Applied Acoustics 112 (2016) 131–139

Contents lists available at ScienceDirect

Applied Acoustics

journal homepage: www.elsevier .com/locate /apacoust

Low frequency acoustic signals associated with rock falls,thunderstorms, and wind turbulences in field environment

http://dx.doi.org/10.1016/j.apacoust.2016.05.0210003-682X/� 2016 Elsevier Ltd. All rights reserved.

⇑ Corresponding author.E-mail address: xuqiang_68@126.com (Q. Xu).

Xing Zhu a,b, Qiang Xu a,⇑, Zhiye Zhao b, Tiantao Li a

a State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, Chengdu 610059, Chinab School of Civil and Environmental Engineering, Nanyang Technological University, Singapore 639798, Singapore

a r t i c l e i n f o

Article history:Received 26 November 2015Received in revised form 20 May 2016Accepted 25 May 2016Available online 1 June 2016

Keywords:Empirical mode decompositionRock fallsThunderstormWind turbulenceLow frequency acoustic signals

a b s t r a c t

This paper characterized the observed low frequency acoustic signals generated by rock falls, thunder-storm, and wind turbulence in large rocky landslide. A digital infrasonic recording system was deployedon site to capture real-time low frequency acoustic signals associated with rock falls. An advanced non-stationary signal analysis method, i.e. Empirical Mode Decomposition (EMD), was applied to get insightto the characteristics of the low frequency acoustic signals induced by the hazards. Joint time–frequencydistribution spectra technique was used to detect distinctive features of the events. The study shows thatthe low frequency acoustic signals can be excited by rock falls, thunderstorm and wind turbulence in thefield environment, but the signal varies in both time domain and frequency domain with different pat-terns depending on the physical processes. The results demonstrated that the EMD-based signal process-ing technique is capable of extracting distinctive features to differentiate acoustic signals in realenvironment.

� 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Rock fall and landslide poses significant risk to the mountainousareas strongly affected by large earthquakes. For example, groundmotion induced by the Wenchuan Earthquake triggered tens ofthousands of landslides and rock falls in southwest of China [1].Understanding the failure mechanism of large rock fall is criticalfor assessing the risk, but such events are sometimes difficult toidentify and document [2]. Geo-hazard mapping based on eyewit-nesses often resulted in underestimation in the volume of the dis-placed material, which may not provide accurate feedback forunderstanding the characteristics and mechanism of the event.Therefore, a tool capable of passive monitoring for rock falls, suchas acoustic or seismic technique, yields great benefits in studyingrock fall activities [3]. Zimmer et al. [2,3] deployed seismic andinfrasound sensors in Yosemite Valley for rack fall detection, andthe study focused on seismic data analysis in time and frequencydomain for individual rock fall occurring at short distances(<1 km). Pankow [4] found that instead of a large earthquake trig-gers landslides, a large landslide can trigger several small earth-quakes. Defined as sub-audible or low frequency acoustic wavewith frequencies ranging from about 0.001 Hz to the lowest

frequency of the human hearing limit (20 Hz or cycles/s) [5–7],infrasound has been employed as a passive monitoring and detect-ing technique in many applications, including snow avalanches[8,9], volcanic eruptions [10], and debris flow [11], as it can prop-agates over long distance with little attenuation and the signals arenot sensitive to high noise disturbance in the field. Researches oninfrasound signals generated by natural and artificial sources areavailable in previous literatures [3,12–15]. However, a few studiesfocused on the low frequency acoustic measurements for rock fallsin field, or on the difference between acoustic signals induced byrock falls and other natural sources.

A digital infrasound detecting equipment was implemented atthe north-west cliff of the Daguangbao landslide in China to studythe characteristics of the infrasonic signals associated with rockfalls occurred at a distance less than 2 km, and the results demon-strated the potential for using such techniques for real-time rockfall monitoring. Meanwhile, low frequency acoustic signals gener-ated by thunderstorm and wind turbulences are detected in thereal environment. The advanced digital signal processing and anal-ysis methods are applied to the detected data to distinguish therock fall from other natural sources on the basis of different signalfeatures registered in the time and frequency domain.

This paper characterizes acoustic signals associated with rockfalls, thunderstorm and wind turbulence to demonstrate thepotential applications of low frequency acoustic monitoring

132 X. Zhu et al. / Applied Acoustics 112 (2016) 131–139

technique for real-time rock fall detection and risk assessment incomplex geological conditions. The methodology section of brieflyintroduces the study area, instrument implementation, and datamining methods for signal processing and features extraction.The monitoring dataset section presents some of the original dataand processed results for three natural acoustic signal sources.Future work and concluding remarks are included in the discussionand conclusion sections.

2. Methods for acoustic signal processing and analyzing

2.1. Study site and acoustic monitoring

On 12 May 2008, Wenchuan Earthquake (Ms7.9) hit the SichuanProvince in southwest China. The earthquake was the largest seis-mic event in more than 50 years in China, and the earthquake trig-gered more than 60,000 landslides [16]. The Daguangbao landslidewas the largest landslide triggered by the Wenchuan Earthquake,of which the displaced material converse an area of 7.8 km2 withthe estimated volume of 7.5 � 108 m3 [1,17]. The location of thelandslide and the epicenter of the earthquake are shown in Fig. 1(a). The landslide is located on the hanging wall of the Yingxiu–Beichuan Fault (YBF) in Anxian town with the distance to fault of7 km [17]. As shown in Fig. 1(b), the north boundary of theDaguangbao landslide is characterized by deep and high cliff withan average slope gradient of 65�, and the maximum vertical heightof the cliff is 1200 m varying between 1850 m and 3000 m a.s.l[17]. A long narrow tensile cracks zone along the boundary wasformed by the strong and long-lasting ground shaking. Rock fallsoccurred frequently along the ridge of high cliff because of thenon-stable geological conditions after the earthquake with addi-tional environmental factors such as rainfall, wind and thunder-storm. Base on the field investigation, many talus cones (i.e. rockfalls and rock mass collapses) are distributed in the bottom ofthe cliff with the width of 300–400 m, as shown in Fig. 1(d). Thenorthern boundary of the landslide is an ideal field observation sitefor the preliminary study of the infrasonic signals generated byrock falls because of its remote location with minimal urbannoises. To study the characteristics of acoustic signals associatedwith the geo-disasters, a digital infrasonic monitoring systemwas employed and deployed in the Daguangbao landslide area asshown in Fig. 1(c), detailed description of the system can be found

Fig. 1. Study site and field deployment of the monitoring devices (a) location of the Dinstallation of the infrasonic monitoring system in the field; (d) detached rock blocks in

in previous study [5]. In this study, the sampling frequency of theacoustic monitoring system was 100 Hz. Fig. 1(d) shows thedetached rock blocks during the monitoring period.

2.2. Signal processing and data analysis

2.2.1. Brief review of empirical mode decomposition (EMD)Hilbert–Huang transform (HHT), which was firstly proposed by

Huang et al. [18], is a novel method for nonlinear and non-stationary signal processing. As a key and fundamental part ofHHT, empirical mode decomposition (EMD) method is an effectiveway to decompose a signal into a finite number of intrinsic modefunctions (IMF) with relatively narrow band signals. An IMF isdefined as a function that meets the two requirements [18–20]:(1) the number of zero-crossings and the number of extrema musteither be the same or differ at most by one, and (2) the envelopesdefined by the local maxima and local minima should be symmet-ric. The IMFs indicate the inner oscillation modes of the signal.

For a given discrete time series xðnÞ, n = 1, 2, . . . N (where Nstands for the length of this signal), the flow chart of EMD algo-rithm to decompose a signal into a collect of IMFs is shown inFig. 2.

The decomposition process of EMD is essentially a sifting pro-cess (see Fig. 2). The purpose of EMD is to divide the signal intoa collection of IMF components with a narrower frequency range,and thus make further processing easier [21]. After EMD decompo-sition, the original acoustic signal xðnÞ can be written as the sum ofall the IMF components and the residual as follows:

xðnÞ ¼ C1ðnÞ þ C2ðnÞ þ � � � þ CkðnÞ þ rðnÞ ¼Xk

i¼1

CiðnÞ þ rðnÞ ð1Þ

where Ci is the i-th IMF, and k is the total number of IMF compo-nents. r is the residual that does not meet the aforementionedrequirements of IMF. The residual r may be a constant value or amonotonic function, which represents the direct current (DC) offsetor the average trend of the signal.

Meaningless frequency components in the original signal mayincrease the complexity of decomposition during the EMD process.As a consequence, excessive decomposition result in boundaryerror accumulation, which even sabotage the physical meaning ofthe EMD [22]. Therefore, the original signal should be de-noised

aguangbao landslide; (b) study region and location of the monitoring system; (c)the study area.

Fig. 2. Flow chart of EMD algorithm.

X. Zhu et al. / Applied Acoustics 112 (2016) 131–139 133

by a band pass filter in accord with the field background noise testand analysis result.

2.2.2. Spectral analysisSpectral analysis is commonly used to analyze non-stationary

and non-linear signals in the frequency domain by examining itsglobal energy-frequency distributions. In this study, the acousticsignal observed in the field was typically non-stationary, andtherefore joint time–frequency representation (TFR) method wasemployed to extract the features of the observed signal in the timeand frequency domain. Short-time frequency transforms (STFT) is avery simple and fast TFR algorithm for a discrete time series x½n�,which can be expressed as follow:

stftðm;xÞ ¼X1n¼�1

x½n�w½n�m�e�jxn ð2Þ

where w½n�m� is the shifting sequence window function used todetermine the local sections of the signal at the sample sequencepoint m. The whole signal is divided into several small segmentsand each of them can be assumed to be stationary. Essentially, itis the Fourier transform of x½n�w½n�m�, a complex function repre-senting the phase and the magnitude of the signal over time andfrequency domain. As a result, local features of the signal overtime–frequency domain can be identified clearly. In this study,the STFT was implemented through a time–frequency toolbox(TFRSTFT function) in MATLAB. The ‘hamming’ window with a widthof 127 sample point was selected as the sampling rate was 100 Hz.

2.2.3. Features extractionThe energy ratio and correlation coefficient between each IMF

and the original signal were measured to obtain the inner relation-ship. The energy ratio Epimf ðiÞ can be calculated as follows:

Epimf ðiÞ ¼Eimf ðiÞEtotal

¼PN

j¼1jIMFiðjÞj2PMi¼1

PNj¼1jIMFiðjÞj2

ð3Þ

where N is the length of each IMF,M is the total number of the IMFsfor a given signal, and i is the order of the IMF. The energy of the sig-nificant IMF component typically takes up a great part of the totalenergy so that we can determine the dominant IMF component ofthe original signal and its key features using the data-driven EMDmethod. The correlation coefficient between IMF and the originalsignal is expressed as:

qi ¼COVðIMFi;xÞ

rirxð4Þ

where x represents the original signal, and ri and rx is the standarddeviation of the i-th IMF and the original signal x respectively.

A stringent threshold for discriminating the relevant and irrele-vant IMFs was proposed by using the results of the correlationcoefficient analysis [23]. The threshold was determined as a func-tion of the maximum correlation coefficient and only effectiveIMFs with correlation coefficients greater than the prescribedthreshold were retained. Large coefficient indicates that IMF con-trols and determines the original signal inner oscillation. The func-tion of threshold is shown as follows:

lth ¼ maxðqiÞ10�maxðqiÞ � 3

; i ¼ 1;2; . . .M ð5Þ

where lth is the threshold, qi is the correlation coefficient of the i-thIMF (see Eq. (4)), and M is the total number of IMFs with maxðqiÞrepresenting the maximum correlation coefficient. Therefore, thedata-driven threshold is very important in discriminating the rele-vant and irrelevant modes. The estimation of original signal canbe partly reconstructed based on the threshold as follows:

~x ¼XMi¼1

IMFiðli P lthÞ ð6Þ

where M is the total number of the IMFs, and i is the order of theIMF that meets the boundary condition. To compare the frequencydistributions of different low natural frequency acoustic sourcespresented in this paper, we estimate the root-mean-square pressurepRMSPSD (defined in Eq. (7)) in three frequency bands was 0–5 Hz, 5–

20 Hz and 20–45 Hz, respectively.

ðpRMSPSD Þ

2 ¼Z fmax

fmin

Psdðf Þdf ð7Þ

where f is the frequency; fmin and fmax represent the lower and upperfrequency, respectively, in the aforementioned three frequencybands. Psdðf Þ is the raw power spectral density (PSD) [Pa2/Hz],which was estimated by using the Welch’s period-gram methodfor the reconstructed signal ~x. pRMS

PSD is the root-mean-square pressure[Pa] in the frequency band between fmin and fmax.

3. Monitoring data and result analysis

An infrasound monitoring system was employed and deployedon the north cliff of the Daguangbao landslide area to identify andanalyze the infrasound signals associated with the occurrence ofrock fall events. The monitoring system can capture the events asthe low frequency acoustic signals propagate long distance withlittle attenuation, and thus the rock fall activities can be monitoredin a large area. In this study, we investigated three kinds of low fre-quency acoustic signals (<50 Hz) associated with rock fall, thun-derstorm and wind turbulence respectively. All dataset wereobtained and stored continually by the field monitoring systemwhereas the research team was located about 1 km away fromthe system for 5 days to test and validate the results.

134 X. Zhu et al. / Applied Acoustics 112 (2016) 131–139

3.1. Environmental background noise level

To effectively distinguish the desired acoustic signals, the levelof background noises was first tested and analyzed using the afore-mentioned methods. Fig. 3 shows the background noise and itsanalysis results. Fig. 3(a) shows the EMD results of the backgroundacoustic noise, and the IMFs components to show that the noisesignal fluctuations between short periods and long periods as wellas from high frequency to low frequency. The correlation coeffi-cients and the energy ratios of the IMFs of the observed back-ground noises are depicted in Fig. 3(a), in which IMF1, IMF2 andIMF5-8 are the main energy components of the field backgroundnoises.

According to energy ratio of the IMFs, it is obvious that the IMF1and IMF2 with high frequency and small amplitude indicate signif-icant features and intrinsic oscillation mode of the backgroundnoise. Fig. 3(b) shows the energy distribution in the joint time–frequency domain of the IMF1, of which the key frequency

Fig. 3. Results of processed background acoustic noises in field environment. (a) EMDcoefficient, and Res is the residual item; (b) TFR of the sum of IMF1 and IMF2; (c) PSD o

constantly focused at about 24 Hz for the majority portion indicat-ing the possibility of a data logger induced signal.

Furthermore, Fig. 3(c) shows the PSD estimations of all IMFsdecomposed from background noises. We found that the dominantfrequencies of IMF1 and IMF2 are located in the higher frequencyrange (�24 Hz) as shown in Fig. 3(c), while energy of the remainingIMFs are mainly in the lower frequency range (<4 Hz). The lowerfrequency components may be associated with the variation ofair pressure and/or other unknown sources at normal state. Thesignal presented in Fig. 3 also contains different low frequencyfluctuation components that produces varying numbers of IMFs.

Abnormal acoustic signals can be identified by processing andanalyzing of the background observations in the field so that thelow cut-off frequency of the band-pass filter can be set to 4 Hzfor other abnormal acoustic observations in this area. A goodknowledge of the features of the local background noise is impor-tant to further identification and study of additional abnormalacoustic signals.

components, Er represents the energy ration of each IMF, Coe is the correlationf the IMFs.

Fig. 4. EMD results of acoustic events generated by rock falls (case 1).

X. Zhu et al. / Applied Acoustics 112 (2016) 131–139 135

3.2. Low frequency acoustic signals associated with rock fall events

The empirical mode decomposition (EMD) results of twoobserved acoustic events (case 1) are shown in Fig. 4, which wereinduced by the strong impact of the falling rock blocks. To differen-tiate from the background noises, remarkable acoustic emissionevents with high signal-to-noise ratio (SNR) was identified in thetime series. Additionally, the correlation coefficient and energyratio of each IMF to the original signal are estimated and depictedin Fig. 5(a). It can be found that the IMF1-4 are the relevant com-ponents of the original signal as their correlation coefficients aresignificantly greater than the data-based threshold of 0.16 withtotal energy ratio reached 99%. At the same time, the results ofthe joint time–frequency representation (TFR) analysis for thereconstructed signals are shown in Fig. 5(b). The analysis wasbased on the sum of IMF1-4.

The impact of rock blocks to the ground surface produces strongvibration and cause energy release and transfer. A significant partof the energy propagated in air as an elastic wave, i.e. acousticwave which result in the sound of rock falling in real environment.As shown in Fig. 5(b), the frequency distributions of acousticevents can be clearly recognized in the TFR image. It is of our directinterest that an inaudible sound portion (�16 Hz) appeared in theacoustic events as infrasound with frequency lower than 20 Hz(human hearing limit) in addition to the audible component(�43 Hz).

Fig. 6 presents an observed acoustic event (case 2) associatedwith the detachment, sliding and impact process of a huge rockblock. IMF1-3 (total energy ratio 97.9%) are the relevant compo-nents of the original signal to reconstruct the estimated signalbased on the calculated threshold lth. The difference of the inten-sity of the rock slide, rock fall and ground surface impact can beclearly depicted in TFR as shown in Fig. 6(b). The characteristic fre-quencies of the acoustic signals generated by the friction betweenthe rock block and sliding surface situated primarily in lower fre-quency range as indicated by E3. The dominant frequencies ofthe acoustic event (E4) related to rock fall and ground shockingwere found in both high frequency (�42 Hz) and low frequency(�15.5 Hz) range.

Fig. 7 shows the analysis results of acoustic signals (case 3)produced by a small rock avalanche after rainstorm. As indicated

in Fig. 7(b), the envelope shape of this type of infrasonic event ischaracterized with longer duration time, which is different fromthe acoustic events associated with the transient rock impact (indi-cated on E2 in Fig. 5(b)). The energy distribution in the TFR is com-plex, which is probably the results of the complexity of thephysical process.

3.3. Robust acoustic signals excited by thunderstorm

Similarly, some acoustic events were observed under severeweather conditions in the study area, such as intense thunder-storm, and the signals were analyzed by using the same method.Fig. 8 shows the two detected acoustic events and their featuresassociated with thunderstorm activities. Above 80% energy of theoriginal signals was located on IMF1. In Fig. 8(b) and (d), it wasfound that the waves can be described by an initial sharp increasein pressure, and followed by gradual decline. The maximum valueof the blast pressure in the time domain reached 5 Pa, which wasseveral times greater than the peak value of the acoustic waveemitted by rock fall. The sensor response amplitude level dependson the energy of the source as well as the distance between thesensor and source. However, the acoustic pressure generated bythunderstorm was more powerful and lasted longer than theobserved signals induced by rock fall. From the time frequency rep-resentations (TFRs) as shown in Fig. 8(b) and (d), the spectral con-tent of the acoustic pressure excited by the thunderstorm is mostlylocated in the higher frequency band of 20–40 Hz. Apart from twodominant frequencies (about 24 Hz and 27 Hz) as indicated in bothTFRs, inaudible infrasound components also appeared in the acous-tic waveform induced by thunderstorms.

3.4. Low frequency acoustic signals induced by wind turbulence

Fig. 9 shows the results of the analysis of the acoustic signalsgenerated by thunder blast originated at far distance and thefollow-up wind turbulence event. It is worth noting that the mostpowerful intrinsic mode components of the signal, which is differ-ent from the signals induced by rock fall and thunderstorm, is IMF3rather than IMF1 (as shown in Fig. 9). The acoustic eventsassociated with the thunderstorm can be extracted successfullyby using the EMD as shown in Fig. 9(c). Fig. 10 shows another

Fig. 5. Results of the acoustic events (case 1). (a) Significant IMF components were selected on the basis of the threshold (lth = 0.16) that was calculated by using thecorrelation coefficient (bar-indicated) of each IMF; (b) TFR of the selected IMFs.

Fig. 6. Results of the acoustic events (case 2) associated with rock slide (E3) and ground impact (E4). (a) Significant IMF components were selected on the basis of thethreshold (lth = 0.16) that was calculated by using the correlation coefficient (bar-indicated) of each IMF; (b) TFR of the selected IMFs.

136 X. Zhu et al. / Applied Acoustics 112 (2016) 131–139

acoustic signal that was excited by wind turbulence. Both of theresults indicated that the spectral content of the acoustic waveassociated with wind turbulence is located mostly in lowfrequency band 1–5 Hz with dominant frequency 2.7 Hz.

4. Discussions

This paper aims to study the infrasound signal information as apre-event warning technique for rock fall, landslide and othertypes of geo-hazards. The preliminary results of the analysis onthe infrasound emission characteristics of rock fall in the real envi-ronment is presented here. Additional abnormal low frequencyacoustic events (e.g. acoustic signals generated by thunderstormand wind turbulence) were monitored in the field and analyzedin our study to clarify the differences between field acoustic signalsthat were excited by rock falls and other natural sources. Thisstudy focuses on the low frequency acoustic signals produced byrock fall events occurred on a cliff of the Daguangbao landslide

area. The study area is a rock fall active region with deploymentof infrasonic instruments for rock fall monitoring. Table 1 showsthe root-mean-square (RMS) pressure estimation (pRMS

PSD ) in threefrequency bands 0–5 Hz, 5–20 Hz, and 20–45 Hz of the air pressure(acoustic vibration) induced by the three natural sources (total 12cases). The peak values of the three frequency bands are high-lighted in bold. It demonstrates that the maximum RMS pressurerange in different frequency bands for different natural physicalevents.

The preliminary results demonstrated that rock fall emitted lowfrequency infrasonic signals can be monitored and differentiatedfrom the signals excited by thunderstorm or wind turbulence.The TFR features of the acoustic signals generated by rock fallsmay be more pronounced than those produced by rock avalancheswith high water content after heavy rainstorm. It was observedthat the envelope of the acoustic event of rock fall can be describedby an initial sharp rise in pressure and followed by a sharp decline,while the envelope of the acoustic events associated with wind andthunderstorm may last longer in time domain. The RMS pressure

Fig. 7. Results of the acoustic signals associated with rock avalanche (case 3). (a) Significant IMF components were selected based on the threshold (lth = 0.17), which wascalculated using the correlation coefficient (bar-indicated) of each IMF; (b) TFR of the selected IMFs.

Fig. 8. Results of the acoustic signals associated with two thunderstorm events. (a and c) Indicate the significant IMFs selections from two recorded thunderstorm acousticevents. The corresponding TFRs are depicted in (b and d), respectively.

X. Zhu et al. / Applied Acoustics 112 (2016) 131–139 137

Fig. 9. Results of the acoustic signals associated with a thunderstorm event and the following transient intense wind turbulences. (a) Significant IMFs selection and energyratio distribution; (b) TFR of the selected IMFs. (c) TFR of IMF1 associated with the thunderstorm event; (d) TFR of IMF2-4 associated the with wind turbulences.

Fig. 10. Results of acoustic signals associated with transient large wind turbulence. (a) Significant IMFs selection and energy ratio distribution; (b) TFR of the selected IMFs.

138 X. Zhu et al. / Applied Acoustics 112 (2016) 131–139

Table 1Root-mean-square pressure estimation of observed acoustic events in the realenvironment (Peak value among the three sub frequency bands is highlighted inbold).

Natural sources Cases no. PRMSPSD (Pa)

0–5 Hz 5–20 Hz 20–45 Hz

Wind turbulences 1 0.208567 0.135647 0.078742 0.042426 0.031623 0.033 0.086603 0.05099 0.0223614 0.376431 0.246577 0.07746

Thunderstorms 5 0.261534 0.642262 1.3259716 0.387427 0.862844 1.5348947 0.545711 1.496028 2.610153

Rock falls 8 0.042426 0.194422 0.1462879 0.04 0.100995 0.074162

10 0.024495 0.055678 0.0489911 0.094868 0.147648 0.09539412 0.04 0.100995 0.074162

X. Zhu et al. / Applied Acoustics 112 (2016) 131–139 139

associated with rock falls is mostly distributed in the middle fre-quency band (5–20 Hz) with dominant frequency of 16 Hz, whilethe RMS pressure induced by wind and thunderstorm ranged pri-marily in the lower band (0–5 Hz) and upper band (20–45 Hz),respectively. The characteristic frequency of the acoustic signalsgenerated by wind turbulence is 2.7 Hz, and the dominant frequen-cies of the acoustic signals emitted by thunderstorm are 24 Hz and27 Hz. The analysis shows that the low frequency acoustic signalscan serve as an approach to monitor unstable rock activities,because the signal is capable of propagating over a long distancewith little attenuation and can be differentiated from other naturalnoises as demonstrated by the results herein. However, largerdataset are required to better understand the relationship betweenthe acoustic signals and the mechanism and dynamics of rock falls.The results show that EMD method is suitable for processing suchacoustic signals associated with complex natural sources.

5. Conclusions

This paper characterized three infrasonic events associated withrock fall (including slide and ground shock), wind turbulence, andthunderstorm. Unique low frequency acoustic signals wereobserved due to rock movements in the Daguangbao landslidearea. The nonlinear signal processing method EMD was appliedto investigate the significant components of the observed signalto minimized the influence of field background noises. Short-time Fourier transform was used to distinguish the spectralfeatures of the vital infrasonic components in the joint time–frequency domain. Meanwhile, every type of low frequency signalis of different features in both time and frequency domain due todifferent physical processes. The research results highlight thepotential application of such field instrument for more accuraterock fall activity assessment. It was concluded that EMD-basedsignal processing technique successfully extracted the features ofdifferential acoustic signals in the real environment.

At present, it is difficult to locate the events based on infrasoundsignal processing and analyzing only, as the transmission of infra-sound signals in the air may be disturbed by temperature vibrationand/or wind activities. Previous studies shown that seismic andinfrasonic signals of huge mass movements are correlated andcomplementary [11,24]. The combination of low frequency acous-tic signals and seismic signals presents great potential in fieldapplication. Seismic signal exhibits strong dependency on geologi-cal condition with high attenuation over distance, however, it ismuch less sensitive to wind and weather disturbances. Meanwhile,infrasound signals are featured with little attenuation in air but

subject to disturbances of background noises induced by wind orsevere wheatear condition. Overall, our results show that low fre-quency acoustic monitoring is effective for detecting rock falls aswell as differentiating the event from other acoustic sources. Thecombination of infrasound and seismic signal analysis providesgreat approach for process identification, automatic detectionand pre-event warning of rock fall or landslide in our future work.

Acknowledgments

This study was financially supported by the National NaturalScience Foundation of China (Grant No. 41502293) and the MajorState Basic Research Development Program of China (973 Pro-grams, Grant Nos. 2013CB733200, and 2014CB744703).

References

[1] Huang RQ, Li WL. Post-earthquake landsliding and long-term impacts in theWenchuan earthquake area, China. Eng Geol 2014;182(Pt B):111–20.

[2] Zimmer VL, Collins BD, Stock GM, et al. Rock fall dynamics and deposition: anintegrated analysis of the 2009 Ahwiyah Point rock fall, Yosemite NationalPark, USA. Earth Surf Proc Land 2012;37(6):680–91.

[3] Zimmer VL, Sitar N. Detection and location of rock falls using seismic andinfrasound sensors. Eng Geol 2015;193:49–60.

[4] Pankow KL, Moore JR, Hale JM, et al. Massive landslide at Utah copper minegenerates wealth of geophysical data. GSA Today 2014:4–9.

[5] Zhu X, Xu Q, Zhou JB, et al. Experimental study of infrasonic signal generationduring rock fracture under uniaxial compression. Int J Rock Mech Min2013;60:37–46.

[6] Johnson JB, Anderson JF, Anthony RE, et al. Detecting geyser activity withinfrasound. J Volcanol Geoth Res 2013;256:105–17.

[7] Dabrowa AL, Green DN, Rust AC, et al. A global study of volcanic infrasoundcharacteristics and the potential for long-range monitoring. Earth Planet SciLett 2011;310(3–4):369–79.

[8] Havens S, Marshall HP, Johnson JB, et al. Calculating the velocity of a fast-moving snow avalanche using an infrasound array. Geophys Res Lett 2014;41(17):6191–8.

[9] Thüring T, Schoch M, Van Herwijnen A, et al. Robust snow avalanche detectionusing supervised machine learning with infrasonic sensor arrays. Cold Reg SciTechnol 2015;111:60–6.

[10] Matoza RS, Fee D, Garcés MA, et al. Infrasonic jet noise from volcanic eruptions.Geophys Res Lett 2009;36(8).

[11] Chou HT, Chang YL, Zhang SC. Acoustic signals and geophone response ofrainfall-induced debris flows. J Chin Inst Eng 2013;36(3):335–47.

[12] Ulivieri G, Marchetti E, Ripepe M, et al. Monitoring snow avalanches inNorthwestern Italian Alps using an infrasound array. Cold Reg Sci Technol2011;69(2–3):177–83.

[13] Kogelnig A, Hübl J, Suriñach E, et al. Infrasound produced by debris flow:propagation and frequency content evolution. Nat Hazards 2014;70(3):1713–33.

[14] Green DN. Seismic and infrasonic signals associated with an unusual collapseevent at the Soufrière Hills volcano, Montserrat. Geophys Res Lett 2005;32(7).

[15] Laštovicka J, Baše J, Hruška F, et al. Simultaneous infrasonic, seismic, magneticand ionospheric observations in an earthquake epicentre. J Atmos Solar-TerrPhys 2010;72(16):1231–40.

[16] Huang RQ, Fan XM. The landslide story. Nat Geosci 2013;6(5):325–6.[17] Huang RQ, Pei XJ, Fan XM, et al. The characteristics and failure mechanism of

the largest landslide triggered by the Wenchuan earthquake, May 12, 2008,China. Landslides 2011;9(1):131–42.

[18] Huang NE, Shen Z, Long SR, et al. The empirical mode decomposition and thehilbert spectrum for nonlinear and nonstationary time series analysis. ProcRoyal Soc London A 1998;454:903–5.

[19] Wu W, Peng H. Application of EMD denoising approach in noisy blind sourceseparation. J Commun 2014;9(6):506–14.

[20] Li SF, Zhou WD, Yuan Q, et al. Feature extraction and recognition of ictal EEGusing EMD and SVM. Comput Biol Med 2013;43(7):807–16.

[21] Li M, Wu X, Liu X. An improved EMD method for time-frequency featureextraction of telemetry vibration signal based on multi-scale median filtering.Circuit Syst Sign Proc 2014;34(3):815–30.

[22] Wang FJ, Li XY, Jiang TM. Life feature extraction based on hilbert marginalspectrum entropy for ADT vibration. Proceedings of the 2012 internationalconference on information technology and software engineering, vol. 211. p.37–46.

[23] Albert AP, Nii AO. A criterion for selecting relevant intrinsic mode functions inempirical mode decomposition. Adv Adap Data Anal 2010;2(1):1–24.

[24] McNutt SR, Thompson G, West ME, et al. Local seismic and infrasoundobservations of the 2009 explosive eruptions of Redoubt Volcano, Alaska. JVolcanol Geoth Res 2013;259:63–76.