1: %\documentstyle[epsfig]{aipproc2}
2: \documentstyle[psfig]{aipproc2}
3: \textheight 22.6cm
4: \textwidth 16.8cm
5:
6: %
7: \def\be{\begin{equation}}
8: \def\ee{\end{equation}}
9: \def\bea{\begin{eqnarray}}
10: \def\eea{\end{eqnarray}}
11:
12: \begin{document}
13: \title{Removal of interference from external coherent signals}
14:
15: \author{Alicia M. Sintes and Bernard F. Schutz}
16: \address{Max-Planck-Institut f\"ur Gravitationsphysik
17: (Albert-Einstein-Institut),
18: Schlaatzweg 1, D-14473 Potsdam, Germany}
19:
20: %National Center for Research\thanks{The National
21: %Center for Research is sponsored by the National
22: %Science Foundation.}, Boulder, Colorado 80307\\
23: %$^{\dagger}$National Standards Institute, Boulder, Colorado 11543}
24: %\lefthead{LEFT head}
25: %\rig436thead{RIGHT head}
26:
27: \maketitle
28:
29:
30: \begin{abstract}
31: We present a technique that we call coherent line removal,
32: for removing external coherent
33: interference from gravitational wave interferometer data.
34: We illustrate the usefulness of this technique applying it to the
35: the data produced by the Glasgow laser interferometer in 1996 and
36: removing all those lines corresponding to the electricity supply frequency
37: and its harmonics. We also find that this method seems to reduce the level
38: of non-Gaussian noise present in the interferometer and therefore, it
39: can raise the sensitivity and duty cycle of the detectors.
40:
41: \end{abstract}
42:
43: \section*{Introduction}
44:
45: In the measured noise spectrum of the different gravitational wave
46: interferometer prototypes \cite{Abra,ga,J}, one observes peaks due
47: external interference,
48: where the amplitudes are not stochastic in contrast to the stochastic
49: noise. The most numerous are powerline frequency harmonics.
50: In this paper we review how to remove these very effectively
51: using a technique we call coherent line removal ({\sc clr}) \cite{AS1,AS2}.
52:
53: {\sc clr} is an algorithm able to remove interference present in the
54: data while preserving the stochastic detector noise. {\sc clr}
55: works when the interference is present in many harmonics, as long as
56: they remain
57: coherent with one another. Unlike other existing methods for removing
58: single interference lines \cite{Th,p3}, {\sc clr} can remove the external
59: interference without removing any \lq single line' signal buried by the
60: harmonics. The algorithm works even when the interference frequency changes.
61: {\sc clr} can be used to remove all harmonics of periodic or
62: broad-band signals (e.g., those which change frequency in time), even when
63: there is no external reference source. {\sc clr} requires little or
64: no a priori knowledge of the signals we want to remove. This is a
65: characteristic that distinguishes it from other methods
66: such as adaptive noise
67: cancelling \cite{anc}.
68: It is \lq safe' to apply this technique to gravitational wave data
69: because we expect that coherent gravitational wave signals will appear
70: with at most the fundamental and one harmonic \cite{Sc2}.
71: Lines with multiple harmonics must be of terrestrial origin.
72:
73: In this paper, we illustrate the usefulness of this new technique by applying
74: it to the data produced by the Glasgow laser interferometer in March 1996
75: and removing all those lines corresponding to the electricity supply frequency
76: and its harmonics. As a result the interference is attenuated or eliminated
77: by cancellation in the time domain and the power spectrum appears
78: completely clean allowing the detection of signals that were
79: buried in the
80: interference. Therefore, this new method appears to be good news as far
81: as searching for continuous waves (as those ones produced by pulsars
82: \cite{Sc2,T})
83: is concerned.
84: The removal improves the data in the time-domain as well.
85: Strong interference produces a significant non-Gaussian component to the noise.
86: Removing it therefore improves the sensitivity of the detector to short
87: bursts of gravitational waves \cite{n}.
88:
89:
90: %\begin{figure}[t!] % fig 1
91: %\centerline{\epsfig{file=got.ps}}
92: %%\centerline{\epsfig{file=got.ps,height=3.5in,width=3.5in}}
93: %\caption{Figure captions are automatically centered and can be
94: %more than one line.}
95: %\vspace*{10pt}
96: %\label{fig1}
97: %\end{figure}
98: %
99: %\begin{figure}
100: %\centerline{\epsfysize=3.0in \epsfbox{figu1.ps} }
101: %\caption{Comparison of the structure of the lines at 250 Hz and at 350 Hz
102: %of the power
103: %spectrum of the Glasgow data.
104: %The broad shape is due to the wandering of the incoming electricity
105: %frequency.}
106: %\label{phidot}
107: %\end{figure}
108:
109:
110: \section*{Coherent Line Removal}
111:
112:
113: In this section, we summarize the principle of {\sc clr}.
114: For further details we refer the reader to \cite{AS2}.
115:
116: We assume that the interference has the form
117: \be
118: y(t)=\sum_n a_n m(t)^n + \left( a_n m(t)^n\right)^* \ ,
119: \label{e3}
120: \ee
121: where $a_n$ are
122: complex amplitudes and $m(t)$ is a nearly monochromatic function
123: near
124: a frequency $f_0$.
125: The idea is to
126: use the information in the different harmonics of the interference
127: to construct a function $M(t)$ that is as close a replica as possible of
128: $m(t)$ and then construct a function close to $y(t)$ which
129: is subtracted from the output of the system cancelling the interference.
130: The key is that real gravitational wave signals will not be present
131: with multiple harmonics and that $M(t)$ is constructed from many
132: frequency bands with independent noise. Hence, {\sc clr} will little
133: affect the statistics of the noise in any one band and
134: any gravitational wave signal masked by the interference can
135: be recovered without any disturbance.
136:
137: We assume that the data produced by the system
138: is just the sum of the interference plus noise
139: \be
140: x(t)=y(t)+n(t) \ ,
141: \label{e6}
142: \ee
143: where $y(t)$ is given by Eq.~(\ref{e3}) and the noise $n(t)$ in the
144: detector is a zero-mean stationary
145: stochastic process.
146: The procedure consists in defining a set of functions $\tilde z_k(\nu)$
147: in the frequency domain as
148: \be
149: \tilde z_k(\nu)\equiv \left\{
150: \begin{array}{cc}
151: \tilde x(\nu) & \nu_{ik}<\nu <\nu_{fk}\\
152: 0 & \mbox{elsewhere}\ ,
153: \end{array}
154: \right.
155: \label{e8}
156: \ee
157: where $(\nu_{ik}, \nu_{fk})$ correspond to the upper and lower frequency
158: limits of the harmonics of the interference
159: and $k$ denotes the harmonic considered.
160: These functions are equivalent to
161: \be
162: \tilde z_k(\nu)= a_k \widetilde{m^k}(\nu) +\tilde n_k(\nu) \ ,
163: \ee
164: where $ \tilde n_k(\nu)$ is the noise in the frequency band of the
165: harmonic considered. Their inverse Fourier transforms yield
166: \be
167: z_k(t)=a_k m(t)^k +n_k(t) \ .
168: \ee
169: Since $m(t)$ is supposed to be a narrow-band function near a frequency $f_0$,
170: each $z_k(t)$ is a narrow-band function near $kf_0$.
171: Then, we define
172: \be
173: B_k(t)\equiv \left[ z_k(t)\right]^{1/k}\ ,\label{e10a}
174: \ee
175: that can be rewritten as
176: \be
177: B_k(t)= (a_k)^{1/k}m(t) \beta_k(t) \ , \qquad
178: \beta_k(t)=\left[ 1+ {n_k(t) \over a_k m(t)^k}\right]^{1/k} \ .
179: \label{e10}
180: \ee
181: All these functions, $\{B_k(t)\}$, are almost monochromatic around the
182: fundamental frequency, $f_0$, but they differ basically by a certain
183: complex amplitude. These factors, $\Gamma_k$, can easily be calculated,
184: and we can construct a set of functions $\{b_k(t)\}$
185: \be
186: b_k(t)=\Gamma_k B_k(t)\ ,
187: \ee
188: such that, they all have the same mean value. Then, $M(t)$ can be
189: constructed as a function of all $\{b_k(t)\}$
190: in such a way
191: that it has the same mean and minimum variance.
192: If
193: $M(t)$ is linear with $\{b_k(t)\}$, the statistically the best is
194: \be
195: M(t)=\left(\sum_k {b_k(t) \over {\rm Var}[\beta_k(t)]} \right) {\Big
196: { /}}
197: \left( \sum_k {1 \over {\rm Var}[\beta_k(t)]}\right) \ ,
198: \ee
199: where
200: \be
201: {\rm Var}[\beta_k(t)]= {\langle n_k(t) n_k(t)^*\rangle\over k^2
202: \vert a_k m(t)^k\vert^2}+ \mbox{corrections} \ .
203: \ee
204: In practice,
205: we approximate
206: \be
207: \vert a_k m(t)^k\vert^2 \approx \vert z_k(t)\vert^2 \ ,
208: \ee
209: and we assume stationary noise. Therefore,
210: \be
211: \langle n_k(t) n_k(t)^*\rangle= \int_{\nu_{ik}}^{\nu_{fk}} S(\nu) d\nu \ ,
212: \ee
213: where $S(\nu)$ is the power spectral density of the noise.
214:
215: Finally, it only remains to determine the amplitude of the different
216: harmonics, which can be obtained applying a least square method.
217:
218: \section*{Removal of 50 Hz harmonics}
219: In this section, we present experimental results that demonstrate the
220: performance of the {\sc clr} algorithm and show its potential value.
221: We apply this method to the data produced by the Glasgow laser
222: interferometer in March 1996 and the electrical interference is
223: successfully removed.
224:
225: In the study of the Glasgow data, we observe in the power spectrum many
226: lines. Some of them are due to thermal noise (which we
227: will not consider here) and many others
228: at multiples of 50 Hz due external interference, where the amplitudes
229: are not stochastic.
230: In long-term Fourier transforms, the lines at multiples
231: of 50 Hz are broad, and the structure of different lines is similar
232: apart from an overall scaling proportional to the frequency. In smaller
233: length Fourier transforms, the lines are narrow, with central frequencies
234: that change with time, again in proportion to one another. It thus appears
235: that all these lines are harmonics of a single source (e.g., the
236: electricity supply) and that their broad shape is due to the wandering of
237: the incoming electricity frequency.
238:
239: %
240: \begin{figure}[b!]
241: \centerline{\vbox{
242: \psfig{figure=f1.ps,height=12.5cm,width=12.5cm}
243: }}
244: \vspace*{10pt}
245: \caption[]{Decimal logarithm of the
246: periodogram of $2^{19}$ points (approximately 2 minutes) of the Glasgow data.
247: (a) One of the harmonics near 754 Hz.
248: (b) The same data after the removal of the interference as described
249: in the text.
250: (c) The same experimental data
251: with an artificial signal added at 452 Hz.
252: (d) The data in (c) after the removal of the interference, revealing that
253: the signal remains detectable. Its amplitude is hardly changed by removing
254: the interference.
255: }
256: \vspace*{10pt}
257: \end{figure}
258: %
259:
260: In the Glasgow data, those lines at 1 kHz have a width of 5 Hz. Therefore,
261: we can ignore these sections of the power spectrum or we can try to
262: remove this interference in order to be able to detect gravitational
263: waves signals masked by them.
264:
265:
266: In order to remove the electrical interference, we separate the data
267: into groups of $2^{19}$ points (approximately two minutes) and, for each of
268: them, the coherent line removal algorithm is applied.
269: A detailed description can be found in \cite{AS2}.
270:
271:
272: In Fig. 1, we show the performance of {\sc clr} on two minutes of data.
273: We can see how {\sc clr} leaves the spectrum completely clean of the
274: electrical interference and keeps the intrinsic detector noise.
275: {\sc clr} is also applied to the true experimental data
276: with an external simulated signal at 452 Hz, that is initially hidden
277: due to its weakness and we succeed in removing the electrical
278: interference while keeping the signal present in the data, obtaining
279: a clear outstanding peak over the noise level.
280:
281: %
282: \begin{figure}[h!]
283: \centerline{\vbox{
284: %\psfig{figure=f2.ps,height=15.5cm,width=10.0cm}
285: \psfig{figure=f2.ps,width=10.0cm}
286: }}
287: \vspace*{10pt}
288: \caption[]{ Comparison of a zoom of the spectrogram.
289: The dark areas correspond to the periods in which the detector is
290: out of lock. (a) is obtained from the Glasgow data. We can observe
291: the wandering of the incoming electrical
292: signal. The other two remaining lines
293: at constant frequency correspond to violin modes. (b) The same spectrogram
294: as in (a) after applying coherent line removal, showing how the
295: electrical interference is
296: completely removed.
297: }
298: \vspace*{10pt}
299: \end{figure}
300: %
301: In Fig. 2, we compare a zoom of the spectrogram for the frequency
302: range between 740 and 760 Hz. There we can see the performance of the
303: algorithm on the whole data stream. We show
304: how a line due to an harmonic of the electrical
305: interference in the initial data is removed.
306:
307:
308: We are interested in studying possible side effects of the line removal
309: on the statistics of the noise in the time domain. We observe that
310: the mean value is hardly changed. By contrast, a big difference is obtained
311: for the standard deviation. For the Glasgow data, its value is around
312: 1.50 Volts. After the line removal, the standard deviation is reduced,
313: obtaining a value around 1.05 Volts. This indicates that a huge amount
314: of power has been removed.
315:
316: Further analysis reveals that values of
317: skewness and kurtosis are getting closer to zero after the line removal.
318: Values of skewness and kurtosis
319: near zero suggest a Gaussian nature. Therefore, we are
320: interested in studying the possible reduction of the level of non-Gaussian
321: noise. To this end, we take a piece of data and we study
322: their histogram, calculating the number of events that lie between
323: different equal intervals. If we plot the logarithm of the
324: number of events versus $(x-\mu)^2$, where $x$ is the central position of
325: the interval and $\mu$ is the mean, in case of a Gaussian distribution,
326: all points should fit on a straight line of slope $-1/2\sigma^2$, where
327: $\sigma$ is the standard deviation. We observe that this is not the
328: case (see figure 3). Although, both distributions seem to have
329: a linear regime, they present a break and then a very heavy tail.
330: The two distributions are very different. This is mainly due to the
331: change of the standard deviation.
332:
333:
334: \begin{figure}[h!]
335: \centerline{\vbox{
336: \psfig{figure=f3.ps,width=10cm}
337: }}
338: \vspace*{10pt}
339: \caption[]{ Comparison of the logarithm plot of the histogram
340: for $3.2\times 2^{19}$ points
341: as a function of
342: $(x-\mu)^2$. The circles correspond to the Glasgow data
343: and the stars to the same data after removing the electrical interference.
344: The Glasgow data is
345: characterized by $\mu=-0.0182$ Volts and $\sigma = 1.5151$ Volts.
346: After the line removal, we obtain the
347: values of $\mu=-0.0182$ Volts and $\sigma = 1.0449$ Volts.
348: In the right-hand corner, there is zoom of the original figure, but
349: rescaled so that
350: the abscissa corresponds to $(x-\mu)^2/2\sigma^2$.
351: If the data resembles a Gaussian distribution,
352: we will expect a single straight line of slope -1. This is not the
353: case for the Glasgow data, but it seems to be satisfied for the
354: clean data up to $4\sigma$. The large number of points in the
355: highest bin of the Glasgow data is an effect of saturating the ADC. These
356: points are spread to higher and lower voltages by line removal.
357: }
358: \vspace*{10pt}
359: \end{figure}
360: %
361:
362:
363: We can zoom the \lq linear' regime and change the scale in
364: the abscissa to $(x-\mu)^2/(2\sigma^2)$. Then, any Gaussian
365: distribution should fit into a straight line of slope -1.
366: We observe that after removing the interference, it follows a
367: Gaussian distribution quite well up to $4\sigma$.
368: The original Glasgow data does not fit a straight line anywhere.
369:
370:
371: In order study the Gaussian character, we have also applied
372: two statistical tests to the
373: data: the chi-square test that measures the discrepancies
374: between binned distributions, and the one-dimensional Kolmogorov-Smirnov
375: test that measures the differences between cumulative
376: distributions of a continuous data.
377:
378: We computed the significance probability for every $2^{12}$ points
379: of the data
380: using both tests and we checked whether the distribution are Gaussian or not.
381: The two tests are not equivalent but in any case, the values of the
382: significance probability would be close to unity for
383: distributions resembling a Gaussian distribution.
384: In both tests, the significance probability increased
385: after removing the electrical interference, showing that this
386: procedure suppresses some non-Gaussian noise, although, generally
387: speaking, the distribution was still
388: non-Gaussian in character, presumably because of the heavy
389: tails, which are not affected by line removal. See \cite{AS2} for details.
390:
391:
392: \section*{Acknowledgments}
393: We would like to thank C. Cutler, A. Kr\'olak, M.A. Papa and
394: A. Vecchio for helpful discussions,
395: and J. Hough and the gravitational waves group at Glasgow University
396: for providing their gravitational wave interferometer data for analysis.
397: This work was partially supported by the European Union,
398: TMR Contract
399: No. ERBFMBICT972771.
400:
401:
402: \begin{references}
403:
404: \bibitem{Abra} Abramovici A., Althouse W., Camp J., Durance D.,
405: Giaime J.A., Gillespie A., Kawamura S., Kuhnert A.,
406: Lyons T., Raab F.J., Savage Jr. R.L., Shoemaker D., Sievers L.,
407: Spero R., Vogt R., Weiss R., Whitcomb S., Zucker M.,
408: {\it Physics Letters A } {\bf 218}, 157 (1996).
409:
410: \bibitem{ga} Maischberger K., Ruediger A., Schilling R., Schnupp L.,
411: Winkler W., Leuchs G., \lq\lq Status of the Garching 30 meter
412: prototype for a large gravitational wave detector", eds.
413: Michelson P.F., En-Ke Hu \& Pizzella G., in {\it Experimental
414: Gravitational Physics}. World Scientific, Singapore, 1991, pp. 316-21.
415:
416: \bibitem{J} Jones G.S., 1996, {\em Fourier analysis of the data produced
417: by the Glasgow laser interferometer in March 1996}, internal report,
418: Cardiff University, Dept. of Physics and Astronomy.
419:
420:
421: \bibitem{AS1} Sintes A.M., Schutz B.F., \lq\lq Removal of
422: interference from gravitational wave spectrum",
423: in {\it proceedings of the 2nd workshop on Gravitational Wave Data Analysis}.
424: Orsay, France, 1998.
425:
426: \bibitem{AS2} Sintes A.M., Schutz B.F.,
427: \lq\lq Coherent Line Removal: A new technique to remove interference from
428: the gravitational wave spectrum", 1998, to be published.
429:
430:
431: \bibitem{Th} Thomson D.J., 1982, \lq\lq Spectrum Estimation and
432: Harmonic Analysis", in {\it Proceedings of the IEEE}, {\bf 70}, 1055-96 (1982).
433:
434: \bibitem{p3} Percival D.B., Walden A.T., {\it Spectral
435: analysis for physical applications}, first edition, Cambridge
436: University Press, 1993.
437:
438: \bibitem{anc} Widrow B., Glover J.R., McCool J.M., Kaunitz J.,
439: Williamns C.S., Hearn R.H., Zeidler J.R., Dong E., Goodlin R.C.,
440: \lq\lq Adaptive Noise Cancelling: Principles and Applications", in
441: {\it Proceedings of the IEEE}, {\bf 63}, 1692-1716 (1975).
442:
443:
444: \bibitem{Sc2} Schutz B.F., \lq\lq Detection of Gravitational Waves",
445: in
446: {\em Relativistic Gravitation and Gravitational Radiation},
447: Marck J.A., Lasota J.P., eds., Cambridge University Press, 1997.
448:
449:
450: \bibitem{T} Thorne K.S., eds. Kolb E.W. \& Peccei R., in
451: {\it Proceedings of Snowmass 1994 Summer Study on Particle and Nuclear
452: Astrophysics and Cosmology}. World Scientific, Singapore, 1995, p. 398.
453:
454:
455: \bibitem{n} Nicholson D., Dickson C.A., Watkins W.J.,
456: Schutz B.F., Shuttleworth J., Jones G.S.,
457: Robertson D.I., Mackenzie N.L., Strain K.A., Meers B.J., Newton G.P.,
458: Ward H., Cantley C.A., Robertson N.A., Hough J.,
459: Danzmann K., Niebauer T.M., R\"udiger A., Schilling R., Schnupp L.,
460: Winkler W.,
461: {\it Physics Letters A} {\bf 218}, 175-180 (1996).
462:
463:
464: \end{references}
465:
466: \end{document}
467:
468: