Recognition of Weak Microseismic Events Induced by Borehole Hydraulic Fracturing in Coal Seam Based on ResNet-10

Abstract

Borehole hydraulic fracturing in coal mines can effectively prevent coal rock dynamic disasters. Accurately recognizing weak microseismic events is an essential prerequisite for the micro-seismic monitoring of hydraulic fracturing in coal seams. This study proposes a recognition method for weak microseismic waveforms based on ResNet-10 to accurately recognize microseismic events generated by borehole hydraulic fracturing in coal mines. To begin with, the background noise and microseismic signals undergo pre-processing through noise reduction and filtering techniques. The preprocessed data are then fed into the ResNet-10 model, and the model parameters are continuously adjusted while the training and test data are updated. The training process stops when the model accuracy rate and loss function value are greater than 99.9% and less than 0.02 for five consecutive times. The model with the highest accuracy rate is then selected to detect the microseismic waveform. The recognition results of ResNet-10 are compared with the threshold value, STA/LTA, and expert recognition results. Finally, the study analyzes flow signal, blasting, and microseismic waveforms. The recognition accuracy rate and recall rate of ResNet-10 are much higher than those of threshold value and STA/LTA, and better than that of the experts. The results of the study show that ResNet-10 can accurately recognize weak microseismic events that are difficult for the threshold value, STA/LTA, and experts to recognize. When water flow signal occurs, it often corresponds to the penetration of hydraulic cracks and the seepage of water. The waveform recognition results demonstrate that the ResNet-10 method has great potential in recognizing weak microseismic waveforms generated by borehole hydraulic fracturing in coal seams.

1. Introduction

Hydraulic fracturing (HF) technology has been widely used to prevent coal rock dynamic disasters, such as rock bursts and coal and gas outbursts [1]. HF involves injecting high-pressure fluid into coal and rock mass, creating fractures in the coal and rock mass under the combined action of earth stress and water pressure to achieve pressure relief [2]. However, the monitoring of the expansion and spatial distribution of HF has not been completely resolved. Rock fracture propagation is a common problem, and the rock fracture process releases energy outward in the form of elastic waves. Microseismic (MS)/acoustic emission (AE) and its related techniques can be used to study the internal structure, stress state, fracture morphology and mechanism of rock mass by using the elastic wave generated by rock fracture [3,4,5,6]. Previous studies have shown that a large number of weak microseismic waveforms are also generated during the HF process in a coal seam [7]. Nevertheless, compared to the surface drilling and fracturing of tight rock strata, such as tight sandstone and shale, HF in coal seams has a relatively small scale and scope. The development of coal joint cracks, soft structure, and stronger heterogeneity make it more challenging to recognize weak microseismic waveforms induced by HF in coal seams [8]. Therefore, accurately recognizing a large number of microseismic events is an essential prerequisite for the microseismic location of HF in coal seams.

The algorithm used for microseismic waveform recognition relies mainly on the characteristics of the signal’s frequency, energy, waveform, and travel time [9]. Expert recognition was the first method used for microseismic waveform monitoring. This approach can precisely recognize waveforms with varying signal-to-noise ratios (SNR), but it becomes inefficient and time-consuming when dealing with a large number of microseismic waveforms [10]. Subsequently, various methods have been proposed for waveform recognition, including short and long window, Akaike information criteria (AIC), and waveform similarity-based template matching methods [11,12,13]. All of these methods have their pros and cons. The long and short window method is effective in recognizing high SNR waveforms, but its accuracy depends on the setting of the short and long window and threshold, making it less effective in recognizing low SNR waveforms [14]. Some methods based on short-time average/long-time average (STA/LTA) have been proposed [15,16], which have improved recognition efficiency and accuracy rates but still cannot meet the requirements of actual detection due to their complicated calculation process and increased computational requirements [17]. AIC is a widely used method for accurately picking up the arrival time of waveforms, but its recognition effect still depends on the setting of the time window and waveform SNR, making it difficult to detect low SNR waveforms [18,19]. Scholars have also proposed some improvement methods based on AIC [20,21]. However, AIC’s recognition effect on a waveform with low SNR is still poor, and the local minimum value of AIC function may vary for different time windows. Additionally, the AIC method can always calculate a minimum point in any time window, regardless of whether the time window contains microseismic waveforms. The template matching method is more effective than the above two methods in detecting weak microseismic waveforms. However, the number of templates used determines its detection accuracy to a large extent, and the calculation amount also increases exponentially with the increase in the number of templates, making it difficult to process massive continuous waveform data rapidly [22]. Recently, scholars have proposed some different waveform recognition methods, such as TFA-DC [23], the fractal feature algorithm [24], and the multi-channel matching tracking algorithm [25], among others. Although these algorithms have their own advantages, they mainly aim to detect various earthquake waveforms, and their effectiveness in detecting borehole HF in coal seams is unknown.

Neural network algorithms are simulations of biological neural networks, which are adaptive nonlinear dynamic network models composed of interconnected neurons [26]. In the 1940s, the first neural network model, the MP model, was developed [27]. This model was a milestone in the development of neural networks. In the 1950s, the single-layer perceptron model was proposed based on the MP model, which marked the beginning of neural networks’ first rise [28]. However, it could not solve the XOR problem. In the 1980s, the Back Propagation Network (BP neural network) was introduced [29] which solved some problems that single-layer perceptrons could not. This event marked the beginning of neural networks’ second rise and was introduced into the automatic recognition of seismic waveforms [30]. In the 1990s, other classification models such as SVM were applied to the field of microseismics and are still in use now [31,32]. With the increase in neural network layers, issues such as overfitting, local optimization, and gradient diffusion began to appear, leading to the slowdown of neural network development.

The 21st century saw a significant improvement in computer functions, leading to the introduction of neural network algorithms and the concept of deep learning [33]. Among the most popular deep learning methods is the convolutional neural network (CNN), which includes classical structures like LeNet, AlexNet, ZFNet, and more. These structures have been improved by scholars [34,35], significantly enhancing the model’s feature extraction abilities. CNN has made great strides in various fields such as image recognition, semantic understanding, and speech recognition [36,37,38,39]. Similarly, microseismic waveform recognition relies mainly on data set selection and parameter optimization, where a large number of samples are used for training, and the weight matrix and bias values of neurons are continuously updated. The loss function is reduced to a minimum, the accuracy rate is increased to a maximum, and an optimal model with strong generalization performance is eventually obtained [40]. In recent years, deep learning convolutional neural networks have been widely used in microseismic detection of earthquakes, oil and gas, rock bursts, etc. [41,42,43]. However, there are few studies on microseismic detection of HF in coal seams.

Microseismic waveforms induced by borehole HF in the coal seam in Xieqiao Coal Mine were studied in this paper. To increase data reliability, enhanced processing was carried out on the data collected by 12 sensors, including noise filtering and expert selection. This resulted in 106,743 background noise and 103,230 microseismic waveforms. A ResNet-10 model was built to recognize microseismic waveforms induced by borehole HF in the coal seam, and it was trained and tested. The experimental results showed that the trained ResNet-10 model had strong generalization ability. Compared to the threshold value, STA/LTA, and expert recognition methods, ResNet-10 can recognize more weak microseismic events, and has better recall rate and accuracy rate. The research results show that using the trained ResNet-10 model can recognize more microseismic events, especially weak microseismic events, which helps to obtain a more complete crack propagation trajectory so as to optimize the HF process and prevent construction risks.

2. Methods

2.1. CNN Model

Convolutional neural network (CNN) is a well-known deep learning neural network which is characterized by weight matrix sharing and sparse connection, leading to lower model parameters and improved operational efficiency. This makes CNN ideal for recognizing microseismic waveforms induced by borehole HF in the coal seam, which has a large amount of data with low waveform signal-to-noise ratio (SNR). The convolution kernel gathers information by recognizing image pixels. Figure 1 below illustrates the convolution process. In the figure, the digit represents the pixel value of each point in the image, while the red part is the input image. The surrounding zeroes represent fill data, so the image size remains unchanged after convolution. The convolution kernel’s depth is determined by the input image’s depth during convolution, and both values are consistent. The number of convolution kernels determines the output image’s depth after convolution, and both values remain consistent. In Figure 1, there are two convolution kernels, and the depth of each convolution kernel is three. There is only one input image with a depth of three, and one output image with a depth of two. The convolution kernel of each layer corresponds to the corresponding position of the image. The first layer’s convolution kernel corresponds to the first layer’s image, the second layer’s convolution kernel corresponds to the lower two-layer images, and the third layer’s convolution kernel corresponds to the third layer’s image. The pixel value above the convolution kernel is dot-multiplied with the pixel value above the specific position of the image, and then added, resulting in the completion of a convolution operation and obtaining the pixel value at the corresponding position of the output image. After that, the convolution kernel continuously moves (usually with a step size of one) until the convolution operation of the entire image is completed.

We used ResNet-10 and ResNet-50 models to train and test the data in the text and found that ResNet-10 had higher accuracy and lowest loss function in training and testing, so we chose the ResNet-10 model.

The training process for this CNN model has two main parts: forward propagation of the waveform and error back propagation [44]. First, the waveforms propagate forward. Prior to training the CNN, initial parameters are set, including weight matrix W^l (parameters of the convolution kernel), Batch value, maximum Epoch, learning rate η, bias value b, and so on. The Batch value is a commonly used unit for updating model parameters, typically set to the NTH power of 2. After repeated debugging, the Batch value for this model is set to 128 based on its training effect. An Epoch means that all training data are trained once. When debugging the Batch value, it is found that the CNN’s accuracy rate stops oscillating around the 25th epoch and remains at an optimal level. To save computing resources, the maximum epoch for this model is set to 57. The learning rate influences the update speed of weight matrix W^l and offset value b^l (l represents the number of layers in CNN). If the learning rate is too large, the training error becomes larger and larger, and there is no trend of fitting. If the learning rate is too small, the gradient may disappear, which means the updating speed of the weight matrix is close to 0. Usually, in the early training, the learning rate is set to be large so that gradient descent can be carried out at a faster speed. In the later stages of training optimization, the value of the learning rate is gradually reduced, which helps the algorithm converge and make it easier to approach the optimal solution. Initial learning rate η of this model is set to 0.0005, and when the test accuracy rate does not improve for three consecutive times, learning rate η is reduced to half of the current value. Offset value b is set to 0.

During the training process of a neural network, a batch of data is extracted from the training dataset. From this batch data, a piece of data ‘x’ is selected, which includes the input data ‘x’ and its corresponding tag ‘y’. These data are then fed into the input layer of the model, and the output parameters of each layer and function of the neural network are obtained.

The output of the input layer is RGB three-channel data ‘x^l’. The output of the convolution calculation or full connection layer is ‘z^l’. The batch normalization function outputs ‘N’, which includes variance (ζ), average value (u), translation (γ), and scaling (β). The activation function σ (relu) outputs ‘a^l’. Lastly, the output of the pooling layer is ‘p’, and its size is ‘p_b’ (where ‘n’ is the image size, ‘f’ is the size of the pooled kernel, and ‘s’ is the step length of the pooled kernel). The maximum pooling layer outputs the maximum value of pixels above the image in each step of the pooling kernel, while the global average pooling layer outputs the average of all pixel values on an image. The output after the activation function and then convolution is represented by ‘z^l’ in Formula (1), and the output after the pooling layer and then convolution is represented by ‘z^l’ in Formula (2).

$${\mathrm{z}}^l=W^l{\mathrm{a}}^{l-1},$$

(1)

$${\mathrm{z}}^l=W^l{p}_{}^{l-1},$$

(2)

$$N^l={\gamma }^l\ast \frac{\left(z^l-u^l\right)}{{\varsigma }^l}+{\beta }^l,$$

(3)

$$a^l=\sigma (N^l),$$

(4)

$$p_b^l=\left(n^l-f\right)/s+1.$$

(5)

During the training process of a model, changing the parameters in one layer can lead to a change in the distribution of the output data. This can cause instability in the distribution of the input data in the next layer, making it difficult for the parameters in that layer to adapt to the new distribution. While data preprocessing can help with the input layer’s distribution, it cannot address the hidden layer’s distribution. This is where batch normalization comes in. Batch normalization normalizes the data distribution at each level to a distribution with a mean of ‘u’ and a variance of ‘ζ’. To prevent the model from becoming worse after normalization, the distribution needs to be shifted by scaling ‘β’ and shifting ‘γ’. When ‘γ’ is equal to ‘ζ’ and ‘u’ is equal to ‘β’, the data are returned to its original state before normalization. Batch normalization enables higher-level inputs to remain relatively stable despite changes in lower-level parameters. This allows gradients the possibility to become more stable, which in turn enables greater learning rates to improve convergence. Additionally, by calculating the mean and variance of the same channel in a batch, all samples in the batch are associated together, preventing overfitting to some extent.

It is common to place the batch normalization function before Relu in a neural network. This is because if Relu is applied first, some neurons may be deactivated during batch normalization, which can cause instability in the function and negatively impact the model’s performance. Batch normalization normalizes the distribution of data back into the unsaturated region, and then the Relu activation function controls how saturated the data is. By following this order, the model can perform more effectively.

First, the error backpropagation calculates loss function C based on output z^l and nominal value y of the neural network. Then, the delta error δ^l is calculated for the output layer using categorical cross-entropy.

$${\delta }^l=\frac{\partial C}{\partial z^l}.$$

(6)

⊙ stands for dot product. The δ^l value of each layer is obtained using the δ^l recurrence formula between adjacent layers. The fully connected layer acts as a "classifier" in the entire convolutional neural network, mapping the "distributed feature representation" to the role of the sample label space. If layer l + 1 is fully connected, then the δ^l error of layer l is obtained.

$${\delta }^l={(W^{l+1})}^T{\delta }^{l+1}\odot {\sigma }^{\prime }(a^l).$$

(7)

ROT180 represents a weight matrix rotated 180 degrees clockwise. If layer l + 1 is a convolution layer, then error δ^l of layer l is

$${\delta }^l={\delta }^{l+1}\ast ROT180(W^{l+1})\odot {\sigma }^{\prime }(a^l).$$

(8)

Upsample indicates the upsampling operation. If layer l + 1 is a pooled layer, then the l-layer error δ^l is

$${\delta }^{\mathrm{l}}=upsample({\delta }^{\mathrm{l}+1})\odot {\sigma }^{\prime }(a^l).$$

(9)

To obtain the derivative of the loss function with respect to the parameters of each layer, we use delta value δ^l of each layer. The derivative of the loss function with respect to the fully connected layer is calculated using Formula (10) while the derivative of the loss function with respect to the convolution layer is calculated using Formula (11).

$$\frac{\partial C}{\partial W^l}=\frac{\partial C}{\partial z^l}\frac{\partial C}{\partial W^l}={\delta }^l{(a^{l-1})}^T,$$

(10)

$$\frac{\partial C}{\partial W^l}={\delta }^l\ast \sigma (z^{l-1}).$$

(11)

When updating the parameters of a neural network, the derivatives of the data in a batch are accumulated and initialized to zero. This process is repeated until the entire batch is trained. The sum of the derivatives is then used to update the parameters using the gradient descent method. To further improve computational efficiency, the ADAM optimization algorithm is applied. Formula (12) represents the parameter update formula.

$$W^l=W^l-\frac{\eta }{batch-size}{\displaystyle \sum _{}^{}\frac{\partial C}{\partial W^l}}.$$

(12)

A ResNet-10 model was developed based on the ‘Le-Net5’ classical CNN model, with the objective of recognizing microseismic waveforms generated by borehole hydraulic fracturing in coal seams. The model comprises 4 blocks and is designed with parameters shown in Figure 1. The convolution kernel size is set to 3 × 3, and each convolution kernel has a batch normalization function layer as well as a ReLU layer. Following the convolution layer, there is either a maximum pooling layer or a global average pooling layer.

In Figure 2, each layer in the ResNet-10 model corresponds to the layer in the model in Figure 3. The top digit of Figure 3 represents the length and width of the feature map of the layer, while the bottom digit represents the depth of the feature map of the layer. As the model contains more layers, the feature maps gradually decrease in size and more of them are required to extract the necessary information. This means that the depth and number of convolutional kernels in Figure 2 need to increase along with the number of layers. After passing through the global average pool layer, the feature map is transformed into a column vector of size 256 × 1. This vector is then transformed into a 128 × 1 column vector through the Relu of the fully connected layer. Finally, the Softmax of the fully connected layer outputs the probability that the waveform segment is a microseismic waveform. Based on this probability, the waveform is classified as either ‘1’ (microseismic waveform) or ‘0’ (background noise).

2.2. Data Collection and Training

Xieqiao Coal Mine is a significant mining site in the Huainan mining area, classified as a coal and gas outburst mine with high gas content and pressure. As per the government’s ‘Rules for the Prevention and Control of Coal and Gas Outburst’, coal mining operations may only begin once the coal seam gas extraction reaches standard levels. To enhance gas extraction efficiency, Xieqiao Mine adopted borehole HF technology, which involves boring through the floor roadway for coal seam pressure relief and permeability improvement. Moreover, a high-precision microseismic monitoring system was installed in the Xieqiao Coal Mine to track the HF in the coal seam. The microseismic system comprises a high-sensitivity wide-band speed microseismic sensor with a 200 v/m/s sensitivity, a frequency band of 4.5 Hz to 1500 Hz, and a system sampling rate of 4 kHz. The sensor is mounted by boreholes to shield background noise. Lastly, to ensure the deep learning ResNet-10 model is effectively trained, a dataset with enough samples covering microseismic waveforms and background noise with various complex characteristics is required.

We monitored the waveform and intercepted a total of 5773 microseismic waveforms, which we marked as 1. Additionally, we recorded 8211 segments of background noise that we randomly selected and marked as 0. To avoid any errors caused by human selection, we double-checked and corrected the selected data. The magnitude of the microseismic events resulting from borehole HF in coal seams is primarily in the range of −3 to 0 (2 to 5.4¹⁰ J), and the duration of microseismic waveforms can reach up to 1.2 s. To include microseismic waveforms of varying magnitudes in waveform segments, we set the length and step size of ResNet-10 sliding window to 1.5 s. To make the dataset more suitable for the deep learning ResNet-10 model, we augmented the dataset with noise and filters using data expansion technology. We added noise to the microseismic waveforms with 12 different gains and applied 5 different degrees of filtering to make the expanded microseismic waveforms and the background noises roughly equal.

The low-pass filter can have five different cut-off frequencies: 400 Hz, 500 Hz, 600 Hz, 700 Hz, and 800 Hz. Adding noise to a waveform is a five-step process that is described as follows:

• In order to generate a sequence of numbers with standard normal distribution, we use the randn function in Matlab (2021 version). The length of this sequence is the same as the length of the waveform segment.
• The series generated by the randn function may not necessarily have a mean value and variance of 0 and 1, respectively. Therefore, we subtract the mean value of the series to reduce the impact of these deviations on subsequent data processing.
• We calculate the db gain of the output signal and the input signal using Formula (15). Next, we multiply this gain by the sum of squares of the waveform segments, and then multiply it by the length of the waveform segments. Finally, we take the reciprocal square root of the result. We perform these operations to ensure that the amount of noise added to each waveform in the segment remains within a reasonable range.
• In order to eliminate the influence of differences between data in the series generated in Step (2), we calculate the standard deviation of the series. We then multiply this standard deviation by each number in the series successively to obtain a new series.
• We divide the result of Step (4) by the result of Step (3) to obtain the noised sequence. Finally, we add this sequence to the waveform fragment to obtain the waveform fragment after noise addition.

$$\mathit{db}=10*{\mathit{log}}_{10}{}^{(P\mathrm{s}/Pn)}.$$

(13)

In Equation (13), P_s represents the output signal’s voltage amplitude, P_n represents the input signal’s voltage amplitude, and db represents the gain. Figure 4 shows a denoising and filtering process for an effective waveform, and Figure 5 shows a denoising process for background noise.

The ResNet-10 method was used to train and test weak microseismic waveforms. After training all the training data once, we tested the data once. The training set consisted of 80% of the expanded data, which included 85,394 background noises and 82,584 microseismic waveforms. The remaining 20% was kept as the test set, which had 21,349 background noises and 20,646 microseismic waveforms. Figure 6 shows that after multiple epochs of training, both the training and test sets had an accuracy of over 99.9%, and the loss function value was below 0.02. This model is suitable for recognizing weak microseismic waveforms. During the initial stage of model training and testing, the accuracy and loss function of the test set fluctuate due to differences between the initial and optimal parameter values of the model (such as the weight matrix and learning rate) and the relatively small amount of data in the test set. However, as the model undergoes continuous training, the model parameters gradually approach the optimal value, and the influence of the small amount of test set data on the model oscillation gradually decreases. By the 26th epoch, the oscillation was significantly reduced, and the accuracy and loss function values of the training and test sets were continually optimized. The model with the best training effect was selected once the accuracy and loss function values were greater than 99.9% and less than 0.02 for 5 consecutive times. The model from the 46th to the 50th epoch, specifically the 47th epoch model, was selected as the actual detection model.

2.3. Detection Procedure

Figure 7 demonstrates the process of recognizing weak microseismic events (low SNR) induced by borehole HF in coal seams using ResNet-10. In the first step, the selected background noise and microseismic waves are extended. The data set is divided into two parts: 80% for training and 20% for testing. If the model’s performance is not satisfactory, the training set is updated, and the model parameters are adjusted. The model is then retrained and tested until satisfactory results are obtained. Once the model’s performance is good, the optimal model is saved and used for microseismic waveform recognition. During the microseismic event recognition process, if more than three channels contain microseismic waveforms in the same time window, it is considered that microseismic events are present in that window. These events are then extracted and saved. However, if less than three channels contain microseismic waveforms in the same time window, the data are discarded.

3. Results

During the detection of continuous half-hour data, 101 microseismic events were recognized using a threshold value, 95 were recognized using the STA/LTA method, 95 were recognized independently by experts, and 100 were recognized using ResNet-10. Based on these results, the expert reviewed and recognized all the real microseismic events within the half-hour time frame. As a result, 130 real microseismic events were recognized. The accuracy (P_e) and recall rates (R_e) of the four methods were compared and analyzed by the experts. Pe and Re are defined as

$$P_e=T_p/(T_p+F_p),$$

(14)

$$R_e=T_p/(T_p+F_n)$$

(15)

In Equations (14) and (15), T_p represents the true example, which means that the microseismic event recognized by the model is the actual microseismic event. If it is a false positive example, then it is represented by F_p. Similarly, T_n represents the true counterexample, which means that the background noise recognized by the model is the actual background noise. If it is a false counterexample, then it is represented by F_n. A high accuracy rate indicates a low false recognition rate, whereas a high recall rate implies that the model misses a low recognition rate. For the model to be practically useful, both accuracy and recall rates need to be high.

Figure 8 shows the statistical results of threshold values, STA/LTA, experts, ResNet-10 methods for identifying effective events. Out of the 101 events that were recognized using a threshold value, only 62 were confirmed to be real, while 39 were wrongly recognized. Meanwhile, 68 real events were wrongly recognized as background noise. When using STA/LTA, 95 events were recognized, of which 80 were real and 15 were misjudged. In total, 50 real events were wrongly recognized as background noise. All 95 events recognized by experts were real events and 35 were recognized as background noise. The accuracy of the threshold value and STA/LTA methods to recognize microseismic events was only 61% and 84%, respectively. However, the accuracy of the expert method and ResNet-10 to recognize microseismic events was 100%. The recall rates of the threshold value, STA/LTA, expert, and ResNet-10 methods to recognize microseismic events were 48%, 61%, 73%, and 77%, respectively. Although none of the four methods could recognize all the real microseismic events, ResNet-10 had the highest recall rate. It was found that ResNet-10 is better than STA/LTA and threshold recognition methods in recognizing microseismic events, and better than expert recognition methods. Due to the poor recognition effect of the threshold value method, it will not be discussed further.

Figure 9 shows a total of 12 microseismic events, all of which were accurately recognized by ResNet-10. However, STA/LTA failed to recognize Event 1, Event 2, Event 7, and Event 10, which were also not recognized by the experts. In contrast, STA/LTA, experts, and ResNet-10 all accurately recognized Event 2, Event 3, Event 4, Event 5, Event 6, Event 8, Event 9, Event 11, and Event 12. Black circle represents the waveform recognized by ResNet-10, red circle represents the waveform recognized by experts, and blue circle represents the effective waveform recognized by STA/LTA.

It has been observed that the number of channels with high SNR microseismic waveforms, in events that have not been recognized by STA/LTA, is usually less than four. In the seventh event, Channels 6, 7, and 10 have microseismic waveforms with high SNR. However, the microseismic waveform in Channel 8 has a low SNR and is not recognized by STA/LTA. Due to this, the event standard of having at least four channels containing microseismic waveforms is not met.

It is often difficult for experts to recognize microseismic events when the number of channels containing high SNR microseismic waveforms is less than three. For instance, in Event 1, only Channel 10 has microseismic waveforms with high SNR, while Channels 1, 8, and 12 have low SNR. Similarly, in Event 10, only Channel 10 has high SNR microseismic waveforms, while Channels 7, 8, and 12 have low SNR. As per the standard, at least four channels should contain microseismic waveforms under the same time window, making it difficult for experts to recognize microseismic events. In Event 2, Channels 6 and 7 contain relatively obvious effective waveforms, which are simultaneously recognized by ResNet-10, experts, and STA/LTA methods. The effective waveforms in Channel 8 and Channel 10 are weak and not recognized by the STA/LTA method.

In events that are recognized by STA/LTA, experts, and ResNet-10, it is often observed that the number of channels containing microseismic waveforms with higher SNR is more than three. For instance, in Event 12, the SNR of microseismic waveforms in Channels 1, 2, 3, and 4 is high. Such events can easily meet the standard of having at least four channels containing microseismic waveforms.

After a comparative analysis by experts, it was determined that the 10 events in Figure 10 were misrecognized as microseismic waveforms by STA/LTA, and the blue circles represent waveforms that are background noises but misrecognized by STA/LTA as effective waveforms. Further review revealed that there was no microseismic waveform present in these 10 events. The waveform in Channel 10 exhibited significant fluctuations, which caused STA/LTA to mistakenly recognize waveforms in Channels 6, 7, 9, and 10 in Events 1–9 as microseismic waveforms.

Event 10 is a signal that interferes with the voltage, and it has only one upward pulse signal. The STA/LTA method wrongly recognizes all channel waveforms in Event 10 as microseismic waveforms. However, it should be noted that the voltage signal can be recognized and eliminated if the short-time window and long-time window length of the STA/LTA are set to be the same. To recognize more weak waveforms induced by borehole HF in coal seam, we tested and found that setting the short-time window and long-time window of the STA/LTA to 60 sampling points and 200 sampling points, respectively, provides the best recognition effect. However, in this case, the voltage signal cannot be recognized.

The figure depicted in Figure 11 displays the 5 min continuous waveform found in Channel 7. The microseismic waveforms are the ones produced by microseismic events. If the waveform in Channel 7 is the only one that is similar to the microseismic waveform, it is not considered a microseismic waveform. The figure contains 20 microseismic waveforms, all of which were recognized by CNN. Among these, the waveform independently recognized by experts is marked with a black circle, which is easy to recognize due to its high SNR. Waveforms that were not independently recognized by experts are marked with red circles. After ResNet-10 recognized the waveforms marked with the red circle, experts amplified and compared them and found that there were indeed microseismic waveforms. This type of waveform has a low SNR and is difficult to recognize.

Expert recognition is usually based on people’s subjective judgment and experience, which may have certain subjectivity and errors. Microseismic wavefroms are often complex and diverse, and expert recognition may have difficulty processing these complex data.

CNN is considered superior in recognizing microseismic waveforms due to the following reasons:

• Capability of feature extraction: Traditional methods require manual design of algorithms which demands a lot of domain knowledge and experience along with repeated attempts to find a good algorithm. However, ResNet-10 can learn the most distinguishing features in microseismic waveforms and automatically extract features without requiring professional knowledge or intervention, thus reducing the difficulty of feature extraction and minimizing human error.
• Ability of pattern recognition: Microseismic waveforms contain complex patterns and structural information, which traditional methods struggle to recognize and understand. In contrast, ResNet-10 has stronger pattern recognition capabilities and can automatically capture pattern and structural information in microseismic waveforms through operations such as convolution and pooling. It gradually transforms this information into higher-level feature representations.
• Adaptability of data: Microseismic waveform characteristics vary due to geological structure, porosity, and other factors. The recognition effect of traditional methods can be limited due to this variability. However, ResNet-10 has good data adaptability and can learn microseismic waveform characteristics under different geological structures and porosity through training. This leads to an improvement in recognition accuracy.

4. Discussion

Figure 12 shows that as water is continuously injected into the coal rock, the water pressure increases, decreases, and then rises again. The rise in water pressure is due to the continuous injection of water into the coal rock while it remains unbroken. The drop in water pressure is caused by the rock breaking and water seeping out. There are two typical water flow signals, Water flow signal 1 and Water flow signal 2. Within 30 min of the water pressure dropping, there were 15 water flow signals. According to the test records, water droplets flowed out of the surrounding rock during the time range of the water pressure drop. This suggests that when water flow signal occurs, it often corresponds to the penetration of hydraulic cracks and the seepage of water.

Figure 13a–c display time-domain pictures of microseismic waveforms. On the other hand, Figure 13f–h represent time-frequency pictures that correspond to Figure 13a, 13b, and 13c, respectively. Figure 13d shows the time-domain image of blasting, while Figure 13i is the time-frequency figure that corresponds to Figure 13d. Figure 13e displays the time-domain picture of water flow signal, and Figure 13j is the time-frequency figure that corresponds to Figure 13e. We selected a display frequency band of 0~250 Hz for the time-frequency figure. The microseismic waveforms induced by borehole HF in coal seam have varying shapes. Figure 13a–c show that the SNR gradually decreases. The duration of the waveforms is also less than 1 s, with waveform duration decreasing continuously with decreasing SNR. The three waveforms have durations of 0.696 s, 0.267 s, and 0.1452 s, respectively, with a main frequency of about 100 Hz.

Figure 13d–i indicate that the duration of the blasting waveform is shorter than that of the microseismic waveforms. This duration is similar to the weak microseismic waveform shown in Figure 13c. The frequency of the blasting waveform is higher, around 400 Hz. This is because the microseismic waveform occurs due to the slow initiation, development, and penetration of microcracks in the coal rock. On the other hand, the blasting waveform has a shorter duration and faster mutation characteristics.

Water flow signal is a type of microseismic waveform that is generated through the friction between water flow with high pressure and rock after a penetrated hydraulic crack is formed. It lasts much longer than microseismic and blasting waveforms, as seen in Figure 13e,j, and can last up to approximately 11.534 s. The main frequency of water flow signal is approximately 75 Hz, which is slightly lower than the main frequency of microseismic waveforms. By using time-frequency transform, the frequency characteristics of the waveforms can be described with the change in time. This allows the waveform dividision into distributed frequency bands, and their frequency characteristics can be accurately quantified through waveform reconstruction.

Figure 14 depicts the energy distributions of three types of waveforms. The blasting waveform has a wide range of frequencies, mainly in higher-frequency bands. On the other hand, microseismic waveforms and water flow signal are primarily distributed in the low-frequency range. Specifically, the blasting waveform energy in the frequency band above 700 Hz is negligible, accounting for only 0.021% of the total energy. The energy of the blasting waveform is mainly distributed in the sub-band of 100~700 Hz, which accounts for 97.2% of the total energy. The energy distribution ratio within 400~500 Hz reaches 35.5%, indicating that this range contains a significant amount of energy.

In the frequency sub-band of 0 to 100 Hz, the proportion of Microseismic waveform 1 to Microseismic waveform 3 gradually increases, while the SNR of Microseismic waveform 1 to Microseismic waveform 3 gradually decreases. The energy of Microseismic waveform 1 is mainly distributed in the sub-band of 0–200 Hz, accounting for 99.1% of the total energy. However, its energy in the band above 200 Hz is very small, only 0.086% of the total energy. The proportion of energy distribution in the sub-band of 0~100 Hz reaches 60.7%. For Microseismic waveform 2, the proportion of energy distribution in each frequency band is roughly the same, fluctuating from 0.95% to 3.38%. In the band above 200 Hz, the proportion of energy of Microseismic waveform 2 is relatively low, accounting for 12% of the total energy. The energy of Microseismic waveform 2 is mainly distributed in the sub-band of 0–200 Hz, accounting for 87.9% of the total energy. The proportion of energy distribution in the sub-band of 0~100 Hz reaches 63.8%.

The energy of Microseismic waveform 3 in the band above 100 Hz is very small, accounting for only 0.079% of the total energy. The energy distribution of Microseismic waveform 3, which has the highest distribution ratio, is mainly in the sub-band of 0~100 Hz, and the energy distribution ratio is 99.9%. The energy of water flow signal is mainly distributed in the sub-band of 0 to 200 Hz, accounting for 97.1% of the total energy. In the sub-band above 200 Hz, the energy of water flow signal is very low, amounting for only 0.029% of the total energy. The proportion of energy distribution in the sub-band of 0–100 Hz reaches 60.6%.

5. Conclusions

The goal of this study is to use a two-dimensional ResNet-10 model to recognize microseismic waveforms caused by borehole HF in coal mines. The ResNet-10 model is compared to the threshold value, STA/LTA, and expert recognition. Additionally, microseismic waveforms, blasting waveforms, and water flow signal are analyzed in both the time and frequency domains. The following conclusions are drawn:

• In terms of accuracy and recall rate, ResNet-10 outperforms the expert and STA/LTA recognition methods. The accuracy rates of STA/LTA, experts, and ResNet-10 in recognizing microseismic events are 84%, 100%, and 100%, respectively. The recall rates of STA/LTA, experts, and ResNet-10 in recognizing microseismic events are 61%, 73%, and 77%, respectively. ResNet-10 can recognize weak waveforms that are difficult for STA/LTA and experts to recognize.
• Within 30 min of the water pressure dropping, there are 15 water flow signals. According to the test records, water droplets flow out of the surrounding rock during the time range of the water pressure drop. This suggests that when water flow signal occurs, it often corresponds to the penetration of hydraulic cracks and the seepage of water.
• In the time domain, the duration of water flow signal is 11.534 s, which is much longer than that of microseismic and blasting waveforms. In terms of time-frequency, the frequency band with the highest energy distribution ratio of the blasting waveform is 400–500 Hz, and the energy distribution ratio is 35.5%. The frequency band with the highest energy distribution ratio of microseismic waveforms is 0–100 Hz, and the energy distribution ratio ranges from 60.7% to 99.9%. The frequency band with the highest energy distribution ratio of water flow signal is 0–100 Hz, and the energy distribution ratio is 60.6%.
• The recognition of more microseismic events, especially weak ones, helps to obtain a more complete fracture propagation trajectory, which can optimize the fracturing process and prevent construction risks. However, how to better match the microseismic waveforms with the model is a difficult task. Next, we will use a more intelligent and more suitable model to complete the classification task of microseismic wavforms.