Noise Analysis System Details
Data Preparation and Processing
The approach of this noise analysis method differs from many previous noise studies
in that we make no attempt to screen the continuous waveforms to eliminate body and
surface waves from earthquakes or transients and instrumental glitches such as data gaps,
clipping, spikes, mass recenters, or calibration pulses. These signals are included in our
processing because they are low-probability occurrences that do not contaminate high-
probability ambient seismic noise observed in the PDFs (see below for details). In fact,
transient signals often are useful for evaluating station performance. Also, eliminating
this event-triggering and removal stage has the benefit of significantly reducing the PSD
computation time by simplifying data pre-processing.
The algorithm used to develop the Albuquerque Seismological Laboratory (ASL) new
low noise model (NLNM) and new high noise model (NHNM) (Peterson, 1993; Bendat
and Piersol, 1971) is used to calculate PSDs for all stations in this study. The processing
steps are detailed below.
Record length. Let a finite length seismic time series, u(t), have N evenly sampled
points at an interval of
Δ
t. For our analysis, we parse continuous time series, for each
station component, into 1-hour (T
h
=3600s) finite-length time series segments,
overlapping by 50 pecent, distributed continuously in time. Overlapping time series
segments are used to reduce variance in the PSD estimate (Cooley and Tukey, 1965). For
this example, we assume that for the broadband seismic data, each 3600s times series
segment is sampled at 40 sample per second (sps), such that
Δ
t••••••s, for a total
N=144,000 data points.
Preprocessing. The PSD preprocessing of each 1-hour time segment consists of
several operations. First, to significantly improve the Fast Fourier Transform (FFT) speed
ratio, by reducing the number of operations, the number of samples in the time series, N,
is truncated to the next lowest power of two, 2
17
, leaving N=131,072, thereby reducing the
series length such that T
h
=3276.8s. Second, in order to further reduce the variance of the
final PSD estimates, each roughly 1-hour time series record is divided into 13 segments,
overlapping by 75 percent, where the length of each new time series segment is now,
T
r
=T
h
/4=819.2s with N=32,768=2
15
. The sample size N is chosen based on the longest
period of interest, T
l
, (lowest frequency, fl). In general, the record length, T
r
=N
Δ
t, is
chosen such that it is 10 times the longest resolvable period, T
l
. Given this,T
l
,=1/f
l
=
T
r
/10=90s. The shortest period, Ts, (highest frequency, fh) is equivalent to the Nyquist
folding frequency, fc=1/2•t=20Hz, and is given by Ts = 1/fc
≤
1/fh
≤
0.05s.
Third, in order to minimize long-period contamination, the data are transformed to a
zero mean value, and any long period linear trend is removed by the average slope
method. If u
n
are the data values in the time series u(t) of length T
r
and N samples, the
data mean is given by:
(1)
mean
u
=
1
N
n
u
n=1
∑
Long period trend, T
lp
, is defined as any frequency component whose period is longer
than the record length, T
r
, and is defined as:
7