Methods for Identifying Effective Microseismic Signals in a Strong-Noise Environment Based on the Variational Mode Decomposition and Modified Support Vector Machine Models

Abstract

The environment for acquiring microseismic signals is always filled with complex noise, leading to the presence of abundant invalid signals in the collected data and greatly disturbing effective microseismic signals. Regarding the identification of effective microseismic signals with a low signal-to-noise ratio, a method for identifying effective microseismic signals in a strong-noise environment by using the variational mode decomposition (VMD) and genetic algorithm (GA)-based optimized support vector machine (SVM) model is proposed. Microseismic signals with a low signal-to-noise ratio are adaptively decomposed into several intrinsic mode functions (IMFs) by using VMD. The characteristics of such IMFs are extracted and used as a basis for the determination of signal validity. The SVM model is optimized by utilizing GA to obtain the optimal penalty factor c and the kernel function parameter g. The availability of IMF components is judged by the optimized SVM model, based on which the effectiveness of microseismic signals is further identified. By applying the algorithm to the microseismic signals with artificially added noise, the effective microseismic signals and ineffective noise are discriminated, verifying the feasibility of the algorithm. After processing the microseismic records collected in the field, we effectively judge the effectiveness of microseismic signals, suppress the interfering noise in the data and greatly improve the signal-to-noise ratio of the seismic records. The results show that the method for identifying effective microseismic signals based on VMD and GA-SVM can well discriminate between effective and ineffective microseismic signals, which is very significant and provides technical support for microseismic monitoring in a strong-noise environment.

1. Introduction

Microseismic technology, as a method for the efficient monitoring of ground pressure in mines and other dynamic disasters, has been widely used in many engineering fields, such as mines, deep tunnels and side slopes. Microseismic monitoring aims to realize the surveillance of the stable state of rocks by analyzing the microseismic events caused by rock fractures. However, due to the nonlinear and nonstationary nature of microseismic signals, the complexity of the signal acquisition environment and the presence of abundant noise in the signals collected by geophones, the identification of effective signals is very difficult [1,2].

The current microseismic monitoring system for automatic identification has a poor effect regarding the identification of microseismic signals and interfering signals, is easily affected by the characteristics of each signal, and requires human intervention for the identification of effective signals, resulting in low efficiency. In addition, the waveforms of microseismic signals containing noise are similar to those of noise signals. This makes it more difficult to automatically identify effective signals and seriously affects the system’s accurate monitoring of microseismic events, which may further lead to misdiagnosis and a loss of early warning opportunities [3,4].

Therefore, accurately distinguishing effective microseismic signals from abundant signals is particularly important and is a basis for utilizing microseismic technology. In recent years, experts and scholars at home and abroad have made satisfactory achievements in the research on the identification of microseismic signals. Currently, common methods for identifying effective microseismic signals include parametric analysis, waveform analysis, wavelet transform, pattern recognition, etc. But most of these methods require manual processing, which is subject to the unstable classification efficiency and affected by the prior experience of processors [5,6,7].

Among the time–frequency analysis methods, the fast Fourier transform (FFT) and wavelet transform are characterized by multi-resolution image analysis, but their parameter settings are based on practical application scenarios and experience, causing some problems [8]. The empirical mode decomposition (EMD) method can adaptively decompose a signal into intrinsic mode functions (IMFs) step by step, but there are problems of modal aliasing and boundary effects. Improved methods such as ensemble empirical mode decomposition (EEMD) and complementary ensemble empirical mode decomposition (CEEMD) solve some of these problems, but problems such as a heavy calculation burden and incomplete spectral decomposition still exist.

Variational mode decomposition (VMD) [9], as a new non-recursive signal decomposition method, effectively overcomes the problems of modal aliasing and boundary effects that are difficult to be solved by traditional methods such as EMD, and boasts better noise robustness and noise reduction [10]. However, it is still difficult to screen IMF components. Manual screening has a high accuracy but is very slow and costly. Traditional screening methods are fast but have some flaws. Existing methods include alternating direction method of multipliers (ADMM), energy entropy, etc., but parameters need to be set according to different signal conditions, which affects the self-adaptability of VMD and leads to the impossible complete automatic operation of the noise suppression process [11,12].

Genetic Algorithm-based Support Vector Machine (GA-SVM) integrates the advantages of genetic algorithm and support vector machines. By introducing global search through genetic algorithms, GA-SVM searches for better model parameters in the parameter space of SVMs and takes full advantage of the excellent generalization performance of SVMs in the training set to ensure good adaptation to unseen data. It is especially suitable for dealing with high-dimensional and complex classification problems by exploiting the synergy between the powerful performance of SVMs and the global search strategy of genetic algorithms to make the model have a stronger generalization capability [13,14,15].

The GA-SVM method can efficiently solve the problem of parameter-level parallel optimization and adaptively screen appropriate IMF components. In this study, the VMD method and GA-SVM algorithm were combined to efficiently and automatically identify the IMF components generated by VMD. This improves the accuracy of extracting the characteristics of microseismic signals with a low signal-to-noise ratio and presents a new idea for the automatic identification of effective microseismic signals.

2. Methodology and Principles

2.1. Principles of the VMD ALGORITHM

Variational mode decomposition (VMD) is a self-adaptive and completely non-recursive method for mode decomposition and signal processing and can determine the number of modes. Its self-adaptivity is reflected by the fact that the number of modes of a given sequence is determined according to the actual situation, and VMD can adaptively match the center frequency and finite bandwidth of each mode in the subsequent searching and solving processes. It may achieve the effective separation of IMFs and the frequency-domain division of signals, further obtaining the effective components of a given signal after decomposition and finally acquiring the optimal solution of a variational problem. VMD can be divided into the construction and solution of variational problems, which involves three critical concepts: classical Wiener filtering, Hilbert transform and frequency mixing.

(1)

Construction of variational problems

Assume that each “mode” is a finite bandwidth with a center frequency. The variational problem is described as a quest for $k$ mode functions $u_k\left(t\right)$, making the sum of the estimated bandwidth of each mode minimized. The constraint condition requires that the sum of all modes should be equal to the input signal $f$. The specific construction steps are as follows:

• Analytic signals of each mode function $u_k\left(t\right)$ are obtained by Hilbert transform to acquire a one-sided spectrum.
• Mix the analytic signals of each mode and predict the center frequency $e^{-j{\omega }_{k}t}$; modulate the frequency spectrum of each mode to the corresponding fundamental frequency band.
• Calculate the squared norm $L^2$ of the gradient of the above demodulated signal, and estimate the signal bandwidth of each mode. The constrained variational problem is as follows:

  $$\left\{\begin{array}{l}\underset{\left\{{{\displaystyle u}}_k\right\}}{\min }\left\{{\displaystyle \sum _k{{\displaystyle ‖{{\displaystyle \partial }}_t\left[\delta (t)+\frac{j}{\pi t}\times {{\displaystyle u}}_k(t)\right]\times {{\displaystyle e}}^{-j{{\displaystyle \omega }}_{k}t}‖}}^2}\right\}\\ s.t.{\displaystyle \sum _k{{\displaystyle u}}_k}=f\end{array}\right.$$

  (1)

  where $\left\{u_k\right\}=\left\{u_1,\cdots ,u_K\right\},\left\{{\omega }_k\right\}=\left\{{\omega }_1,\cdots ,{\omega }_K\right\}, \sum _k =\sum _{k-1}^K$ [16].

(2)

Solution of variational problems

By introducing the quadratic penalty factor α and Lagrangian operator $\lambda \left(t\right)$, the constrained variational problem is changed into an unconstrained variational problem. The quadratic penalty factor can ensure accurate signal reconstruction in the presence of Gaussian noise. The Lagrangian operator makes the constraints remain stringent. The extended Lagrangian expression is stated as follows:

$$\begin{array}{c}L(\{{{\displaystyle u}}_k\},\{{{\displaystyle \omega }}_k\},\lambda )=\alpha {\displaystyle \sum _k{{\displaystyle ‖{{\displaystyle \partial }}_t\left[\delta (t)+\frac{j}{\pi t}\times {{\displaystyle u}}_k(t)\right]\times {{\displaystyle e}}^{-j{{\displaystyle \omega }}_{k}t}‖}}^2}\\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}+{{\displaystyle ‖f(t)-{\displaystyle \sum _k{{\displaystyle u}}_k(t)}‖}}_2^2+〈\lambda (t),f(t)-{\displaystyle \sum _k{{\displaystyle u}}_k(t)}〉\end{array}$$

(2)

In VMD, the Alternating Direction Method of Multipliers (ADMM) for multiplication operators is used to solve the above variational problem. Alternately update $u_k^{n+1}$, ${\omega }_k^{n+1}$ and ${\mathsf{\lambda }}^{n+1}$ to seek the saddle point for extending the Lagrangian expression [16].

The calculation of $u_k^{n+1}$ can be expressed as follows:

$$\begin{array}{c}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}{{\displaystyle u}}_k^{n+1}=\underset{{{\displaystyle u}}_k\in X}{{\displaystyle \arg \min }}\left\{\alpha {{\displaystyle ‖{{\displaystyle \partial }}_t\left[\delta (t)+\frac{j}{\pi t}\times {{\displaystyle u}}_k(t)\right]\times {{\displaystyle e}}^{-j{{\displaystyle \omega }}_{k}t}‖}}_2^2\right.\\ +\left.{{\displaystyle ‖f(t)-{\displaystyle \sum _i{{\displaystyle u}}_i(t)}+\frac{\lambda (t)}{2}‖}}_2^2\right\}\end{array}$$

(3)

where ${\omega }_k$ is equivalent to ${\omega }_k^{n+1}$; $\sum _i{u}_i\left(t\right)$ is equivalent to $\sum _{i\ne k}{u_i\left(t\right)}^{n+1}$ [16]. Fourier isometric transform (Parseval/Plancherel’s Theorem) is used to convert Formula (3) into the frequency domain. $\omega$ is replaced by $\omega -{\omega }_k$, and the above formula is converted into the form of integrals in the non-negative frequency range. The solution of quadratic optimization problems is expressed as follows:

$${{\displaystyle \widehat{u}}}_k^{n+1}(\omega )=\frac{\widehat{f}(\omega )-{\displaystyle \sum _{i\ne k}{{\displaystyle \widehat{u}}}_i(\omega )+\frac{\widehat{\lambda }(\omega )}{2}}}{1+2\alpha {{\displaystyle \left(\omega -{\omega }_k\right)}}^2}$$

(4)

The value of the center frequency is converted to the frequency domain according to the same process, and the method of updating the center frequency is obtained:

$${{\displaystyle \widehat{\omega }}}_k^{n+1}(\omega )=\frac{{\displaystyle \underset{0}{\overset{\infty }{\int }}\omega {{\displaystyle \left|\left.{{\displaystyle \widehat{u}}}_k(\omega )\right|\right.}}^{2}d\omega }}{{\displaystyle \underset{0}{\overset{\infty }{\int }}{{\displaystyle \left|\left.{{\displaystyle \widehat{u}}}_k(\omega )\right|\right.}}^{2}d\omega }}$$

(5)

where ${\widehat{u}}_k\left(\omega \right)$ is equivalent to the result of the Wiener filtering of the current residual amount $\widehat{f}\left(\omega \right)-\sum _i\widehat{u_i}\left(\omega \right)$; ${\widehat{\omega }}_k^{n+1}$ refers to the focus of the power spectrum of the current mode function; after the Inverse Fourier Transform of $\left\{{\widehat{u}}_k\left(\omega \right)\right\}$, its real part is $\left\{u_k^{n+1}\right\}$ [16].

2.2. GA-Based Optimization of SVM

The use of support vector machines (SVMs) for discrete variables and the introduction into kernel functions require the determination of a penalty factor c and kernel function parameter g. The choice of the factor and the parameter directly affects the accuracy of the classification and generalization capability of SVMs. In the standard form of SVMs, the penalty factor c does not appear directly in a separate formula, but is used as part of a regularization parameter to control the model complexity and avoid overfitting. A larger c value represents a higher degree of fitting training samples. But an excessively high c value may lead to an increased empirical risk and overfitting, while the contrary situation will cause a small penalty for empirical error and underfitting. In this paper, the radial basis function (RBF) is used as the kernel function of SVM. g is a built-in parameter of the function, which implicitly determines the distribution of the data after mapping to a new feature space. The formula is expressed as follows:

$$k\left(x,z\right)=\exp \left(-\frac{{{\displaystyle d\left(x,z\right)}}^2}{2{{\displaystyle \sigma }}^2}\right)=\exp \left(-g{{\displaystyle d\left(x,z\right)}}^2\right)$$

(6)

Formula (6) can lead to

$$g=\frac{1}{2{{\displaystyle \sigma }}^2}$$

(7)

A smaller g value means more support vectors and a faster training and prediction, but an excessively small g value will lead to a poor anti-interference capability of classifiers; conversely, a larger g value represents fewer support vectors and a slower response of classifiers. The value range of g trained in this paper is (0,1000); the use of SVMs for handling discrete variables and the introduction into kernel functions require the determination of penalty factor c and kernel function parameter g. The choice of the factor and the parameter directly affects the accuracy of classification and generalization capability of SVM.

The parameter selection in SVM is, in essence, an optimization issue. At present, the methods for SVM parameter selection mainly include an empirical selection, an experimental trial-and-error method, a grid search, etc. It is difficult for the random selection of the parameters of SVM classifiers to achieve the desired results. Empirical selection and experimental trial-and-error methods can not guarantee the acquisition of optimal values of parameters, and the experimental process highly relies on the operator’s experience. The grid search-based parameter optimization method can help find the optimal solution within a specified range, but to improve the accuracy or to extend the search scope will take more time.

Currently, heuristic algorithms have been successfully applied in parameter optimization. The genetic algorithm (GA), as a global optimization method, has many advantages such as simplicity, parallel processing, absence of the need for explicit mathematical equations and derivative-related expressions, unlikely falling into the locally optimal solution, etc. [17]. In this study, the GA is used to achieve an automatic iterative optimization of SVM parameters. The algorithm begins with an initial population and adopts the self-designed fitness function to evaluate the individual fitness. On this basis, it selects excellent individuals from the population and creates a new generation of individuals by mutation and crossover. This process is iterated until the termination condition is reached. The specific operations are stated as follows:

• Population initialization: A certain number of an initial population is created. The real number encoding is adopted for individual coding. Each individual has two random numbers, constructing a group of SVM parameters.
• Fitness function: The accuracy rate of training samples after cross-validation is used for the fitness function for individual evaluation.
• Selection: The roulette wheel selection method is used to choose operators and ensure a higher probability of high-quality individuals’ being selected. The strategy of optimal individual retention is used, and optimal individuals are directly selected to become members of the next generation.
• Crossover: Crossover or recombination of 2 real-coded genetic loci between individuals is performed.
• Mutation: Select individuals randomly based on the mutation probability to modify values of loci.

After parallel and efficient optimization by genetic algorithm for selection, the optimal SVM parameters c and g can be obtained.

3. Introduction to Algorithms

3.1. Selection of Characteristics

Regarding the extraction of the characteristics of seismic signals (such as geometric, kinematic, kinetic, and statistical characteristics), the characteristics or characteristic combinations of seismic signals that reflect the essential characteristics of such signals and are independent from each other are preferentially selected by using human experience or mathematical methods so as to improve the effect of processing and interpretation methods related to seismic signal [18]. At present, there are many selectable characteristics of seismic signals. The direct use of all these characteristics for signal analysis and related calculations will inevitably cause a larger amount of calculation work. Therefore, it is necessary to select the characteristics that can reflect the essential features of seismic signals and are independent from each other.

A single characteristic index after processing by the GA-SVM model may cause the incorrect identification of components that do not contain effective signals. Several different signal features can reflect differences between effective components and noise components in varied aspects and manifest abundant information of signal components. So, the introduction of a variety of signal features into the model for identification may ensure the accuracy of identification [19]. In this paper, the attributes such as average energy, energy ratio, dominant frequency, skewness, kurtosis, and margin are chosen as the characteristics for the identification of effective IMFs.

(1)

Average energy: the average value of energy. The energy used in this paper refers to the average value of energy from the amplitude of seismic signals. The average energy of noise is generally lower, while the average energy of effective signals is relatively higher. This can distinguish between the two to some extent.

(2)

Maximum energy ratio: the maximum value of the signal energy ratio within two adjacent time windows. Effective signals will experience a significant change in energy within the time window with the arrival of seismic waves, while the overall energy of noise will not show much change throughout the signal duration.

(3)

Dominant frequency: the frequency of dominant waves in the signals. Most noises have a higher frequency than effective signals. So, the features of the dominant frequency can help recognize effective signals from the perspective of the frequency domain.

(4)

Skewness and kurtosis: skewness and kurtosis are dimensionless parameters used to make comparisons with normal distribution curves and reflect the degree of deviation of the signal probability distribution compared with the normal distribution. The third-order central moment of vibration signals is used to calculate the skewness, and the fourth-order central moment is used to calculate the kurtosis. The skewness and kurtosis of effective signals are generally larger than those of noise signals.

The skewness is defined as follows:

$$Skew(X)=E\left[{\left(\frac{X-\mu }{\sigma }\right)}^3\right]$$

(8)

The kurtosis is defined as follows:

$$Kurt(X)=E\left[{\left(\frac{X-\mu }{\sigma }\right)}^4\right]$$

(9)

where μ refers to the mean value of a seismic signal, and σ is the standard deviation of the signal.

(5)

Margin: the ratio of the peak value of a signal to the RMS amplitude, which can help detect the presence or absence of shocks in signals. The margin will increase significantly when effective waves appear [20].

3.2. Algorithm Flow

In a strong-noise environment, the identification of effective microseismic signals becomes more complicated. In order to overcome the interference from noise, we use the VMD technique to decompose a signal into intrinsic mode components (IMFs). Then, by extracting features from each IMF and inputting these features into the SVM model, we are able to determine whether it contains effective components, thus recognizing the whole microseismic signal. The specific processing flow is explained as follows:

• VMD processing of the acquired signal to be processed to obtain separated IMF components;
• Extraction of the characteristics of all IMF components;
• The extracted characteristics are input into a pre-trained GA-SVM model to automatically determine whether each IMF component contains effective information;
• The results of each IMF of a single microseismic signal are considered comprehensively to determine whether the signal is effective.

4. Verification of Algorithm Feasibility

To verify the algorithm’s feasibility, we used microseismic signals with a high signal-to-noise ratio and artificially added noise to obtain a microseismic dataset with a low signal-to-noise ratio and input the dataset into an algorithm model. The noise added in this study included random noise and power frequency interference, which have a frequency of about 50 Hz. Before processing microseismic signals, we first trained the SVM by genetic algorithm for optimization to obtain the best SVM parameters c and g. Then, the microseismic signals in the dataset were processed by VMD to extract the characteristics of each IMF. Effective microseismic signals were labeled as effective components, while ineffective components and those components after noise interference and decomposition were labeled as ineffective ones. Finally, a training set was formed. The training process is shown in Figure 1. The accuracy rate of the algorithm was improved after 100 iterations and finally reached 97%.

By cross-validation, we continuously updated and optimized the SVM parameters and finally acquired the optimal penalty factor c and kernel function parameter g. The number of IMFs from decomposition was determined according to the actual signals and noise. In general, a higher number of IMFs means a better decomposition effect of a signal, but an excessively high number of IMFs will lead to a significant reduction in the computing speed. After testing, five mode components were chosen, which showed better performance in practice.

The algorithm here was used for the processed dataset. Typical signals and processing results are shown in Figure 2 and Figure 3.

The dataset here contained 300 pieces of data, of which 100 pieces of data were effective signals and 200 pieces of data referred to noise. The data here were randomly selected from a large data set of 300 data. Each piece of data was decomposed by VMD into five IMFs and one residual, and there were 1800 sets of data in total. In this study, the number of signal sampling points was 4096, and the signal waveform was a velocity waveform.

The analysis in Figure 2 and Figure 3 reveals that VMD can effectively separate effective signals from the microseismic signals with a low signal-to-noise ratio. In Figure 2, the first three IMFs are ineffective, while the fourth and fifth IMFs are effective. The GA-SVM successfully identifies and reconstructs effective components and obtains the denoised microseismic signals. However, all components after decomposition in Figure 3 are ineffective. SVM failed to find effective IMFs after identification, leading to the absence of signals in the reconstruction results.

In order to evaluate the effect of recognition and processing by the algorithm here on effective microseismic signals, we calculated the signal-to-noise ratio (SNR) and root mean square error (RMSE) of the processed effective microseismic signals. The signal-to-noise ratios and root mean square errors of some typical signals are presented in Figure 4 and Figure 5.

SNR was increased by 19.11~36.62, and SNR was between 37.78 and 52.51 after the algorithm’s processing. RMSE was reduced by 1.65 × 10⁻²~7.96 × 10⁻², and RMSE after algorithm processing was between 0.06 × 10⁻² and 0.64 × 10⁻². Therefore, after the algorithm processing, the signal-to-noise ratio of the signal was greatly improved, and the root-mean-square error was effectively reduced.

The results of the datasets after network training and cross-validation are listed in

Table 1. The results of dataset network training and cross-validation.

Signal Type	Data Quantity	Number of Correctly Identified Components	Accuracy Rate of Component Identification	Accuracy Rate of Signal Identification
Effective components of effective signals	200	196	98%	97%
Ineffective components of effective signals	400	387	96.75%
Components of noise	1200	1174	97.83%	98.5%
Total	1800	1757	97.61%	98%

: the accuracy rate of recognizing effective signals is 97%, the accuracy rate of recognizing noise signals is 98.5%, and the accuracy rate of identifying the signals in the whole dataset is 98%.

5. Field Tests

The field tests were conducted at a test site in Hefei, Anhui Province, China. The presence of a large number of mechanical devices operating continuously in the monitoring environment produced a significant increase in background noise and power frequency interference. When the energy of interference exceeds the pre-set threshold of the system, it is recorded as an event and seriously affects the quality of effective microseismic signals collected by the microseismic monitoring system. In this case, the automatic identification of effective microseismic signals from a huge microseismic data set becomes very difficult. Therefore, it is especially necessary to use the method for recognizing effective microseismic signals with a low signal-to-noise ratio proposed in this paper to accurately identify effective signals.

The algorithm was applied to the microseismic signals collected on-site in the above experimental field to verify the ability of the algorithm to process actual signals. Typical signals and processing results are shown in Figure 6 and Figure 7.

According to Figure 6 and Figure 7, the decomposition of seismic signal maps by VMD can efficiently extract IMFs, regardless of effective signals or ineffective signals. After processing, the trained GA-SVM model was able to adaptively extract the components of effective microseismic signals from IMFs and form microseismic signals with a high signal-to-noise ratio by reconstruction. In contrast, no effective components were recognized in the noise signals during the decomposition process, thus showing an empty set after signal reconstruction.

In order to evaluate the effect of recognition and processing by the algorithm here on effective microseismic signals, we calculated the signal-to-noise ratio (SNR) and root mean square error (RMSE) of the processed effective microseismic signals. The signal-to-noise ratios and root mean square errors of some typical signals are presented in Figure 8 and Figure 9.

As shown in Figure 8 and Figure 9, after processing by the algorithm, the SNR is greatly enhanced, from 23.9~34.55 to 51.31~59.74, and the RMSE is significantly reduced, from 35 × 10⁻⁴~114 × 10⁻⁴ to 1.9 × 10⁻⁴~7.6 × 10⁻⁴, which proves the good recognition and processing of effective signals by the algorithm and verifies the effectiveness of the algorithm.

6. Engineering Application

In order to further validate the applicability of the algorithm, it was implemented in the identification of effective microseismic signals in practical engineering scenarios. The data originated from a microseismic monitoring project conducted in the Jiu Mining area of Hebi City, Henan Province, China. The observation system comprised 16 geophones with a spacing of 60 m between channels. Given the project’s nature of long-term monitoring, a substantial volume of microseismic signal data was collected daily. In this study, the effective signal identification was conducted solely on 15 sets totaling 240 microseismic signals collected over a period of three hours. Both the raw signals and their processed results for one of the sets are depicted in Figure 10 and Figure 11, respectively.

From Figure 10 and Figure 11, it is evident that invalid signals in the data are transformed into non-signals through algorithmic processing, while noise within the valid signals is effectively suppressed. Among the 15 sets of microseismic signals, as is shown in

Table 2. The accuracy rate and recall rate of effective signal recognition of the algorithm.

	Data Quantity	Number of Effective Signals	Number of Noise Signals	Accuracy	Recall
Actual engineering data	240	135	75	98.75%	97.78%
Algorithmic processing result	240	132	78

, there are 135 valid microseismic signals, of which the algorithm successfully identifies 132. The accuracy reaches 98.75%, and the recall reaches 97.78%. These results demonstrate that the algorithm can achieve satisfactory performance in practical engineering applications, effectively accomplishing the task of identifying valid microseismic signals.

7. Discussions

By identifying and processing effective microseismic signals from a large number of original microseismic signals, we confirmed the good effect of the VMD and GA-SVM-based method for the identification of effective microseismic signals. Compared with traditional identification methods, the algorithm proposed in this paper has better adaptivity and robustness, and can adaptively distinguish between effective microseismic signals and noise signals. Moreover, it is able to suppress the noise in effective microseismic signals, thus helping obtain the microseismic signals with a higher signal-to-noise ratio and better quality. Therefore, the algorithm is practical in judging the effectiveness of microseismic signals, which provides a new idea for the effective discrimination of microseismic signals [21,22,23,24,25,26].

Despite its many advantages over similar methods, the algorithm still has many problems to be solved. It involves machine learning-based result classification and prediction and requires the introduction of abundant effective microseismic signals (including effective signals and noise) into the GA-SVM model for training. This means that a certain preliminary work is needed to obtain accurate prediction results. In addition, a special case was found during the study: IMFs of some effective microseismic signals were recognized as ineffective components after the VMD processing, resulting in the identification of such signals as ineffective signals. But this was a rare case and did not greatly influence the accuracy rate of the algorithm. Hence, no further improvements were made for this case. In the future, we may carry out related research to address this issue.

8. Conclusions

For dealing with the difficult identification of effective microseismic signals in a strong-noise environment, a method based on VMD and GA-SVM is proposed. The GA-SVM adaptively identifies effective IMFs generated by the VMD processing of microseismic signals, and, thus, recognizes the effectiveness of signals.

By using the method for the identification of effective microseismic signals proposed in this paper, effective microseismic signals with artificially added noise were successfully recognized. The accurate reconstruction of the signals suppresses the influence of noise and improves the signal-to-noise ratio, which proves that this method can not only recognize effective signals but also effectively deal with strong noise environments in practical situations.

In the context of the strong noise collected in the test field and actual engineering projects, we judged the effectiveness of microseismic signals and successfully distinguished between effective and ineffective components of decomposed signals. In other words, we separated effective signals from ineffective ones and accomplished the identification of the effectiveness of microseismic signals. This would help to discover effective microseismic signals with a high signal-to-noise ratio for the subsequent localization and interpretation of microseismic events.