Nanoseismic.Monitoring.fills.the.gap.pdf

7/28/2019 Nanoseismic.Monitoring.fills.the.gap.pdf

1/8

Leveraging Technology

2008 EAGE www.firstbreak.org 117

special topicfirst breakvolume 26, June 2008

Monitoring seismic aftershocks in a 1000 km2 search

area is a key investigation method for the planned

on-site-inspections (OSI) of the Comprehensive

Test Ban Treaty Organization (CTBTO) when

searching for potential, nuclear underground explosions

(/link1/, Zucca et al., 1996). The inspections will take placeweeks after any suspicious event, and aftershocks are expect-

ed to be rare and weak, if any. Thus the political demand is

a completeness threshold ML

2.0 for seismic signal moni-

toring which translates to low-SNR seismograms under any

surface site noise condition.

Geophysics derives plentiful information from moni-

toring fracture processes, e.g., the global earthquake cov-

erage unveils the borders of plate tectonics, regional and

local studies quantify seismic hazard, and resolve Benioff

zones of subducting plates. Near-source borehole stations

map the fracture growth in hydraulic fracturing, and ultra-

sonic, piezo transducers resolve material failure in non-destructive testing. In common for all these applications, a

sufficient SNR of the recorded seismograms allows for une-

Nanoseismic monitoring fills the gap betweenmicroseismic networks and passive seismic

Manfred Joswig* introduces the concept of nanoseismic monitoring as a third method (micro-seismic networks and passive seismic being the others) to record upper Crust or surface layerfracture signals, and to locate them in space and time. First results of low-SNR aftershockmonitoring for nuclear arms control purposes demonstrate the method and its potential.

quivocal onset phase determination, being followed by the

non-linear, iterative Geiger approach for hypocentre cal-

culation. Fig. 1 gives an example for hydrofrac monitor-

ing, and exhibits sample seismograms of sufficient SNR.

Table 1 summarizes the typical characteristics of microseis-

mic networks, e.g., according to Lee and Stewart (1981),while details for the Geiger inversion may be found in many

seismological textbooks, e.g., Lay and Wallace (1995). If

source processes turn small, monitoring distances must

shrink respectively. The recorded signal frequencies scale

accordingly, to tens of kHz for acoustic emission, and hun-

dreds of kHz for non-destructive testing.

The need for borehole stations in hydrofrac monitoring is

an obvious example of the limits of microseismic networks,

which motivated the search for possible alternatives. Passive

seismic is an emerging new technique that offers the chance

to record hydrofracs from surface stations. This demands a

large number of stations, usually available by the dual use of4D seismic layout operations for the instrumented oil field.

Fig. 2 gives an example from Kochnev et al. (2007), while

microseismic

networks

nanoseismic

monitoring

passive seismic

typical application

area, completeness magn.

#stations (typical)

select analysis segment

noise forensics

perm. local network

10000 km2 ML

1.0

30 single 3c

STA/LTA & voting

optional

temp. fault mapping

100 km2 ML

-1.0

3 SNS arrays

Sonograms & PR

essential

instrumented oil field

1 km2 ML

-3.0

100+ array traces

(continuous)

none

Signal-to-Noise Ratio

status of onset phases

process solution

test/improve solution

> +15 dB (5:1)

clear

pick all batch

new run

> 0 dB (1:1)

questionable

pick live update

slide any parameter*

> -15 dB (1:5)

not visible

automated stack

(not applicable)

solution info

improve by master event

identify effects of single

parameter* to joint solution

hypo, t0, M

L, M

possible

indirect by time residuals to

LMS solution

hypo, t0, M

L

possible

fully resolved in location

domain by jack-knifing

statistics

(not applicable)

(not available)

* phase picks, forced depth, layer model

Table 1 Properties of microseismic networks, nanoseismic monitoring, and passive seismic.

*Universitt Stuttgart, Germany, E-mail: [email protected]


2/8


www.firstbreak.org 2008 EAGE118

special topic first breakvolume 26, June 2008

of moment tensor M for resolution of fracture orientation

and dominant force regimes. The needs for nanoseismic

monitoring rose by the demands of OSI for CTBTO when

ML

2.0 events must be discovered in a 1000 km2 search area,

with the logistics of some 10 crew members working for a

few weeks in the field. The work plan includes near real-time

processing and potential direction of further inspections teams,

e.g., radionuclide or visual inspection. Thus nanoseismic moni-

toring will be presented here along with examples performed

during the directed exercise DE04 of CTBTO in 2004.

However, nanoseismic monitoring, acting like a moderate-

effort seismic microscope of previously unavailable sensitivity,

offers many new chances to resolve ambient fracture process-

es. Wust-Bloch and Joswig (2006) derive a process identifica-

tion of pending sinkhole collapses at the Dead Sea, Hge and

Joswig (2008) report on improved resolution of seismicity dur-

ing an inter-swarm period at Vogtland, Czech Republic, andWalter and Joswig (2008), in a forthcoming issue ofFirst Break,

describe the first-time discovery of cracks in a creeping, clayey

landslide during heavy rainfall in Vorarlberg, Austria.

DE04 of CTBTO was conducted on a military camp ground

in Stupava, Slovakia (Fig. 3). Its purpose was to determine the

distances at which weak aftershocks could be detected in the

ambient seismic noise. The OSI source target of aftershocks

from an underground, unclear explosion was mimicked by a

series of small explosions with 75, 150, 200, 400, 475, and 600

g explosive, respectively. Lines of single, three-component (3C)

geophones were laid out at 500 m spacing each. Three small

arrays, so-called seismic navigating stations (SNS) were sited at1.0, 1.5, and 2.5 km distances from the explosion site. Each SNS

consists of an eight-channel data logger, a central, 3C geophone,

and a tripartite array of vertical sensors centred around the 3C

site as an equilateral triangle with some 100 m aperture; two

Table 1 summarizes the relevant properties. The most nota-

ble restriction is the limitation to just statistical descriptions

of fracture energy release, i.e. no single-event bulletin will

be available.

Nanoseismic monitoringAnother significant alternative is presented here - nano-

seismic monitoring with its properties according to Table

1. Like passive seismic, it resolves events in much poorer

SNR than microseismic networks; like the latter it produces

individual event bulletins that just lack the determination

Figure 1 Microseismic network monitors hydrofrac signals at Ekofisk (compiled

from Oye and Roth, 2001; 2003). Despite the unusual station layout in a VSP

string, the application features the essential characteristics of microseismic

networks which process station signals of sufficient SNR for individual onsetphase determination and polarization analysis.

Figure 2 Passive seismic monitors hydrofrac at 2445 m from surface (from

Kochnev, 2007). Note the colour code for scaled amplitude related to relative

energy release of fractures per unit time interval of stacking. Individual event

seismograms could not be observed at surface records.

Figure 3 The Stupava site of DE04. Explosions were detonated by army staff

at red star. Layout of single 3C geophones was in to lines (1-8) and (A-G), SNS

small arrays were installed at red triangles.


3/8




mands speed up the interaction. Each observational parameter,

like onset time or amplitude, may be set by mouse click, and

continuously shifted by arrow keys. All derived constraints

are displayed and updated in real-time. Likewise, all results of

event processing, like epicentre, depth, origin time, magnitude,

array back-azimuth, and slowness, may be set or shifted with anaccompanying, real-time update of all simulated observations,

e.g., theoretical phase onset times, or maximum amplitudes.

The same fully interactive scheme applies for the intrin-

sic parameters of event location, like velocities and layer thick-

ness of different underground models, or the magnitude-dis-

tance correction curve. The spirit of interactive event loca-

tion may best be explored by downloading the software

(/link2/), with the related data set of DE04. The purpose of inter-

action is to test and play the many different possibilities of phase

identification which exist due to the poor SNR, to check poten-

tial solutions on their parameter plausibilities, and to explore

auxiliary channels may record air pressure and wind speed (notavailable at DE004). Fig. 4 shows how the weakest explosion of

75 g yield is detected, or missed depending on the local weath-

er conditions, by the single 3C geophone at shortest distance of

500 m. Fig. 5 summarizes the results in a distance-dependent

detection threshold; in the case of a 500 m sensitivity per site,

one would need some 1000 network stations to cover the 1000

km2 search area. In contrast, a single SNS performs well until 3

km yielding a total of 30 SNS to be distributed in the search area

to reach the ML

2.0 monitoring threshold.

Principles of operationHow does nanoseismic monitoring perform so well? The results

cannot be explained just by the utilized array approach: thegain in SNR is a mere 6 dB by stacking the four vertical traces,

and the sparse spatial sampling does not allow for any kind of

f-k analysis. The key contribution for success is via an innova-

tive, highly interactive software approach focusing on a real-

time updated display of hypocentre location constraints, plus

sophisticated diagnosis tools for human event analysis including

noise forensics, instead of automated detections by off-the-shelf,

sub-optimal STA/LTA approaches. The software tools are imple-

mented in the event analysis program HypoLine: Fig. 6 shows

a representative screen dump from a single-SNS campaign. A

fixed-frame layout eases orientation, and single keystroke com-

GlossaryMacroseism: Intelligence investigation on earthquake dam-

age which results in an earthquake intensity or Mercalli

scale 1-12, similar to Beauford scale for wind speed.

Microseism: Dominant peak of Earth noise in the range of

5-15 Hz caused by ocean waves and weather turbulenc-

es, known since the early days of seismology (Guten-

berg, 1931).

Microseismicity: Earthquakes below the level of human

sensitivity, say ML

3.0, recorded locally (within 100 km)

or at regional scale (up to 3000 km).

Microseismic network: Distribution of geophones at local

or regional scale to locate and identify seismicity; may

also be applied for acoustic emission and non-destruc-

tive testing.

Nanoearthquake: Suggested phrase for earthquakes belowM

L0.0; however, not yet commonly accepted by the

seismological community (Butler, 2003).

Nanoseismic monitoring: Location and identification of

low-SNR fracture processes, e.g., nanoearthquakes by

jackknife analysis of tripartite array networks.

Acoustic emission: Short distance recording of fracture sig-

nals at high frequencies, e.g., in mines. Magnitudes may

well reach ML

3.0.

Non-destructive testing: Laboratory-scale experiments for

sample deformation, records fracture progress until the

eventual probe failure.

Passive seismics: Location of energy release from fractureprocesses by means of seismic exploration-like equip-

ment and software tools, e.g., in the concept of instru-

mented oil fields.

Forensic seismology: Location and identification of non-

seismic sources by seismic networks, e.g., airplane

crashes, submarine explosions (Zucca, 1998).

Figure 4 Explosions of 75 g were recorded at nearest 3C geophones (A, 1 of Fig.

3) in 500 m distance (compiled from Labak et al., 2005). Depended on different

ambient noise conditions, the signal was either recognized (top), or masked

completely (bottom).


4/8




displays the time segments of Fig. 6 as sonograms which help to

guide the phase picks, and gives the location results; the event

has ML 2.1 in 1.4 km (slant) distance.

Jackknifing for robustnessA central role for processing weak signals is in the realization of

outlyer-resistant statistics. For event location, we have selected

the approach of jack-knifing (see box) that can identify the con-

tribution of single, erroneous parameters to the joint, averaged

solution. For this, all hypocentre-relevant information is broken

down into graphical location constraints; for the ideal solution

they all would meet in a single point.

Lets first start with the simple case of surface stations, and

a homogeneous half-space model for the underground. Then

the multitude of similar quality event solutions, e.g., close to dif-

ferent layer boundaries for the hypocentre depth determination.

Within a few minutes, the analyst may test on hundreds of alter-natives to locate and identify an event which is just marginally

above the ambient noise level.

Event detection and type diagnosis are supported by f-t sig-

nature analysis in sonograms (Fig. 7), where sonograms can be

understood as self-adaptive, optimum filters of non-linear ener-

gy display based on spectral estimates of noise median and frac-

tile variance for noise muting (Joswig, 1990, 1995). These signal

representations may be used for automated pattern recognition

in standard observatory work; however, the spirit of OSI is the

search for the single, suspicious exception which prohibits the

restriction of routine processing by simple scanning tools. Fig. 8

Figure 5The curves for distance-dependent detection thresholds summarize

the results of DE04. The single station approach by 3C geophones demands less

than 500 m inter-station distance to ensure detection of ML 2.0 events. Thistranslates to more that 1000 stations necessary to cover the 1000 km2 search

area of OSI. Tripartite SNS arrays perform close to 3 km reducing the number

of systems to some 30 units for full area coverage.

Figure 6 Screen layout of analysis software HypoLine showing a candidate

event at the threshold of processing capabilities. The seismograms were

acquired by the four SNS stations sketched in the zoom map; optimum filters

to enhance SNR were already applied.

Figure 7Seismogram, related power spectral densitiy (psd) matrix, and sono-

gram. The psd matrix is obtained by sliding FFT, and binned logarithmically for

frequency and amplitude. The sonogram adds pre-whitening and noise mut-

ing, and clearly enhances the display of weak, short-term signal energy.

Figure 8Processing results for the candidate event of Fig. 6. The sonograms

guide the phase picking for the four weak onsets, the jackknifing gives four tri-

ple junctions (red dots in the zoom map). For the adjustment of the epicentre,

additional information from the tS-t

Pcircle (dotted green circle segment) and

the two array beams for P and S onset (yellow fans) is considered. The event

has ML

2.1 in 1.4 km (slant) distance.


5/8




ure for the picked phase onsets. For extreme cases, symmetry

will cause two triple points, or too large time differences pro-hibit the existence of hyperbolae at all. In the spirit of a jack-

knife solution, the spread of triple points, instead of hyper-

bolae smear, will characterize the potential contradictions in

the parameter space of an over-determined equation system,

i.e. more than three tP

station readings.

For any number of stations N 3 with P onsets tP, the maxi-

mum number of hyperbolae His by Eq. (1)

the time difference tS-t

Pat any single station yields a semi-sphere

with constant radius that constrains the underground source

location in 3D spatial coordinates (Fig. 9). From now on, we

will take the estimate of hypocentre depth as an external param-

eter that is modified by the analyst, instead of being determined

by data inversion. The rational for this decision is that vari-

ations in the estimate of epicentre (x,y) relate to time incre-

ments of opposite sign at any station layout of a comprehen-

sive spread; thus a clear minimum of residual times for iterative

or grid-search location procedures can be identified. Change of

depth, on the other hand, will affect all surface stations by time

increments of equal sign, and the same effect may also indicate a

mere shift of origin time. Thus depth determination depends on

second order residuals, which makes source depth poorly con-

strained in most seismic bulletins.

For depth as an external parameter, each semi-sphere reduc-

es to a circle, as the intersection of sphere and the plane of depthconstraint. Likewise, any t

P-t

Pdifference between two distinct

station onset times describes a semi-hyperboloid that reduces

to a hyperbola by intersection with the depth-plane (Fig. 10).

Permutating pair-wise all station onset times tP

forms the jack-

knife ensemble of hyperbolae that will constrain the epicentre,

and will guide the reasonable depth assumption by graphically

minimizing the spread of curves. However, in the 3D solution

space (x,y,t0) for source epicentre and origin time, the two sta-

tion onset times tP

for any single hyperbola form an underdeter-

mined equation system; the derived solution gets scaled against

a third, undetermined parameter. In our case of a hyperbola for

source location, the free parameter is the source time t0 whichruns symmetrically to earlier origin at both outer legs (Fig. 11).

To get a mathematically exact solution for 3D parameter space,

one needs three constraints which translate to three station

onset times tP

in our case. Thus the related, three hyperbolae

will always match in one triple point regardless of any spe-

cific parameter selections; its existence is nota quality meas-

Jackknifing explained

Jackknifing performs outlyer-resistant statistics to solve

over-determined equations, and helps to trace the influence

of single (erroneous) parameters to the joint solution. The

principle is explained here by the task of finding the linear

trend of seven observations. In the left figure, the red dashed

line is determined by standard LMS analysis; the result is

significantly offset by the single blue outlyer. Jackknifing

instead breaks the dimensionality of parameter space to theminimum requirements for the linear solution, i.e., from

seven to two points which define a single straight line. Per-

mutating the seven observations pairwise gives 21 solutions,

or lines. Six of them are affected by the one outlyer while 15

remain undisturbed. Graphically one recognizes the great

spread of all disturbed solutions, and may average the main

trend to the green dashed line in the left figure. Compared

to the red LMS solution, the green line obviously improves

the estimate of a correct linear trend.

Figure 9 Example of three tS-t

Pspheres which constrain the hypocentre solution.

For depth as an external parameter, the opaque-red depth plane intersects the

spheres as circles. Their projection to surface visualizes location quality; perfect

touch in one point demands altered depth estimate. The deviation from per-

fect spheres comes by refracted paths in a layered-Earth model.

Figure 10 The location constraint of onset times from two distinct stations is

a hyperboloid with rotational symmetry around the stations connecting line.

Intersection with the depth plane forms a hyperbola as long as symmetry axis

and depth plane do not intersect. For subsurface stations, or inclined depth

planes, one may get (deformed) ellipses instead.


6/8




Latest with N= 8, the multiplicity of hyperbolae can not be

resolved in any area plot. One could escape to cell-hit counts,

like for tomographic resolution analysis, and Fig. 12 gives the

appropriate example. Alternatively, the hypolines (hyperbolae,

circles, beams) could be smeared as probability densities, and

added in fuzzy logic like manner (Liu and Saanford, 2001).

One could interpret the situation of 8+ stations as the break

even point of graphical jackknife location to the iterative Gei-

ger inversion, i.e. the residual analysis of travel time differences

for searching the best location solution in a least squares error

sense. But even then, jackknife analysis could still be beneficial

since the dense spread of triple point clouds is by far a more real-

istic measure of location accuracy than shaping the contour of

a 99% error ellipse.

At the other end of station numbers, already the records of

one SNS are sufficient to locate weak events in space and time.

For epicentres within some five times of SNS aperture, a reason-able depth estimate can be derived too since hyperbolae and cir-

cles shrink in opposite manner when the user-given depth plane

is altered up or down (Fig. 13).

Layer models and array processingOnce we abandon the simple half-space assumption, the semi-

spheres degrade to sliced spheres due to the effect of refraction

paths (Fig. 9), and hyperboloids must be constructed by inter-

section of these sliced spheres with respective discontinuities of

the first derivative (Fig. 11). Changing to subsurface stations

will introduce intersections of hyperboloids and the depth plane

that are not parallel to the hyperboloid rotation axis, and thusmay result in ellipses or even circles, instead of hyperbolae, as

2D intersection curves (Fig. 10). Our current limit for real-time

(1)

given that any time difference stays below the ratio station dis-

tance to velocity (else no hyperbola exists). Likewise, the upper

limit for the number of triple points Tis given by Eq. (2)

(2)

These formulae govern a strong increase of permutations, as

displayed in Table 2 for N= 3...12, where N= 4 describes

the situation of one SNS.

N 3 4 5 6 7 8 9 10 11 12

H 3 6 10 15 21 28 36 45 55 66

T 1 4 10 20 35 56 84 120 165 220

Table 2 Stations N, hyperbolae H, and triple points T by jackknife analysis.

Figure 11 Any hyperboloid may be constructed by intersection of two spheres

which grow due to the assumption of earlier origin time. When spheres

degrade due to refracted paths (see Fig. 9), the resulting hyperboloid gets

discontinuities in its first derivative (not shown here).

Figure 12 Event location by jackknifing for a local, six-station network. The

large number of hypolines already degrades the visibility of maps. Instead, cell-

hit counts can be colour-coded and get displayed in the inlet. The red circles

mark the automatedly determined maximum concentration of hypolines.

Figure 13 Estimation of optimum depth and half space vP by interactive slid-

ing of parameters. Variation of depth causes opposite changes of hyperbolae

and circles; circles finally vanish if wave fronts cant reach surface for the given

travel times. Variation of vP

will increase the spread of triple points if the opti-

mum solution is altered.


7/8




processing. The conducted explosions could not just be detected

but located in most of the cases. The determined magnitude

suggests a yield of 100 g explosive for the envisioned ML

2.0

threshold of OSI monitoring. Fig. 15 gives an example for the

first, 75 g explosion of Table 3. The event with ML

2.1 is at

the border of processing capabilities, phase onset times could

only be derived by array analysis. Fig. 16 relates the determined

magnitudes to the detonated yields where the extrapolation to

established magnitude-yield curves of large, nuclear explosions

(Khalturin et al., 1998) seems reasonable. The change in slope

may indicate the degraded fraction of radiated seismic energy

since large explosions will cause a high amount of evaporation.

The concept of nanoseismic monitoring has proven successful

in further field tests and training exercises of CTBTO, and will

be tested full scale with 30 SNS during the large, integrated field

modelling is the description of inclined layers that break hyper-

boloid symmetry, and demand 3D raytracing for determination

of the fastest travel path (Fig. 14).

The handling of array information in HypoLine stays

straight forward under all the above mentioned conditions.

Conforming to the idea of jack-knifing demands the pertur-

bation of array station phase readings into triplets which each

describe back-azimuth and slowness. For the tripartite SNS

with centre 3C site, one gets four beams per SNS. The spread

of beams is extremely sensitive to any variation in parameters,

and its minimum yields a reliable estimate for accurate phase

picking of related onsets even under poor SNR conditions. The

details for magnitude determination, specifically the extension

of the classical ML

scale to distances below 10 km, are described

in Wust-Bloch and Joswig (2006). The scheme for master event

correlation and relative hypocentre determination by HypoLine

is elaborated in Hge & Joswig (2008). A detailed tutorial aboutall aspects of nanoseismic monitoring can be found in the docu-

mentation part of /link2/.

DE04 results and conclusionsFor the event processing of DE04, Table 3 summarizes the detec-

tion sensitivity while Table 4 lists the results of one day SNS

3-component stationsYield [g] detection threshold [m]

75

150

200

400

475

600

340-500

340-500

1200

1900

2600

2600

Tripartite mini-arrays

Yield [g] location capability [m]

75

150

200

400

475

600

1500

1500

2500

2500+

2500+

2500+

Table 3 Sensitivity results of 3C single stations versus small arrays.

Turkish Hill 1.0 km Opposite Hill 1.5 km Hill near station E 2.5 km

OT

08:43:45

08:47:20

08:50:5508:54:40

08:58:25

09:02:05

09:35:35

09:39:20

09:43:30

09:47:10

09:51:30

09:55:50

Load[g]

75

150

200400

475

600

75

150

200

400

475

600

Ml

-2.1

-1.7

-1.4-1.3

-1.2

-1.2

-2.1

-1.8

-1.7

-1.4

-1.5

-1.3

comments

+ car

+ acoustic

+ acoustic

Ml

-

-2.0

-1.7-1.5

-1.4

-1.3

-2.4

-1.9

-1.7

-1.6

-1.6

-1.4

comments

detected

+ acoustic

Ml

-

-

-1.5-1.4

-1.4

-1.3

-

-1.9

-1.9

-1.5

-1.5

-1.4

comments

too weak

detected

detected

Results of Small Array Processing (SNS and HypoLine) of 7. Oct. 2004

Table 4 Single-event bulletin by small array processing.

Figure 14 Example of graphical location for subsurface stations and an inclined

layer model. The horizontal intersection curves are determined along an

adaptively inclined depth layer (red line); the vertical cut displays the strong

deformation of tS-t

Pcircles.


8/8




inclined layer models and subsurface stations, and provided

Figs. 9-11, 14.

ReferencesButler, R. [2003] The Hawaii-2 observatory: observation of nanoearth-

quakes, Seism. Res. Lett.74, 290-297.

Gutenberg, B. [1931] Microseisms in North America, Bull. Seism. Soc. Am.

21, 1-24.

Hge, M. and Joswig, M. [2008] Spatiotemporal characterisation of inter-

swarm period seismicity in the focal area of Nov Kostel (West Bohe-

mia/Vogtland) by a short-term microseismic study, Geophys. J. Int. (sub-

mitted).

Joswig, M. [1990] Pattern recognition for earthquake detection, Bull. Seism.

Soc. Am.80, 170-186.

Joswig, M. [1995] Automated classification of local earthquake data in the

BUG small array, Geophys. J. Int.120, 262-286.

Khalturin, V.I., Rautian, T.G. and Richards, P.G [1998] The seismic signalstrength of chemical explosions, Bull. Seism. Soc. Am.88, 1511-1524.

Kochnev, V.A., Goz, I.V., Polyakov, V.S., Murtayev, I.S., Savin, V.G., Zom-

mer, B.K. and Bryksin, I.V. [2007] Imaging hydraulic fracture zones from

surface passive microseismic data, First Break, 25(10), 77-80.

Labk, P., Joswig, M., Fojtkov, L., Dewez, P. and Guendel, F. [2005] Detec-

tion capability of 3-component seismic stations and tripartite mini-

arrays: CTBT monitoring of artificial nanoevents with M

Date post:	03-Apr-2018
Category:	Documents
Upload:	putih138242459
View:	214 times
Download:	0 times

Nanoseismic.Monitoring.fills.the.gap.pdf

Documents