix is obtained by normalizing the attention weight matrix, and the normalized attention weight matrix is as shown in formula (2):

[0070] (2)

[0071] wherein, denotes normalization processing, denotes the attention weight matrix;

[0072] the attention weight matrix is obtained by performing conditional attention weight calculation on the fusion feature, and jointly adjusting feature interaction by a spatial constraint mask, to obtain the attention weight matrix , as shown in formula (3):

[0073] (3)

[0074] wherein, denotes the number of rows, is used to adjust dot product scale, an adaptive adjustment factor, a spatial constraint mask;

[0075] a spatial constraint mask a spatial constraint mask centered at the trigger location ttrigger with a radius of concern δ sampling points , suppressing noise interference far from the trigger location region, as shown in equation (4):

[0076] (4)

[0077] wherein, denotes the corresponding arbitrary sampling point on the waveform, denotes the trigger location, denotes the radius of concern.

[0078] The conditional attention mechanism principle diagram of the conditional attention layer is shown in Figure 2 .

[0079] The input of the training module is the predicted P-wave arrival time data and the standardized P-wave arrival time label data set, and the training module is used to evaluate the performance of the dual-channel conditional attention model. When the performance of the dual-channel conditional attention model meets the set accuracy, the P-wave pickup model is output.

[0080] The design of the conditional attention layer realizes the cooperative optimization of data-driven and traditional verification algorithm, which not only retains the abstract ability of neural network to complex features, but also suppresses the interference of irrelevant regions through the trigger point location prior.

[0081] In addition, a regularization method is also used, specifically L2 regularization (kernel_regularizer='l2') to prevent overfitting.

[0082] In the training of the dual-channel conditional attention model, the MSE loss function and the Adam optimizer are used, combined with early stopping (patience=25), model checkpointing and learning rate decay (factor=0.5, patience=15). In the training process optimization, the batch size is 256, balancing memory occupation and gradient update stability; the number of iterations is 250 epochs, and the early stopping strategy prevents overfitting; the learning rate scheduling is that the initial learning rate is 1e-3, and when the validation set MAE does not decrease for 15 consecutive rounds, the learning rate is halved.

[0083] S7. Input the microseismic event waveform to be processed into the P-wave pickup model to obtain the pickup result of the P-wave arrival time.

[0084] This example compares the performance of different models in picking P-wave arrival times on the same dataset in terms of MAE mean absolute error, RMSE root mean square error, and MAPE mean absolute percentage error. The specific results are shown in Table 1:

[0085] Table 1 Performance of P-wave arrival time picking results of different models

[0086]

[0087] It can be seen that the MAE, RMSE and MAPE of the method in this embodiment are all the best, which are 8.5%, 5.9% and 7.6% lower than the single-channel conditional attention model with the best performance among traditional methods. The error is one order of magnitude higher than that of the traditional STA / LTA method, which verifies the overwhelming advantage of the method in this embodiment. Figure 3 Figure 2 shows the distribution characteristics of the P-wave arrival timing errors of the five models. The P-wave arrival timing error distribution of the P-wave arrival timing model is highly concentrated within the range of ±18 sampling points (accounting for 85.76%), significantly outperforming the single-channel conditional attention model (80.28%) and the single-channel general attention model (78.06%). The STA / LTA method has the widest error distribution, with only 17% of the error concentrated within ±18 sampling points. Specifically, the dual-channel model exhibits symmetrical concentration characteristics in P-wave arrival timing, with the error peak located near zero. The STA / LTA method exhibits systematic delays in P-wave arrival timing (mainly distributed between 18 and 36 sampling points).

[0088] Accurately picking the arrival time of microseismic signals (P) in mines is a core technical challenge for rock fracture mechanism analysis and disaster early warning. To address bottlenecks such as the low accuracy of traditional STA / LTA methods in low signal-to-noise ratio environments and the insufficient feature coupling of existing deep learning methods, this embodiment proposes a microseismic signal P-wave picking method based on a dual-channel conditional attention mechanism. This innovatively constructs a dual-channel input architecture: the primary channel focuses on waveform time-frequency feature extraction, while the auxiliary channel embeds the STA / LTA (short-time window and long-time window energy ratio algorithm detection) trigger point location. These channels process denoised waveform features and trigger point location features separately, and combine the STA / LTA constraint-guided conditional attention mechanism to enhance features during critical time periods. This method has significant application potential in fields such as real-time monitoring of deep mine dynamic hazards.

[0089] The embodiments of the present invention are presented for purposes of illustration and description and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments are chosen and described in order to better illustrate the principles of the invention and its practical application and to enable those skilled in the art to understand the invention and design various embodiments with various modifications as suited for specific applications.

Claims

1. A microseismic signal P-wave picking method based on a dual-channel conditional attention mechanism, characterized in that: The steps include: S1. Collect microseismic events in the mine and obtain a raw waveform dataset; S2. Performing multiple filtering on the original waveform in the original waveform data set to obtain a waveform data set consisting of waveforms; manually picking up the original waveform in the original waveform data set when the P wave arrives, and marking the P wave arrival to obtain a P wave arrival marking data set; S3 based on the short window and long window energy ratio algorithm to detect the trigger point position of each waveform signal in the waveform data set to obtain the trigger point position data set; S4. The waveform data set, trigger point position data set and P-wave arrival time mark data set are respectively enhanced and standardized to obtain a standardized waveform data set, a standardized trigger point position data set and a standardized P-wave arrival time mark data set; S5. Constructing a dual-channel conditional attention model, the dual-channel conditional attention model includes an input layer, a dual-channel feature module, a feature fusion and conditional attention module, an output layer, and a training module; S6. Inputting the standardized waveform dataset, the standardized trigger point position dataset, and the standardized P-wave arrival time marker dataset into the dual-channel conditional attention model, training the dual-channel conditional attention model, and obtaining a P-wave picking model after the training is completed; S7. Inputting the microseismic event waveform to be processed into the P-wave picking model to obtain a picking result of the P-wave arrival time.

2. A microseismic signal P wave picking method based on a dual-channel conditional attention mechanism according to claim 1, characterized in that: In step S1, the first 6000 sampling points are intercepted from each microseismic event as the original waveform, the corresponding duration of the sampling points is 1 second, the sampling rate is 6000 Hz, and the portion of the original waveform with a signal length less than 6000 sampling points is zero-filled to obtain the original waveform data set.

3. The microseismic signal P wave picking method based on the dual-channel conditional attention mechanism according to claim 1 is characterized in that: In step S2, the multiple filtering processing includes bandpass filtering, notch filtering and Savitzky-Golay filtering. The bandpass filtering adopts a fourth-order Butterworth filter, and the passband frequency range is set to 2 Hz to 200 Hz. The bandpass filtering is used to suppress low-frequency mechanical vibration noise and high-frequency electromagnetic interference in the original waveform data set; the notch filtering adopts an IIR notch filter with a center frequency of 50 Hz and a quality factor of 35. The notch filtering is used to eliminate the power frequency interference remaining after the bandpass filtering; the Savitzky-Golay filtering adopts a smoothing filter with a window length of 101 and a polynomial order of 4. The Savitzky-Golay filtering is used to reduce random noise while retaining the high-frequency details of the signal after the notch filtering.

4. The microseismic signal P wave picking method based on the dual-channel conditional attention mechanism according to claim 1 is characterized in that: In the short-time window and long-time window energy ratio algorithm described in step S3, the short-time window length is 72 sampling points, the corresponding duration of each sampling point is 12ms, the sampling rate is 6000Hz, the long-time window length is 900 sampling points, the corresponding duration of each sampling point is 150ms, the sampling rate is 6000Hz, and the energy ratio threshold is 2; the preset trigger condition of the trigger point position is that the energy ratio of three consecutive sampling points exceeds the energy ratio threshold, and the maximum ratio point among the three consecutive sampling points is located as the trigger point position corresponding to the current filter.

5. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 1 is characterized in that: In step S4, the data enhancement includes performing time offset enhancement on the waveform dataset, the trigger point position dataset, and the P-wave arrival time mark dataset, respectively, maintaining the relative relationship between the trigger point position and the P-wave arrival time mark of the same waveform, and expanding the amount of data to obtain a waveform enhanced dataset, a trigger point position enhanced dataset, and a P-wave arrival time mark enhanced dataset; The standardization processing includes performing Z-score standardization processing on the waveform enhancement data set to obtain a standardized waveform data set; performing Gaussian distribution heat map conversion on the trigger point position enhancement data set to obtain a standardized trigger point position data set; The P-wave arrival time markers in the P-wave arrival time marker enhanced dataset are mapped to the interval [-1, 1] to obtain a standardized P-wave arrival time marker dataset.

6. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 1 is characterized in that: In step S5, the dual-channel feature module includes a main channel and an auxiliary channel, the main channel includes a first main convolutional layer, a first main pooling layer, a second main convolutional layer, a second main pooling layer, and a main bidirectional long short-term memory layer connected in sequence, and the auxiliary channel includes a first auxiliary convolutional layer, a first auxiliary pooling layer, a second auxiliary convolutional layer, a second auxiliary pooling layer, and an auxiliary bidirectional long short-term memory layer connected in sequence; The feature fusion and conditional attention module includes a conditional attention layer, a bidirectional long short-term memory layer and a fully connected layer connected in sequence.

7. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 6 is characterized in that: In step S5, the input layer includes a first channel and a second channel, the first channel is connected to the main channel, and the second channel is connected to the auxiliary channel.

8. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 6 is characterized in that: In step S6, the input of the main channel is the standardized waveform data set, the main channel is used to extract the characteristics of the waveform, and the output is the waveform characteristics; the input of the auxiliary channel is the standardized trigger point position data set, the auxiliary channel is used to extract the position characteristics of the trigger point, and the output is the trigger point position characteristics, and the output of the output layer is the predicted P wave arrival data.

9. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 6 is characterized in that: In step S6, the conditional attention layer fuses the waveform feature with the trigger point position feature to obtain a fusion feature. The conditional attention layer dynamically weights the hidden state of the fusion feature through a normalized attention weight matrix to obtain a weighted fusion feature, as shown in formula (1): （1） in, represents the query vector, , represents the fusion feature matrix, , Represents the waveform characteristics, Represents the trigger point position feature and fusion feature matrix The size is T rows × D in List, Represents the query vector weight matrix, query vector weight matrix The size is D in Row × D out List, represents the key vector, , Represents the key vector weight matrix and the query vector weight matrix The size is D in Row × D out List, represents a value vector, , Represents the value vector weight matrix, the value vector weight matrix The size is D in Row × D out List, represents the normalized attention weight matrix; The normalized attention weight matrix To normalize the attention weight matrix, we get the normalized attention weight matrix , as shown in formula (2): （2） in, express Normalization processing, represents the attention weight matrix; The attention weight matrix In order to calculate the conditional attention weights of the fused features, the joint spatial constraint mask dynamically adjusts the feature interaction to obtain the attention weight matrix , as shown in formula (3): （3） in, Indicates the number of rows, Used to adjust the dot product scale, represents the adaptive adjustment factor, represents the spatial constraint mask; The spatial constraint mask As shown in formula (4): （4） in, Indicates the corresponding arbitrary sampling point on the waveform, Indicates the trigger point position, Indicates the attention radius.

10. The microseismic signal P wave picking method based on dual-channel conditional attention mechanism according to claim 8, characterized in that: In step S6, the input of the training module is the predicted P-wave arrival data and the standardized P-wave arrival mark data set. The training module is used to evaluate the performance of the dual-channel conditional attention model. When the performance of the dual-channel conditional attention model meets the set accuracy, the P-wave picking model is output.

Images