Shock and Vibration

Volume 2019, Article ID 5803184, 9 pages

https://doi.org/10.1155/2019/5803184

## Automatic Classification of Microseismic Signals Based on MFCC and GMM-HMM in Underground Mines

^{1}School of Resources and Safety Engineering, Central South University, Changsha, Hunan 410083, China^{2}Digital Mine Research Center, Central South University, Changsha, Hunan 410083, China

Correspondence should be addressed to Zhengxiang He; moc.361@eh_gnaixgnehz

Received 24 December 2018; Accepted 28 May 2019; Published 11 June 2019

Academic Editor: Zhixiong Li

Copyright © 2019 Pingan Peng et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

In order to mitigate economic and safety risks during mine life, a microseismic monitoring system is installed in a number of underground mines. The basic step for successfully analyzing those microseismic data is the correct detection of various event types, especially the rock mass rupture events. The visual scanning process is a time-consuming task and requires experience. Therefore, here we present a new method for automatic classification of microseismic signals based on the Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) by using only Mel-frequency cepstral coefficient (MFCC) features extracted from the waveform. The detailed implementation of our proposed method is described. The performance of this method is tested by its application to microseismic events selected from the Dongguashan Copper Mine (China). A dataset that contains a representative set of different microseismic events including rock mass rupture, blasting vibration, mechanical drilling, and electromagnetic noise is collected for training and testing. The results show that our proposed method obtains an accuracy of 92.46%, which demonstrates the effectiveness of the method for automatic classification of microseismic data in underground mines.

#### 1. Introduction

Microseismic monitoring is becoming a popular technology with wide and successful applications in petroleum, mining, and geotechnical engineering [1–3]. In general, the microseismicity is recorded by accelerometers and/or geophones buried in the surrounding rock. Thus, by processing the recorded waveforms, microseismic monitoring can provide important insight into a rock mass and quantify where a certain magnitude of induced rock fracturing is occurring within the volume [4]. With decades of applications, microseismic monitoring plays a vital role in generating valuable information and mitigating economic and safety risks during mine life [5–7]. The system detects all types of vibration events (e.g., rock mass rupture, blasting vibration, mechanical drilling, and electromagnetic noise) in its monitored area. These records are then processed to evaluate the rock mass stability. However, identification of suspicious microseismic events is the first key step of processing microseismic data, which is usually done by experienced analysts through visual scanning of waveforms manually. Thus, microseismic records classification is a time-consuming and tedious task, needs experience, and may suffer from the subjective view of the observers [8]. For these reasons, an automatic technique of identification of event types is in great request.

Throughout the years, many automatic classification methods have been proposed to address this problem in seismic and microseismic fields [9–14]. For example, Scarpetta et al. use neural networks to distinguish volcano-tectonic earthquakes from local seismic signals [15]. Provost et al. propose a random forest supervised classier to identify the type of seismic sources based on 71 seismic attributes [16]. Ruano et al. build a classifier using support vector machines, aiming at distinguishing local and regional earthquakes and explosions from the other possibilities in earthquake early-warning system [17]. Shang et al. propose a hybrid technique based on principal component analysis and artificial neural networks (PCA-ANN) to discriminate between microseismic events and quarry blasts [18]. The PCA-ANN is trained on a dataset with 1600 events, and 22 source parameters are extracted from each event, such as corner frequency, seismic moment, energy, source radius, and static stress drop.

The results of these works are very encouraging because they demonstrate an alternative way to do the tedious scanning work. However, there is still room for further improvement. Most of these classifiers are trained on a large number of parameters [18–21], which are acquired through experienced processing. In other words, these algorithms cannot classify an event unless basic processing procedures (e.g., *P*-wave arrival picking and epicenter location) are done. Thus, these algorithms are not suitable for real-time processing and early-warning system. In this article, we focus on presenting an automatic classification system that requires a minimum amount of training data while enabling to recognize highly variable event patterns. Mel-frequency cepstral coefficient (MFCC) is a widely used feature that has been successfully applied in speech recognition and volcano classification [22]. Hidden Markov Model (HMM) is a powerful tool in modeling any time-varying series [22, 23]. In particular, microseismic records can also be modeled as a time sequence of different microseismic events. Gaussian mixtures are capable of clustering data into different groups as a collection of multinomial Gaussian distributions. Each microseismic signal can be devised as a collection of multinomial distribution and HMM can model the intraslice dependencies between each time period. Therefore, in this research, we propose to utilize Gaussian Mixture Model-based HMM (GMM-HMM) for microseismic signal classification using only MFCC features extracted by waveforms.

The rest of the paper is organized as follows. Section 2 presents the implementation details of automatic classification of microseismic signals based on MFCC and GMM-HMM. In Section 3, we test our proposed method using the field data recorded in an underground copper mine. Finally, Section 4 summarizes the main conclusions.

#### 2. Methods

##### 2.1. MFCC

The first step of the classification process is feature extraction, which converts the microseismic waveform to a parametric representation with less redundant information. Feature extraction consists of choosing those features which are most effective for preserving class separability. Davis and Mermelstein [24] first proposed the MFCC, which is a representation of the short-time energy spectrum of the signal waveform; it is obtained by projecting the logarithmic power spectrum of the microseismic signal onto the nonlinear Mel scale by linear cosine transformation. The transformation relationship between the Mel scale and the frequency is as follows:where *f* is the frequency of the signal. The MFCC is well suited to compensate for signal distortion.

The steps for extracting the MFCC are as follows:(1)High-pass filtering: Passing the waveform of microseismic signal through a high-pass filter can effectively enhance the high-frequency portion, reducing the spectral fluctuation of the signal waveform and allowing any band of the spectrum, regardless of high or low frequencies, to be obtained based on a similar signal-to-noise ratio.(2)Waveform framing: The microseismic waveform is segmented every *N* sampling points to form a new waveform unit, called a frame. According to the requirements of feature extraction and the length of the signal, the value of *N* is generally selected as 256 or 512. In addition, the front and back frames of each frame are intersected with a small portion thereof to avoid an excessive difference between consecutive frames.(3)Adding the hamming window: To increase the continuity between each frame and its adjacent frames after the microseismic waveform is framed, each frame of the waveform is multiplied by the Hamming window. Assuming that the microseismic signal is , multiplying the signal by the Hamming window gives where *N* is the number of frames of the framed microseismic signal.(4)Fast Fourier transform: The difference in energy distribution can represent the features of different signals, so the microseismic signals are converted into an energy spectrum in the frequency domain for comparison. After the microseismic signals are continuously overlapped (50% overlap between successive windows) and framed, the fast Fourier transform is performed on each of the decomposed signal frames to calculate an energy spectrum in the frequency domain.(5)Triangular band-pass filtering: Using a series of triangular band-pass filters *H*_{m}(*k*), the energy spectrum obtained by fast Fourier transformation is converted to the Mel scale to obtain a set of coefficients *m*_{1}, *m*_{2}, …, *m*_{P}. The series of filters are a series of triangular windows that are spaced evenly with overlapping on the Mel-frequency axis.(6)Calculation of the logarithmic energy spectrum: Calculate each filter bank, and take the logarithm of the result. The obtained value is the logarithmic energy, and the logarithmic power spectrum of the corresponding frequency band can be obtained. The calculation formula is as follows:where *X*(*k*) is the energy spectrum of the microseismic signal and *H*_{m}(*k*) is the filter bank, where *m* = 1, 2, …, *P*, *P* is the number of filters and *s*(*m*) is the logarithmic energy.(7)Discrete cosine transform: The discrete cosine transform is used to transform the spectrum from the frequency domain to the time domain. The result is the standard MFCC. The mathematical formula for calculating the cepstral coefficient is as follows:where *n* is the number of frames calculated, with , and *m* is the number of Mel-frequency cepstral coefficients, with .

As the lower order coefficients contain most of the information about the overall spectral shape of the source-filter transfer function, it has become customary in many signal applications to select the first 12 MFCCs because they are considered to carry enough discriminative information in the context of various classification tasks. Consequently, we use 12 features calculated by equation (4) and 12 difference cepstral parameters to form a 24-dimensional feature parameter vector in order to improve the classification performance. The difference cepstral parameters obtained by the difference operation show the variation of the original 12 Mel-frequency cepstral coefficients in the time domain. The method for calculating the difference cepstral parameter *D*_{t}(*n*) is as follows:where *D*_{t}(*n*) represents the *t*-th first-order differential cepstral parameter of the Mel-frequency cepstral coefficient *C*(*n*) calculated in the *n-*th frame of the signal. Θ denotes the time difference of the first derivative in the expression; generally, Θ = 2, and .

Following the above steps, the 24-dimensional feature parameter vector can be successfully extracted from the original complex microseismic signal.

##### 2.2. GMM-HMM

To realize the automatic classification of microseismic signals, the basic procedure is to extract the features of the waveforms by using algorithms and then use these features in combination with machine learning. The classification system presented in this paper is based on GMM-HMM.

The HMM is a probabilistic model of time series. An HMM can typically be represented by five parameters: *λ* = (*N*, *M*, *π*, **A**, **B**), where *N* is the size of the Markov state chain in the HMM and is a fixed value in actual use. Let *N* states be *θ*_{1}, *θ*_{2}, …, *θ*_{N}; the state at time *n* is then *q*_{n}, where *q*_{n} ∈ (*θ*_{1}, *θ*_{2}, …, *θ*_{N}). *M* is the number of observations that may correspond to each state in the Markov state chain. Let *M* observations be *V*_{1}, *V*_{2}, …, *V*_{M}; then, the observed value at time *n* is *o*_{n}, where *o*_{n} ∈ (*V*_{1}, *V*_{2}, …, *V*_{M}). *π* is the initial state probability distribution, and *π* ∈ (*π*_{1}, *π*_{2}, …, *π*_{N}), where *π*_{i} = *P*(*q*_{1} = *θ*_{i}) and 1 ≤ *i* ≤ *N*. The parameter **A** is the state transition probability matrix: **A** = [*a*_{ij}]_{N×N}, where *a*_{ij} = *P*(*q*_{n+k} = *θ*_{j} | *q*_{n} = *θ*_{i}) and 1 ≤ *i*, *j* ≤ *N*, indicating that at any time *n*, if the state is *θ*_{i}, then the probability of the state at the next time instant is *θ*_{j}. **B** is the observed value probability matrix: **B** = [*b*_{ij}]_{N×M}, where *b*_{ij} = *P*(*o*_{n} = *V*_{k} | *q*_{n} = *θ*_{i}) with 1 ≤ *I* ≤ *N* and 1 ≤ *j* ≤ *M*, indicating the probability that the observed value *V*_{k} is acquired at any time *n* if the state is *θ*_{i}.

To better identify complex microseismic signals, the GMM-HMM is constructed with the probability density function of the observed values by using the Gaussian Mixture Model based on the original HMM technique, and *b*_{jk} is modified to the Gaussian distribution probability density function between the current state and observation, that is,where *μ*_{jm} is the mean, *U*_{jm} is the variance, and *c*_{jm} is the Gaussian distribution weight; thus, a GMM-HMM is constructed.

##### 2.3. Classification Strategy: Implementation Details

Our proposed automatic classification method is divided into three major steps:(1)Feature extraction of microseismic signals: In this paper, according to the extraction process in Section 2.1, the microseismic signals are framed, and the Mel-frequency cepstral coefficients of each frame are calculated to obtain 12 Mel-frequency cepstral coefficient values. Then, the differences between the 12 Mel-frequency cepstral coefficients are calculated. As a result, a 24-dimensional feature vector is extracted from a microseismic signal to provide a database for automatic classifier.(2)Training of an automatic classification model: The microseismic signals marked in a certain period of time are selected to extract the feature vectors by the MFCC, and the feature vectors are used as training data set for the corresponding event type. After the model parameters *N*, *M*, *π*, **A**, and **B** are obtained by iterations, the GMM-HMM classifier is constructed. The flowchart for microseismic signal feature extraction and classification is shown in Figure 1.