PSSegNet: Segmenting the P- and S-Phases in Microseismic Signals through Deep Learning

Abstract

Microseismic P- and S-phase segmentation is an influential step that limits the accuracy of event location, parameter inversion, and mechanism analysis. Therefore, an improved Unet named PSSegNet is proposed to intelligently segment the P- and S-phases. The designed masks are used as the outputs of PSSegNet, which is used to obtain the time–frequency features of the P- and S-phases. As a result, the MSE (mean square error) between the predicted mask and the actual labeled mask is concentrated below 2.5, and the AE (accumulated error) of the reconstructed P/S-phase based on the predicted mask is concentrated below 1.0 × 10⁻³. Arrival picking results show that the overall error of the entire test set is less than 50 ms and most of the errors are less than 20 ms. Data with SNR (signal to noise ratio) < 2, 2 ≤ SNR < 3, PSR (P-phase to S-phase ratio) < 1, or 1 ≤ PSR < 2 in the dataset were selected for arrival picking and their errors were counted. The statistical results show that PSSegNet is robust at low SNR and PSR. The P- and S-phase segmentation based on PSSegNet has excellent potential for use in various applications and can effectively reduce the difficulty of obtaining the P/S-phase arrivals.

1. Introduction

Environmental perception is an essential link in the construction of smart mines. The definition of a mine environment is extremely broad, including air quality [1], personnel and equipment location [2,3], rock mechanical state [4], and so on. Among them, the mechanical state of the rock mass is a crucial factor for mine production safety. Currently, microseismic monitoring technology is an advanced method to perceive the mechanical state of rock masses. Microseismic monitoring techniques can determine the mechanical activity of rock masses by monitoring the propagation of mechanical waves in rock masses in real time. However, to reflect mechanical activity from mechanical waves, it is necessary to process the collected waveform data. Furthermore, the treatment of microseismic wave shapes is dominated by the analysis of the P- and S-phases.

The microseismic monitoring region of the rock mass is considerably smaller than the natural seismic region. Although there is a difference in velocity between the P- and S-phases, the microseismic wave shape commonly appears as a partial overlap of the P- and S-phases due to the short propagation path. Therefore, in microseismic analysis, researchers typically focus solely on P-phase analysis, with a common task being P-phase arrival picking. Among the classical approaches for P-phase arrival picking are window-based, non-window-based, and hybrid approaches. The window-based methods include the short- and long-time average (STA/LTA) ratio [5,6,7], the modified energy ratio (MER) algorithm [8,9], the modified Coppens’ method (MCM) [10,11], the phase arrival identification kurtosis (PAI-K) [12,13,14], and the short-term kurtosis to long-term kurtosis (S/L-Kurt) ratio. The non-window-based methods include the Akaike information criterion (AIC) [15,16,17,18,19], the cross-correlation-based techniques [20,21,22,23], the image processing techniques [24,25], the global optimization-based techniques [26], and the beamforming (delay and stack) of waveforms [27]. Window-based and non-window-based methods each have advantages and disadvantages. Hybrid approaches have been proposed to combine the benefits of these approaches. These hybrid approaches include the wavelet transform-based approaches [28,29], Akazawa’s method [30], and the joint energy ratio and AIC (JER-AIC) [31], joint STA/LTA-polarization-AIC (TYFH’s method) [32], and EMD-AIC [33]. In addition, artificial intelligence (AI) techniques have begun to be utilized for P-phase arrival picking. The most representative methods include machine learning methods [34], deep recurrent neural networks (RNN) [35,36], convolutional neural networks (CNN) [37,38,39], and capsule networks (CapsNet) [40,41]. These efforts, which leverage AI technology, have yielded remarkable outcomes.

Compared to the investigation of the P-phase, the exploration of the S-phase in microseismic monitoring techniques is considerably more limited. For microseismic signals, the existing S-Phase picking is realized through multi-step pickup based on P-phase picking [42]. However, for seismic signals, there are various S-phase picking methods. Diehl et al. combined three common phase detection and picking methods (AR-AIC algorithm, STA/LTA ratio, and polarization analysis) to form a robust S-phase picking method [43]. Kuperkoch et al. [44] proposed an automatic S-phase arrival time determination algorithm for local, regional, and remote seismic events based on waveform autoregressive (AR) prediction. Lois et al. [45] obtained the first S-phase arrival time estimate by applying the kurtosis criterion to the designed characteristic function. Recently, Zhu et al. [46] proposed that PhaseNet selects the arrival times of the P wave and the S wave. PhaseNet takes a three-component seismic waveform as input and generates as output a probability distribution of P-arrivals, S-arrivals, and noise, where the peaks in the probability distribution provide accurate arrival times for the P- and S-phases. Zhou et al. [47] developed a hybrid algorithm that uses CNN and RNN, respectively, to select P- and S-phases in two steps from archived continuous waveforms. Mousavi et al. [48] proposed a global deep learning model for simultaneous seismic detection and phase selection by combining the staged and full-waveform information of seismic signals using a hierarchical attention mechanism. Soto and Schurr [49] introduced DeepPhasePick, an automatic two-stage approach in which the first stage uses CNNs to detect the phase and the second stage uses two RNNs to select two phases. Zhu et al. [50] introduced a new end-to-end approach consisting of three subnetworks: a backbone network that extracts features from the original waveform, a phase-picking subnetwork that extracts P- and S-wave arrivals based on these features, and an event detection subnetwork that aggregates features from multiple sites and detects earthquakes. The aforementioned S-phase picking methods mainly study seismic signals, and the adopted data are based on three-component waveforms. These methods mainly treat different seismic signal phases as different types of signals for classification studies. However, for microseismic signals, the P- and S-phases tend to partially overlap, so segmentation studies are also necessary.

Therefore, aiming at the particularity of microseismic signals (P- and S-phase partial overlap, single-component acquisition, low signal-to-noise ratio (SNR)), in this paper, a P- and S-phase segmentation method called PSSegNet is proposed based on mask design and deep learning. PSSegNet is designed with UNet as the basic framework, where features are extracted by downsampling and the final result is output through upsampling. The microseismic signals are used as input for deep learning after time–frequency analysis, and the corresponding masks are used as output through mask design in the P- and S-phases. Finally, the P-phase and S-phase in the microseismic signal are obtained by multiplying the mask predicted by PSSegNet witsh the time–frequency features of the original signal, respectively. PSSegNet can effectively separate overlapping P- and S-phases from microseismic signals with low SNR, which provides a data base for additional research on microseismic monitoring techniques.

2. Materials and Methods

2.1. Dataset

To be able to train and test PSSegNet, we collected 1100 microseismic data to form a dataset. The data came from a microseismic monitoring system deployed at the Huashugou Copper Mine. As shown in Figure 1, the Huashugou Copper Mine is located in the Sunan Yugur Autonomous County of Gansu, China. To understand the seismicity within the rock mass, a microseismic system was used to perform continuous 24-h monitoring. A total of 32 single-component accelerometers with a sensitivity of 10 V/g and a sampling frequency of 10 kHz were embedded in the Huashugou Copper Mine. Their coordinates are shown in Figure 2 and listed in

Table 1. Coordinates of the sensors.

Sensor	X	Y	Z	Sensor	X	Y	Z
R-2820-01	12,110.576	8615.488	2821.972	R-2700-01	12,236.783	8583.922	2702.058
R-2820-02	12,000.213	8670.474	2822.144	R-2700-02	12,241.372	8629.458	2701.662
R-2820-03	11,958.893	8761.138	2818.54	R-2700-03	12,191.075	8622.778	2702.345
R-2820-04	12,018.628	8843.893	2821.774	R-2700-04	12,164.519	8675.117	2702.125
R-2820-05	12,051.731	8804.489	2822.346	R-2700-05	12,108.346	8685.187	2702.04
R-2820-06	12,090.328	8744.155	2822.469	R-2700-06	12,085.612	8726.988	2702.401
R-2760-01	12,181.087	8573.237	2759.917	R-2700-07	12,138.957	8759.831	2702.275
R-2760-02	12,094.117	8619.037	2759.961	R-2700-08	12,047.007	8741.941	2702.096
R-2760-03	12,039.003	8679.226	2758.961	R-2700-09	12,025.402	8792.233	2702.281
R-2760-04	11,990.874	8733.964	2759.257	R-2700-10	11,966.233	8887.393	2702.286
R-2760-05	11,931.208	8875.974	2758.237	R-2700-11	11,921.350	8964.683	2702.219
R-2760-06	11,880.954	8917.851	2757.815	R-2700-12	11,842.552	9050.731	2701.619
R-2760-07	11,968.433	8859.945	2758.346	R-2700-13	11,833.117	9118.063	2702.046
R-2760-08	12,039.310	8810.887	2758.791	R-2640-14	12,150.326	8702.537	2647.752
R-2760-09	12,096.173	8750.540	2759.295	R-2640-15	12,085.930	8660.176	2648.047
R-2760-10	12,142.769	8704.456	2758.919	R-2640-16	12,185.542	8661.853	2647.841

. In addition, an example of a detected microseismic event is shown in Figure 3.

As shown in Figure 4, we present an example of a microseismic signal and the results of its time–frequency analysis. The signal distinctly exhibits several characteristics of microseismic signals, namely, P- and S-phase partial overlap, single-component acquisition, and low signal-to-noise ratio. In addition, we performed a simple statistical analysis of the overall characteristics of the dataset, and the results are shown in Figure 5. The overall distribution of data in the dataset is reasonable and can represent most of the actual collection results. In manual analysis, it is frequently segmented by looking at the difference between the P- and S-phases in the frequency range of the time–frequency domain. In the time–frequency domain, it is clear that the frequency range of the P-phase is narrower than that of the S-phase. Based on the difference in the time–frequency domain of the P- and S-phases, all data were manually labeled with the positions of the P- and S-phases, which served as the basis for the training and validation of PSSegNet.

2.2. Mask Design

According to the characteristics of the P- and S-phase overlap in microseismic signals, microseismic signal X(t) can be expressed as

$$X(t)=P(t)+S(t),$$

(1)

where P(t), and S(t) denotes the P- and S-phases in the microseismic signal, respectively. On this basis, to perform a time–frequency analysis of X(t), the following must be used:

$$X(t,f)=TFT\{X(t)\}=P(t,f)+S(t,f),$$

(2)

where TFT is the time–frequency transform. Accordingly, we use TFT⁻¹ to represent the inverse time–frequency transform.

Therefore, we can define two individual masks, $M_P(t,f)$ and $M_s(t,f)$, for signal and noise, respectively:

$$M_P(t,f)=\left[\frac{1}{1+\frac{|S(t,f)|}{|P(t,f)|}}\right],$$

(3)

$$M_S(t,f)=\left[\frac{\frac{|S(t,f)|}{|P(t,f)|}}{1+\frac{|S(t,f)|}{|P(t,f)|}}\right].$$

(4)

Each mask has the same size as the input time–frequency representation, $X(t,f)$, and contains values between 0 and 1.

The mask is designed to serve as the output of PSSegNet. Assuming that the time–frequency features of the signal are fed to PSSegNet, which is able to predict the mask efficiently, then we have

$$\widehat{P}(t)=TF{T}^{-1}\{M_P(t,f)X(t,f)\},$$

(5)

$$\widehat{S}(t)=TF{T}^{-1}\{M_S(t,f)X(t,f)\},$$

(6)

where $\widehat{P}(t)$ and $\widehat{S}(t)$ are the P- and S-phases after segmentation by PSSegNet, respectively. Therefore, the objective of P- and S-phase segmentation is to minimize the expected error between the P(t), S(t) and $\widehat{P}(t)$, $\widehat{S}(t)$:

$$error=E{\Vert \widehat{P}(t)-P(t)\Vert }_2^2+E{\Vert \widehat{S}(t)-S(t)\Vert }_2^2.$$

(7)

With the mask design, we set the objective function of the segmentation process as:

$$\begin{array}{cc}\min & -[\beta p\log \widehat{p}+(1-p)\log (1-\widehat{p})]+\frac{1}{2}{\Vert w\Vert }^2,\end{array}$$

(8)

where $p$ is the true mask and $\widehat{p}$ is the mask output by the model. $\beta$ is a positive sample-weighting parameter that reduces false negatives when greater than 1 and false positives when less than 1; w represents the parameters in the deep learning network. Therefore, we have the foundation to build a deep learning network.

2.3. PSSegNet

After the mask-assisted implementation of P- and S-phase segmentation, we proceed to build PSSegNet. The input to PSSegNet is the time–frequency domain’s microseismic signal after time–frequency analysis, and its output is a P- and S-phase mask. In this paper, we use the short-time Fourier transform to transform the microseismic signal in the time domain. In the constructed dataset, the microseismic signal was sampled for 0.2 s at a sampling rate of 10,000 Hz. Therefore, after transforming the time–frequency domain, it becomes a 33 × 35 time–frequency matrix in which the elements are in complex form (as shown in Figure 4b,c). The real part of the time–frequency matrix is separated from the imaginary part to form two matrices as input to PSSegNet. Thus, for PSSegNet, the input becomes 33 × 35 × 2. In addition, since the size of the mask is the same as the input, the output size of PSSegNet should be 33 × 35 × 2.

By analyzing the input and output sizes, it can be found that the final desired mask should also be a matrix of the same size and dimension as the time–frequency information matrix of the microseismic recording. In the field of deep learning, a common network model structure for such problems is the encoder–decoder architecture. The fully convolutional network (FCN) is an earlier neural network model with a coding–decoding structure. The main feature of the CNN is that it turns the last fully connected layer into the convolutional layer, uses the convolutional network to realize the up-sampling calculation in the convolutional process, and applies the skip layer structure to the network results, which effectively preserves the details in the convolutional process. UNet is a refinement of FCN, and the structure of the entire network is an upper-case English letter U, hence the name UNet [51]. In contrast to FCN, UNet uses concatenation as the fusion method for feature maps. Compared to the skip-layer structure of FCN, UNet deepens the field of view of feature layers by concatenating the number of channels, giving more retention of detailed features, ensuring the replenishment of vital information during up-sampling, and improving the resolution of the results.

As shown in Figure 6, we design a PSSegNet model for generating P/S wave masks based on UNet; the left half of the network is the encoding part to extract deeper features of the input time–frequency information, while the right half is the decoding part to up-sample and concatenate the extracted features. Its main components are convolutional layers, deconvolutional layers, and pooling layers. A convolutional layer extracts the features of the input data through a convolutional kernel, and the output of the convolutional layer is

$$y_j^l=relu\left({\displaystyle \sum _{i\in M_j}{x}_{ij}^l}\times k_{ij}^l+b_j^l\right),$$

(9)

where $x_{ij}^l$ represents the input of the jth neuron of the l layer, $k_{ij}^l$ is the convolution kernel of the j-th neuron of the l layer, $M_j$ represents the selected input, $b_j^l$ represents the bias of the j-th neuron of the l layer, relu is the activation function, and its calculation process is

$$relu(x)=\max (0,x).$$

(10)

The pooling layer reduces the resolution of the feature maps and prevents overfitting due to the precise location of the feature maps, which negatively affects the performance. The output of the pooling layer can be expressed as

$$y_j^l=relu\left({\beta }_j^l\mathrm{pooling}\left(x_j^{l-1}\right)+b_j^l\right).$$

(11)

To enable the generation of the masks of P- and S-phase for the time–frequency information of microseismic recordings, the proposed network model contains 18 convolutional layers, 4 pooling layers, and 4 deconvolutional layers. The specific parameters of the model are listed in

Table 2. Parameters of PSSegNet.

Number	Layer	Output	Kernel/Stride	Copy and Crop
1	Conv 1	33 × 35 × 64	3 × 3/1
2	Conv 2	33 × 35 × 64	3 × 3/1
3	Max pooling 1	16 × 17 × 64	2 × 2/2
4	Conv 3	16 × 17 × 128	3 × 3/1
5	Conv 4	16 × 17 × 128	3 × 3/1
6	Max pooling 2	8 × 8 × 128	2 × 2/2
7	Conv 5	8 × 8 × 256	3 × 3/1
8	Conv 6	8 × 8 × 256	3 × 3/1
9	Max pooling 3	4 × 4 × 256	2 × 2/2
10	Conv 7	4 × 4 × 512	3 × 3/1
11	Conv 8	4 × 4 × 512	3 × 3/1
12	Max pooling 4	2 × 2 × 512	2 × 2/2
13	Conv 9	2 × 2 × 1024	3 × 3/1
14	Conv 10	2 × 2 × 1024	3 × 3/1
15	Up-Conv 1	4 × 4 × 1024	2 × 2/1	Conv 8
16	Conv 11	4 × 4 × 512	3 × 3/1
17	Conv 12	4 × 4 × 512	3 × 3/1
18	Up-Conv 2	8 × 8 × 512	2 × 2/1	Conv 6
19	Conv 13	8 × 8 × 256	3 × 3/1
20	Conv 14	8 × 8 × 256	3 × 3/1
21	Up-Conv 3	16 × 17 × 256	2 × 2/1	Conv 4
22	Conv 15	16 × 17 × 128	3 × 3/1
23	Conv 16	16 × 17 × 128	3 × 3/1
24	Up-Conv 4	33 × 35 × 128	2 × 2/1	Conv 2
25	Conv 17	33 × 35 × 64	3 × 3/1
26	Conv 18	33 × 35 × 64	3 × 3/1

.

3. Results

3.1. PSSegNet Traning

Based on the designed deep learning network model, the model is trained on the Pytorch framework and the Nvidia RTX 2080 GPU platform. To train the model, 1100 microseismic recordings were collected as a dataset, each manually labeled with the arrival of the P- and S-phases. Based on the position of the calibrated P- and S-phases, each signal can be divided into separate P- or S-phase, and the P- and S-phase masks of each signal can be computed as training labels. During the training process, the model parameters are adjusted using the weighted binary cross-entropy (WCE) loss function, which is defined as

$$WCE(p,\widehat{p})=-[\beta p\log \widehat{p}+(1-p)\log (1-\widehat{p})],$$

(12)

where $p$ is the true mask and $\widehat{p}$ is the mask output by the model. $\beta$ is a positive sample-weighting parameter that reduces false negatives when greater than 1 and false positives when less than 1. Since the distribution of the mask’s own values is not balanced, better results can be obtained by using weighted parameters.

The dataset is randomly divided into training and test sets according to a 10:1 ratio. Based on this dataset, the designed network model is trained. Some hyperparameters still need to be determined before training. In deep learning, model parameters are updated iteratively during learning, and various hyperparameters need to be determined before training. These parameters mainly include learning rate, learning rate decay method, number of training rounds, batch size, optimizer, etc. The learning rate and the optimizer are crucial factors that affect the convergence of the training process. Therefore, in this paper, we discuss the effects of different learning rates and the choice of optimizer, with reference to the choice of optimal hyperparameters.

First, the learning rate of different orders is tested, the convergence rate of different learning rates is tested without considering the accuracy, and the accuracy of the same number of training rounds (epochs) is tested. The test results are shown in Figure 7a,b. It can be seen that among the experimental results of different comparison schemes, the optimal learning rate is 0.0001 with moderate convergence rate and extreme accuracy. Similarly, the optimizer selection method is consistent with the learning rate selection method. Figure 7c,d show the results of the experiment; the RMSprop optimizer performs better and converges faster for the problems investigated in this paper.

The RMSprop algorithm is designed to suppress zigzag descent of gradients, but in contrast to Momentum, RMSprop does not require manual configuration of the learning rate hyperparameters, which is done automatically by the algorithm. More importantly, RMSprop can choose a different learning rate for each parameter. RMSprop is a single iteration for each parameter w, and each iteration is completed according to the following formula:

$${\nu }_t=\rho {\nu }_{t-1}+(1-\rho )\ast g_t^2,$$

(13)

$$\Delta {\omega }_t=-\frac{\eta }{\sqrt{{\nu }_t+ϵ}}\ast g_t,$$

(14)

$${\omega }_{t+1}={\omega }_t+\Delta {\omega }_t,$$

(15)

where $\eta$ is learning rate, $v_t$ is the Exponential Average of squares of gradients, and $w_t$ is the gradient at time t along w.

After the above experiments, the hyperparameters were determined as follows: the learning rate is 0.0001, the learning rate decay method is exponential decay, the number of epochs to achieve full convergence is 500, the batch size is 10, and the optimizer is the RMSprop optimizer. The model is trained based on the selected hyperparameters. The training process is shown in Figure 8. After the model training is completed, a PSSegNet model that can intelligently generate P-phase or S-phase masks is obtained.

3.2. Testing Results

After the PSSegNet model is trained, the test set is used to test the effect of the model. Figure 9a is a microseismic recording from the test set, on which the time–frequency analysis was performed. The time–frequency features are shown in Figure 9b. The real and imaginary parts, respectively, represent the overall characteristics of the signal. After time–frequency characterization (Figure 10a,b) and normalization (Figure 10c,d) are carried out, and then substituted into the trained PSSegNet model, the masks of the P- and S-phases can be obtained, respectively. Finally, the separated P- and S-phases can be obtained through inverse short-time Fourier transform. Normalized inputs can speed up the training of the model and improve the final accuracy.

As a comparison of the training effect of the model, the original sample signals are divided into P- and S-phase after manual calibration. An example microseismic event signal containing P- and S-phases is shown in Figure 11a. In addition, a calibrated mask can be obtained from the P/S-phase mask calculation formula, as shown in Figure 11b,c. The calibrated mask is the criterion for evaluating the accuracy of the masks generated by PSSegNet.

Figure 12 shows the P/S-phase mask obtained after feeding the above example signal into the trained PSSegNet model. It can be found that the overall mask is fundamentally consistent with the actual mask and there are errors at the junction of the two signals. Based on the P/S-phase mask generated by the model and the time–frequency information matrix of the example signal, the respective time–frequency information of the P/S-phase can be obtained. The simple P- and S-phases can be obtained by applying inverse short-time Fourier transform to the time–frequency information of P- and S-phase, respectively, and the results are shown in Figure 13. As can be seen from the experimental results, the generated mask-based PSSegNet model can effectively separate the P- and S-phases from the microseismic recording signal.

The effect of PSSegNet is analyzed and evaluated based on the entire test set. The main evaluation strategy is to calculate the mean square error (MSE) between the predicted mask and the labeled mask and the accumulated error (AE) between the reconstructed P/S-phase and the labeled P/S-phase. Among them, the MSE is calculated as follows:

$$MSE=\frac{{\displaystyle \sum _{N_j}{\left(y_i^{(2)}-y_i^{(1)}\right)}^2}}{N_j},$$

(16)

where N_j is the total number of matrix elements of the mask; $y_i^{(2)}$ is the ith sampling point of the reconstructed signal; and $y_i^{(1)}$ is the ith sampling point of the original signal. The AE is calculated as follows:

$$AE={\displaystyle \sum _{N_c}(y_i^{(2)}-y_i^{(1)})},$$

(17)

where N_c is the length of the original signal.

The results are shown in Figure 14 for mask generation, P/S-phase splitting, and computational errors for all data in the test set. The MSE between the generated mask and the actual labeled mask is concentrated below 2.5, and the AE of the reconstructed P/S-phase based on the predicted mask is concentrated below 1.0 × 10⁻³. Overall, the errors in the reconstruction of both the mask and the P/S-phase are relatively minor. Therefore, it can be argued that the P/S-phase segmentation based on PSSegNet has excellent practical results.

3.3. Generalization

The generality of PSSegNet is analyzed using the P/S-phase arrival picking errors in the test set. An error is defined to be

$$E=\left|p_1-p_2\right|,$$

(18)

where $p_1$ is the picked arrival and $p_2$ is the labeled arrival. Moreover, the results are analyzed at a different signal-to-noise ratio (SNR) and P-phase-to-S-phase ratio (PSR) to determine the robustness of the proposed method. SNR and PSR are defined as

$$SNR=10\times \lg (\frac{E}{E_{noise}}),$$

(19)

$$PSR=10\times \lg (\frac{E_P}{E_S}),$$

(20)

where E is the energy of the signal, E_noise is the energy of the noise, E_P is the energy of the P-phase, and E_S is the energy of the S-phase.

In Figure 15, the errors of the P/S-phase arrival picking are shown separately. It can be found that the overall error of the whole test set is less than 50 ms and most of the errors are less than 20 ms. This precision has excellent performance for P/S-phase arrival picking. In order to additionally test the robustness of our method, data with SNR < 2, 2 ≤ SNR < 3, PSR < 1, or 1 ≤ PSR < 2 in the dataset were selected for arrival picking and their errors were counted. These conditions further interfere with the signal during the arrival picking. The experimental results are shown in Figure 16. In the presence of major signal interference, the arrival time error is still controlled within a limited range, fundamentally within 10 ms for most of the data. It can be argued that this method is relatively reliable for P/S-phase arrival picking.

3.4. Comparison with Other Methods

The current mainstream microseismic monitoring systems and applications mainly adopt the combined algorithm of AR-AIC (Auto Regress Akaike Information Criterion) and STA/LTA (Short Time Average/Long Time Average), namely, the AR-AIC+STA/LTA algorithm. Based on the AR-AIC+STA/LTA algorithm, P/S-phase arrival picking is performed on the same test set. The same error picking analysis was performed, and the results are shown in Figure 17. It can be seen from the picking results that in the P/S-phase arrival picking based on the AR-AIC+STA/LTA algorithm, the P-phase arrival picking is more accurate, but there are more errors in the S-phase arrival picking, with numerous errors greater than 20 ms. In addition, the S-phase arrival picking errors in three sets of test data exceed 50 ms. Therefore, it can be seen that the P/S-phase segmentation based on PSSegNet has an excellent effect during application, as it can effectively reduce the difficulty of obtaining P/S-phase arrivals.

4. Discussion

After the above experiments and analysis, PSSegNet can efficiently segment the P/S-phases. We discuss the successful practice of this approach and explore the theoretical foundations behind it. In previous studies, a large body of literature has demonstrated differences in the frequency distribution of the P/S-phases. In the current manual treatment, experienced researchers find the P/S-phase boundary by looking at the distribution of signal frequencies in the time–frequency domain. Therefore, the time–frequency properties of microseismic signals are taken as input in the design of deep learning methods for intelligent P/S-phase segmentation. Through the powerful data mining capability of deep learning, the mapping of the time–frequency domain to the P/S seismic phase segmentation domain (referred to in this paper as the mask) is realized. PSSegNet successfully constructs this mapping by improving UNet and training using the dataset. This is the theoretical basis for the design of the method in this paper.

The success of PSSegNet in separating the P- and the S-phase is exciting and means that we can segment the components of more complex signals. In this paper, the microseismic signal is reduced to a simple superposition of the P- and S-phases, but the actual microseismic signal also contains additional components such as noise. Therefore, in future studies, we can design and implement a theory and method to segment the entire component of the microseismic signal with reference to the method in this paper. This will be highly effective in improving the intelligent processing of microseismic signals. In addition, the results of the time–frequency analysis of the signal are used as the input to the model and the mask design is used as the output. This approach limits the application of the model. In future research, this formalism can be improved to enable direct input of signals and direct output of results.

5. Summary and Conclusions

Microseismic P- and S-phase segmentation is an influential link that limits the accuracy of microseismic event location, microseismic parameter inversion, and caustic mechanism analysis. Therefore, a deep learning method named PSSegNet is proposed in this paper to intelligently segment the P- and S-phase in microseismic signals. Unlike the previous research (e.g., [48,49,50]) mentioned in this paper, the problem of signal overlapping is treated properly here. The method is designed based on the differences between the P- and S- phases in the time–frequency domain. Time–frequency features based on short-time Fourier transform are used as input to PSSegNet. Then, based on the time–frequency features and artificial labels, a mask matrix of the same size is designed as the output of PSSegNet, which is used to obtain the time–frequency features of the P- and S-phases. Finally, based on UNet, we redesign the network model suitable for this paper based on the input and output dimensions. PSSegNet contains 18 convolutional layers, 4 pooling layers, and 4 deconvolutional layers, and achieves mapping from time–frequency features to segmentation masks via copy and crop.

To verify the effectiveness of PSSegNet for microseismic P/S phase segmentation, we trained and tested PSSegNet on 1100 microseismic records. The predicted mask is compared with the label mask using MSE and AE. The MSE between the generated mask and the actual labeled mask is concentrated below 2.5, and the AE of the reconstructed P/S-phase based on the predicted mask is concentrated below 1.0 × 10⁻³. Overall, the errors in the reconstruction of the mask and the P/S-phase are relatively minor. Arrival picking results show that the overall error of the entire test set is less than 50 ms and most of the errors are less than 20 ms. This precision has excellent performance for P/S phase arrival picking. In order to additionally test the robustness of our method, data with SNR < 2, 2 ≤ SNR < 3, PSR < 1, or 1 ≤ PSR < 2 in the dataset were selected for arrival picking and their errors were counted. The commonly used AR-AIC+STA/LTA algorithm is used as a comparison method. Based on the AR-AIC+STA/LTA algorithm, the P-phase arrival picking is more accurate, but the S-phase arrival picking has additional errors, with a large number of errors larger than 20 ms. In addition, the S-phase arrival picking error exceeds 50 ms for the three sets of test data. Therefore, it can be seen that the P/S-phase segmentation based on PSSegNet has excellent potential for use in various applications and can effectively reduce the difficulty of obtaining the P/S-phase arrivals.