Académique Documents
Professionnel Documents
Culture Documents
A laser radar whose resolution is greater than 1 pum reported. We present the radar results when they is are used for such purposes as determining the size of a void inside a silicon wafer, profiling a cross-sectional pattern of an optical fiber, studying the birefringence of a lithium-niobate crystal, or finding a fault in an optical guide in an optical integrated-circuit wafer. Neural-network theory was used in processing the radar signal. Radar processing based on neural-network theory gave significantly superior resolution compared with Fourier-transform-based processing.
Introduction
The resolution of the step-frequency radar increases with the range of frequency steps. Laser diodes can shift their oscillating frequency by tens of terahertz. The obtainable resolution of a radar with such a wide range of frequency steps is of the order of micrometers. This kind of high-resolution radar opens up new applications in measurement in the field of optics. An immediate application of such an imaging radar would be the prescreening of either silicon or GaAs wafers to locate possible internal flaws. It could also be implemented in a device for profiling the cross section of an optical fiber while the fiber is being drawn without interruption of production. It would be possible to use the radar for locating faults even in optical integrated circuits (IC's). One of the most popular optical-fiber fault locators to date is the optical time-domain reflectometer.1 Extending the resolution of this method to an order of micrometers would be difficult. Obtaining a -pm resolution requires femtosecond pulses. Some of the delicate optical IC's might not withstand the peak intensity of such pulses. Two types of optical frequency-domain reflectometer2 - 5 have been proposed; one modulates the frequency of the amplitude modulation, and the other
modulates the carrier frequency of the laser light. The former type does not necessitate coherent laser light, but the obtainable resolution is limited to an order of meters. While a resolution of a meter may be sufficient for an optical-fiber fault locator, it is not good enough for an optical IC fault locator. The latter type of optical frequency-domain reflectometer can achieve the micrometer resolution needed for an IC fault locator; however, source coherency requirements are more demanding. Another interesting approach is the optical coher6 ence-domain reflectometer, ' 7 which makes use of the fact that, if an incoherent light source is used for a Mach-Zehnder interferometer, an interference pattern exists only at the location where the path from the source is exactly identical to that from the reference arm. The optical coherence-domain reflectometer, however, requires a moving stage and a source with a higher degree of incoherency. The reported resolutions are of the order of 10 m.
Principles of Operation
K. lizuka is now with the Ontario Laser and Lightwave Research Center, Department of Electrical Engineering, University of Toronto, Toronto, Canada M5S 1A4. S. Fujii now is with the Communications Research Laboratory, Okinawa Radio Observatory, Nakagusuku-son, Nakagami-gun, Okinawa 901-24, Japan.
Received 25 November 1992.
The neural-network laser radar uses a cw. The frequency of the transmitted wave is changed stepwise, and at each step of frequency the phase and amplitude of the wave scattered from the target are measured. One extracts the distance information by analyzing the changes in phase and amplitude of the received wave as the frequency of the transmitted wave changes. Figure 1 shows the layout of the step-frequency radar. The carrier frequency of the laser beam is stepped. The operating frequency f of the nth frequency is
0003-6935/94/132492-10$06.00/0.
0 1994 Optical Society of America.
fn = o +nAf.
(1)
2492
Xo
Xk
@
XN-1
. . . .
f
time
--
:t=
where n = 0, 1, 2, . . , N - 1,fo is the frequency o f the initial step, and Af is the frequency step width. The amplitude of the transmitting light is E. The received signal is the sum of the signals scattered from all the scattering centers. Let the scattering centers be located at
Xk = x 0 + kAx
form
N-i
exp(j2nk N)(7)
(2)
k.
Sk
is the desired
The
Two methods of obtaining Sk were considered: (a) fast-Fourier-transform (FFT) signal processing and (b) neural-network signal processing.
Fast-Fourier-Transform Signal Processing
Hn = N- 1E
k=O k-s
fo + nAf
kAx) .
V
(3)
Since Eq. (7) is of the form of an inverse discrete Fourier transform, one immediate method of obtaining Sk from En is the execution of the discrete Fourier transform of En. The discrete Fourier transform provides values of
SOW SD 2) .. * Ski .. *
(4)
SN-1-
kAx j4Tr -f Hn exp( j4m vx0) = NEi exp orE Jr -yv k=O ,k x exp j2,r l V
N-1
(5)
The left-hand side of Eq. (5) has only*n as a parameter, whereas the sum of the first factor of the right-hand side has only k as a pa rameter. The second factor in Eq. (5) has the form of the weighting function of the discrete Fourier transf orm. Let us put
Nonzero values of Sk indicate the existence of scatterers at Xk = x0 + kAx. One of the attractions of FFT processing is that FFT algorithms are readily available, which simplifies the implementation. Another interesting feature of FFT processing is its zooming capability, such as the zoom function of a camera. The value of x0 in Eq. (2) can be arbitrarily set to a desired value, and a higher resolution can be maintained in the x > x0 region. This zoom capability is desirable in applications such as the optical IC chip fault locator, where the length of the input guide of the IC chip is of the order of millimeters, but one desires the resolution inside the IC to be of the order of micrometers.
Neural-Network Signal Processing
En= Hnexp(-j4r
x0),
kAx
Sk = kE exp j4'Tr - fo .
(6)
The neural network similar to the Hopfield type was used to find Sk from E,,. Such a network is shown in Fig. 2. It resembles an array of identical amplifiers. Two special features of the array are the nonlinearity of the input-output function of the amplifiers and the intensive network of feedback. The input-output
2493
with where U,and Viare the input and the output voltages, respectively, at the ith amplifier. Ii is the external input current from the ith terminal, and g(Uj) is the input-output function of the amplifier and monotonically increases with Uj. The negative sign was placed in front of the summation in Eq. (8) to indicate the negative feedback explicitly. In the special case in which the input to only one port is distinctively predominant and the coupling constants are all positive real numbers, the status of
*=
C
V0
V,
V2
$9
C
SI
Xl
XOt
1 N-1 N-1
L=-2 a
function of an amplifier is nonlinear, but the output always monotonically increases with the input signal level. The output from each amplifier is fed back to the inputs of all other amplifiers including to itself through couplers with coupling constant Tip The output from the array is determined by the coupling constants. The Ti. consists of N x N constants, which are in fact the same number as the weighting functions of the Fourier transform given by Eq. (7). The weighting functions of the Fourier transform are made so that twice the integral transform brings back the original function (except for the negative sign in the variable). At issue here is whether this special feature of the FFT is essential when it is used for radar signal processing. On the other hand the values of Tij are calculated from a given geometry of the radar so that the difference between the measured and theoretically calculated values is minimized.
XTumi - IXii
+R I
fN-1 Vi
(V)dV
(9)
where g-i(V) is the value of U when expressed in terms of V. Note here that energy is a hypothetical quantity and we are not referring to energy measured in joules. If an expression that represents the error between the experimentally measured values and the theoretically calculated values is reformulated to be of a form that is similar to the Lyapnov energy function, the status of the outputs converges to the values of the minimum error. This characteristic of the Hopfield neural network is utilized to find the outputs that optimumly fit with the theoretical results. Thus below we find first the error function and then reformulate it in the form of the Lyapnov energy function to find the values of Tij for our present case. Once the values of Tij are found, the numerical solution of the differential equation, Eq. (8), which represents the function of the circuit in Fig. 2, gives the converging values of Vithat minimize the error. Short-hand notation is used, namely,
en = exp j2,r
Ii - C dt dt - R -TijVj=, ~j=0
2494 APPLIED OPTICS / Vol. 33, No. 13 / 1 May 1994
dU
U,
N-1
(8)
n( I
N} ,
(10)
where the summation sign has been removed with the understanding that the appearance of the same subscripts twice means the summation with respect to that subscript from 0 to N - 1. Equation (7), which is the measured quantities at frequencies fo, fi, f2,... ) fn*... * fN-l are now rewritten by use of Eq.
(10) as
eOk
elk
e2k
ek =
enk
(15)
e(N- l)k
E= enkSk,
Equation (14) is recast into the form of Eq. (9). The Euclidean norm is the sum of the squares of all matrix elements: 1 N-1 QN-1 E n=O m=O
EN-, = e(N-l)kSk.
(11)
N-1
-
I
i=O
S,12ei'ei.
(16)
(12)
The derivation of Eq. (16) from Eq. (14) is shown in Appendix A by an example with N = 2. Before we compare Eqs. (9) and (16), the third term of Eq. (9) needs special attention. A simpler input-output function g(U) of a neuron is desirable from the viewpoint of computation. We found that the function of such a simple shape as shown in Fig. 3 serves our purpose. It is expressed as
AU U> 0
g(U)= 0
U< 0'
(17)
elkSk
where A is the gain factor of the amplifiers. With Eq. (17) the third term of Eq. (9) becomes (V)dV R izoJog 2 RAIX j 0 0 0
R N1 Vi
1 1
N-1
P=EEt=
enkSk
(13)
L=2 I
1 N-1
N-1
i=O j=O
(.~~ + Tij +
8zJ
We now show that the Euclidean norm Q of the difference between the covariance matrix obtained experimentally and theoretically can be reformulated to the Lyapnov energy function so that the Hopfieldtype neural network can be used for optimization.9
Q is expressed as
NV 2.
(18)
E-O ivi.
(19)
QIXF
1
-
N-i
2 (14)
0 Input U
klekek t k=O I7
where ek is a vector defined by the kth target, 11-11 denotes the Euclidean norm, and the bar over P indicates the measured values. The elements of the vector are phases at different frequencies for the kth
2495
AR
If the values of the parameters given in Eqs. (20) are used in the neural network shown in Fig. 2, the output values are such that the Euclidean norm of the error function is minimized. In conclusion the purpose of the neural-network processing procedure is (1) to calculate Tij from Eqs. (10) and (20) and to calculate Ii from the measured P, (2) either to implement the circuit shown in Fig. 2 with the values of the components calculated above or to rely on a computer to solve the differential equation, Eq. (8), with the parameters obtained above. In this paper we took the latter approach, and we solved the differential equation numerically by the Runge-Kutta method. 10 We should emphasize that once the coupling constants Ti3 are calculated, they do not have to be changed according to the type of target; only the values of Ii have to be calculated from the covariance matrix whose elements are obtained from the measured Envalues.
Layout of the Experimental Apparatus
as a source. The wavelength could be varied from 1.5 to 1.6 pm, which corresponds to 12.5 THz of frequency shifting. By means of an acousto-optic modulator (AOM)the laser beam is split into a probe (object) beam, shown as a solid line, and a reference beam, shown as a dashed line. The AOM not only splits the beam into two but also shifts the carrier frequency of the reference beam by the driver frequency of 40 MHz from that of the probe beam. The probe beam is focused into the target by way of a nonpolarizing beam splitter. The return beam scattered from the scatterers in the target takes the same route and is deflected by the nonpolarizing beam splitter into the photodiode mixer. The reference beam whose carrier frequency is shifted from that of the probe beam by 40 MHz is guided by two mirrors into the photodiode mixer mentioned above. The mixer diode puts out a 40MHz cw signal whose amplitude is proportional to the reflection from the target and whose phase is the same as that of the probe beam. The advantage of such a detection scheme is that no matter how much frequency of the light source is shifted, all one needs is a 40-MHz amplifier. The amplitude and phase information at each step of frequency are fed into the processor. A He-Ne visible-light laser beam is also put into the system for easier alignment of the object beam with the target. The target is mounted on an electronic precision translator stage for easy alignment. Two visible-range stereoscopic microscopes and an IR camera were found useful for coupling the object beam into the optical guide on the optical IC. A microscope objective focused the probe beam to a
Wave meter
AOA
Xk
Laser
7
BS
-
-A
fn
2496
DiscreteFourier Transtorm
Neural network processing
0.8
05 4
uLr
7
0.0
(a)
Probe beam
(b)
HFig. 5.
321
target.
specific region of interest. A large numericalaperture microscope lens, however, distorts the relative intensities of the scattering along the distance. The intensity of the scattered wave at the distance that the probe beam is focused is highlighted. At each step of the frequency the wavelength of the probe beam was measured by a seven-digit waveme-
The first target was a quartz glass plate with a 27-pm-thick air void. We made the void by inserting 27-p1m-thickspacers between two glass plates. Figure 5 shows the display of the radar output. The vertical axis is the scattering intensity and the horizontal axis, the distance. The range of frequency (N - 1)Afwas 12.5 THz, and the number of steps was N = 512. The distance that appears in the display depends on the wavelength in the medium. The index of refraction of each layer is necessary to find the exact thickness of the layers. The locations of the four peaks clearly correspond to the interfaces between air and glass. We show the display obtained by FFT processing with a Blackman Harris window by the solid curve and that obtained by the neural network by the dotted curve. A significant reduction of the sidelobes is observed in the result processed by the neural network. Figure 6 demonstrates the power of neural-network processing over the FFT processing. Using the same quartz glass target, we repeated the measurements with a reduced frequency range (N - )Af = 4.23 THz and compared the measured result with (N - 1)Af= 12.5 THz. The top two graphs in Fig. 6 are the results processed by the FFT with these two frequency ranges. The resolution of the discrete-Fouriertransform processing is given from Eq. (4) by
V
Ax= 2fN'
(21)
ter.
(N-1) Af = 12.5THz
FFT
a
a)
0.6
U) 0.2)-
-20
ono
0.8 i
1.0 _
.L
Neural Network
C ) 0.6
O0
.C= 0.4 0
Z cn 0.2k
0.00
-
1~~I,
-200 0.0 200
Distance(arbitrary
Fig. 6. Comparison of the radar displays with FFT processing and with neural-network processing.
2497-
longer resolve the center peaks, and the total number of peaks is reduced from four to three. The two lower curves show the results when neuralnetwork processing was applied to these two frequency ranges. The lower right curve was made with (N - 1)Af= 4.23 THz. The center peak, which could not be resolved as two peaks with FFT processing, can now be resolved as two peaks with neuralnetwork processing. The widths of the peaks of the lower right curves are sharpened further when the frequency range is raised to 12.5 THz, as shown in the figure at the bottom left. To quantify the improvement in the resolution, we compare the widths of the peaks at the distance of -200 units shown on the graph compared for FFT processing and neuralnetwork processing. The ratio of the widths is 6, which means that neural-network processing yields a factor-of-6 improvement over FFT processing, and the resolution of the radar with neural-network processing in this medium is 1 jim. Next Fig. 7 shows the result when an optical fiber is used as a target. The fiber is an oversized 250-jimdiameter, step-index, multimode fiber. The indices of refraction of the core and cladding layers are 1.59 and 1.52, respectively. The probe beam was incident normal to the wall of the cladding layer. A possible application of such a radar is for profiling the cross section of an optical fiber while the optical fiber is being drawn from the nozzle. The production is not interrupted for the sake of measurement. Figure 8 shows the results when the radar was used
1.0
| |
L I--
.- 0.5
C0
,.
0.0
0.5
0.0
I I i
0.0
200
400
Distance (arbitrary)
Fig. 8. Demonstration of the change in the apparent thickness of an anisotropic crystal as the crystal is rotated: x, direction of propagation of the probe beam; E, electric field of light; C, optical axis of the crystal.
to measure the birefringence of a crystal. A 500-,jmthick lithium-niobate (LiNbO3 ) crystal was used as a target. The probe beam was incident perpendicular to the surface that contained the crystal axis. The effective thicknesses were measured as the crystal was rotated around the incident probe beam. Figure 8(a) shows the case in which the direction of the polarization of the probe beam is parallel to the
crystal axis (the e wave is excited). Figure 8(b) shows
. *
*S 0.6-
the case in which the crystal is rotated so that the crystal axis is at 45 deg from the direction of the polarization of the probe beam (both the e and o waves are excited). Figure 8(c) shows the case in which the crystal is rotated so that the crystal axis becomes perpendicular to the direction of the polarization of the probe beam (the o wave is excited). The ordinary index of refraction n 0 of the LiNbO3 crystal is no = 2.2113, whereas the extraordinary index of refraction ne is ne = 2.1361 at = 1.6 jm.li The thickness of the same crystal measured by the ordinary wave appears longer (in terms of wavelength in the medium) than that measured by the extraordi-
0020
0.0
-200
(a)
.. ...........
0 200
Distance (arbitrary) Core
(b)~ ~
250 gm
Fig. 9. Indicatrix of a uniaxial crystal.
Fig. 7.
target.
2498
0.2
02
o.1
0.0~
Distance (arbitrary)
Figure 10 shows the result when the radar was used for locating a score made on an optical IC wafer. Several straight single-mode optical guides were made on a LiNbO3 wafer. The guides were scored at the same location a few times perpendicularly to the optical guides. The cross section of the score measured by a profiler is shown at the top left of Fig. 10. The width of the score is 100 jim, and the depth is 20 im and located 300 jim from the endface of the wafer. The middle graph shows the radar display. A corresponding microscope photograph is shown at the bottom of Fig. 10 to verify the locations of the peaks in the display. The peak midway between the front and back edges of the score may be due to the irregularity of the score caused by repeated scoring.
Conclusions
End face
Score
Optical guide
Fig. 10. Radar display when an optical guide on an optical IC is used as a target.
nary wave. A comparison between the curves in Figs. 8(a) and 8(c) verifies this fact. The ratio of the indices of refraction is nol/ne= 1.035, and the ratio obtained from the experimental data is 1.033 0.004. Good agreement was obtained even with such a thin sample. Referring to the optical indicatrix shown in Fig. 9, we review the law of propagation in a birefringent crystal. When the light is incident on an anisotropic crystal along any direction ON of propagation, only two directions of polarization are permitted; they are the directions of the minor and major axes of ellipses D1 and D2. No other direction of polarization is
permitted.
12
A laser radar that employs neural-network signal processing was constructed, and its performance was evaluated. The resolution of the radar by the neuralnetwork processing was demonstrated to be at least 6 times better than that processed by FFT. The resolution of the radar with neural-network processing is less than 1 im in the lithium-niobate crystal. A few possible applications such as a flaw locator for a silicon wafer, cross-sectional profiling of an optical fiber, determination of the birefringence of a crystal, and a fault locator for optical IC's have been explored.
Appendix A
The derivation of Eq. (16) from Eq. (14) for the case of two frequency steps with two scatterers is as follows: P = EEt
=
with
ekSk
)(eok*Sk*, elk*Sk*),
eOkSk = eSo + eolSl, elkSk = e1oSo+ e11S 1, L (eooSo+ eoiSi)(eoo*So* + eoi*Si*) (eloSo + elSi)(eoo*So* + eoi*Si*) (eooSo+ eoiSi)(eio*So* + ell*Sl*)1 (eloSo + e1iSi)(eio*So* + eln*Si*)] '
2 ISOi eooelo*+
I SoI 2 + IS,
+SOS1*eooeo1*+SSo*eoeoo*
IS 0 12+ IS1J2+SoS1*eioe11*+SSo*e1elo*
Let us focus on the twin peaks in Fig. 8(b) at the location of the back wall of the crystal. One of the twin peaks in Fig. 8(b) lines up with the e-wave peak in Fig. 8(a), and the other twin peak in Fig. 8(b) lines up with the o wave in Fig. 8(c). The appearance of the twin peaks, rather than a single broader peak, confirms the above-mentioned law of propagation in a birefringent crystal.
The difference between the left- and right-hand sides of the equation above becomes even larger if either S0 or S1 has a value of zero. In the case of quantized
1 May 1994 / Vol. 33, No. 13 / APPLIED OPTICS 2499
points with a relatively small number of scatterers, the left-hand side becomes much larger than the right-hand side:
P=ISOI2(1
eooelo*)
2Q
1(e
P= IS1 eoeo*
1
I ) + IS 12ellel*
1,/
= IS12eoeOl + IS12elell
=
k=O
X I Sk 12 ekekt,
2 +
2
2 I Sk
2
where e, e, and ek are phasor vectors associated with the zeroth and first target, respectively, and are
defined as
eoo eok
ISk 2 eOkeOk* |
I
k=O
eOkelk*
2
elkeOk*
k=0
2 ISk1 elkelk*
= elo '
(el
1
2
= F+ G + H.
We assume the following to be Hermitian:
0* P 00 , P 01 * = P 01 , P 1 =
10 *
P 10 , P 11 *
Q=
21Qoo 12=
PT E _
Po -
k=O
1
Sk 2ekekt
2
-2
k=O
ISkI 2 {Pooe0k*eok +
le0k*elk
k=O
1
=
h I2 eokeOk*
2 IPool I
Po
1
k=O
2
ISk
I 2eok *eok 2
-2
k=O
oo* I
o
2
|Sk
k=O
1 2
eOkeOk* +
+ elk*(PloeOk+ Pllelk)}
1(
=-2
'~
21 Qol
I
k=O
|Sk
1
eOkelk =
k=O
(~k ek*
elk*)(P
-2
IPO-
Po
1
I Sk 12eOk*elk k=O
k= =00
|Sk 2 (eok*
P1
i,
poy*I
ISk 2 eOkelk* + |
2
G = -2
,
k=O
1
k=O
ISk
2 1 eOkelk*
k=O
IXISkl 2 ektPek.
Next H is obtained:
1 1
2 S 1 eoj*eoj
1Q1oI =
P10k=O
| I Sk
elkeok
H
ISkI2 elk*eOk k=O
1
IX IX ISEI 2eoieoi*
i=0 j=o
1 1
P 1 2 - p10
1
-
+ 2 12elkeOk* ,
IX
i=o j=0
1 1
S I 2eeoieli* o
Sj
eoj*elj
Po*
IX I=SkIelkeok*
1 s
+
2
I ISk
+
1
I
i=0 j=0
1
1 2 = Pll 21~
12 lkl
k
k=
-1 - _l p11
Sk le
lk
+
1
I I
i=o j=0
1
ISI 2elie1i*
Sj 2elj*ej
Pl-l,*
k=0
1 |SkIekelk +
k=0
X ISk
= z o=0 j=o
IX|Sj I2 Ilj
2[eoieoj*(eoi*eoj + ei*elj)
elkelk* ,
+ elieV*(eoi*eoj + eli*eV)]
2500
E =E j=0 i=0
1 1
+ eli*elj)(eoieoj*
+ elielj*)
References 1. M. K. Barnoski and S. M. Jensen, "Fiber waveguides: Opt. 15, 2112-2115 (1976).
a novel
i=oj=o
1 1
+ eli*eljI
2. W. Eickhoff and R. Ulrich, "Optical frequency domain reflectometry in single-mode fiber," Appl. Phys. Lett. 29, 693-695
(1981). 3. H. Ghafoori-Shiraz 160-162 (1985). and T. Okoshi, "Optical-fiber diagnosis
i=0 j=o
1 1
IX
(eoi*
eli*)
(eoi) i2
using optical-frequency-domain reflectometry," Opt. Lett. 10, 4. J. Nakayama, K. Iizuka, and J. Nielsen, "Optical fiber fault locator by the step frequency method," Appl. Opt. 26,440-443
(1987). 5. K. izuka, Y. Imai, A. P. Freundorfer, R. James, R. Wong, and
i=oj=o
The authors are grateful to I. Kitano of Nippon Sheet Glass, T. Sueta, M. Izutu, and H. Nishihara of Osaka University, Y. Sakauchi of Sanyo Electric Company, and J. Minowa of Sumitomo Cement Company for providing us with various kinds of target to try with our fault locator. They are also grateful to Y. Imai for suggesting that we use the fault locator as a device to determine the anisotropy of a crystal. K. Kawashima of Optical and Radio Communication Research Laboratories was helpful in preparing the samples. H. Murata of Furukawa Electric Company suggested the possible use of a monitor during the fabrication of the fiber. They acknowledge Mary Jean Giliberto for assistance in preparing the manuscript. The authors are grateful to T. Manabe and H. Shimodahira of ATR for technical discussions on neural-network theory and Y. Furuhama of ATR for enthusiastically supporting this project.
S. Fujii, "Optical step frequency reflectometer," J. Appl. Phys. 68, 932-936 (1990).
6. R. C. Youngquist, S. Carr, and D. E. N. Davis, "Optical a new optical evaluation coherence-domain reflectometry:
8. J. J. Hopfield, "Neurons with graded response have collective computational abilities like those of two-state neurons," Proc.
Natl. Acad. Sci. USA 81, 3088-3092 (1984).
9. T. Manabe and S. Fujii, "Array processingwith neural networks for multiple emitter bearing estimation," in 199OIEEEAntennas and Propagation Symposium Digest (Institute of Electrical and Electronics Engineers, NewYork, 1990), pp. 1458-1461.
10. R. W. Southworth and S. L. Deleeuw, Digital Computation and Numerical Methods (McGraw.Hill, New York, 1965).
2501