You are on page 1of 4

Proceedings of the 9th European Radar Conference

Accuracy limits of a K-band FMCW radar with


phase evaluation
Steffen Scherr, Serdal Ayhan, Mario Pauli, Thomas Zwick
Institut für Hochfrequenztechnik und Elektronik, KIT,
Kaiserstr. 12, 76131 Karlsruhe, Germany
{steffen.scherr, serdal.ayhan, mario.pauli}@kit.edu

Abstract—The FMCW radar technology offers the possibility approach, as demonstrated in [2], [3]. Even though different
to achieve very high target accuracy mainly limited by the FMCW radars for high accuracy measurements have been
SNR. In this paper the CRLB for an FMCW radar with phase realized, there is a lack of theory showing the maximum
evaluation is derived and different algorithms reaching the CRLB
are presented. The algorithms are tested in a simulation and the capabilities of an FMCW radar with phase evaluation. In this
simulation results are compared with measurement results of an paper the accuracy limits for such a radar will be shown and
existing FMCW radar. The used algorithms are the zero padded calculated for a realized radar.
FFT, the chirp z-transform and the esprit algorithm. The FMCW The paper has the following structure: Section II derives the
radar works in the K-band with a center frequency of 24 GHz general Cramer-Rao lower bound (CRLB) with respect to the
and a bandwidth of 1 GHz. An accuracy of 65 µm is reached.
Index Terms—CRLB, FMCW radar, range detection, high FMCW radar application. In section III the used algorithms
accuracy. and the simulation results are presented. Section IV compares
the theoretical accuracy limits with the accuracy reached in
I. I NTRODUCTION measurements. In section V, the most important results are
summarized.
The accurate measurement of positions is necessary for
numerous applications such as liquid level measurements, II. CRLB FOR FREQUENCY AND PHASE
reference or guiding systems and controlling of industrial An FMCW radar measures distances by sending linear
machines. In general contact-free devices are favored since ramps (frequency modulated signals) to a target. To simplify
they are less prone to damage than cable-connected or probe- the calculation of the FMCW limits, the target is not moving
based sensors. Additionally non contacting sensors are the only so that no doppler shift in the frequency occurs. The received
choice if explosive, volatile or high temperature substances are signal is time delayed and damped, but has the same linear
involved. Furthermore the devices need to work properly even ramp form as the transmitted signal if no nonlinear distortion
if the measurement path is blocked by smoke, dust or flying occurs. Both signals are mixed together and the resulting signal
debris. is a sine wave with the amplitude A, the phase ϕb and beat
Several solutions exist for this task: laser sensors, ultrasonic frequency fb .
devices and microwave (radar) sensors. Laser sensors offer an
unmatched accuracy [1] but may not work if a measurement s(t) = A sin(2πfb t + ϕb ) (1)
through a medium different from air (e.g. oil in hydraulic
systems) is required. Also they might be difficult to be in- The beat frequency is directly related to the target distance
tegrated into an industrial machine due to the necessary space B 2r
around the sensor head. Ultrasonic devices suffer from similar fb = (2)
T vgr
problems. A good choice for hostile measuring environments
is a microwave radar system. Microwave signals can penetrate depending on the bandwidth B, sweep time T , target
most substances with acceptable losses and the realization distance r and group velocity vgr . Hence the problem of
is possible in advantageous monolithic microwave integrated achieving a high accuracy stems from the frequency estima-
circuit (MMIC) technology. tion. As sampling the frequency is a discrete process, eq. 2
Continuous wave (CW) and frequency modulated continous can be rewritten to
wave (FMCW) radars are often used for high accuracy mea-
fb B 2r
surements. CW radar devices allow relative distances to be = (3)
fs N vgr
determined with high accuracy, but the unambiguous range
is limited by the wavelength. FMCW radars, however, may with N being the number of samples taken in the ob-
resolve multiple targets but the accuracy is insufficient in re- servation window T and the sampling frequency fs . The
lation to the increasingly higher industrial requirements. Hence CRLB defines the limits of estimating a discrete sine wave
the combination of both systems by combining frequency [4] depending on the SNR η and may be applied to radar
and phase evaluation methods has proven to be a successful problems [5], [6]:

978-2-87487-029-3 © 2012 EuMA 246 31 Oct - 2 Nov 2012, Amsterdam, The Netherlands
  phase estimation is more effective in increasing the accuracy
fb 12 than an improved frequency estimation. For a typical SNR of
var ≥ (4)
fs (2π)2 ηN (N 2 − 1) 30 dB the maximum accuracy would be around 290 µm for
Eq. (4) holds true if only one sine wave in white gaussian the frequency estimation and around 7 µm if the additional
noise exists and no other additional effects like interference, phase estimation is used.
leakage or nonlinear distortions occur. The CRLB for complex
sine waves is different [7], but as no IQ demodulator is used in −10
the FMCW radar, the assumptions are sufficient for a proper CRLB frequency
comparison. Using eq. (3), the variance of the discrete beat CRLB phase
−20

3σ deviation in dBmeter
frequency can be expressed as
 
fb 4B 2 −30
var = 2 2 var {r} (5)
fs N vgr
−40
By combining eq. (4) and eq. (5) the simplified CRLB
(N >> 1) for an FMCW radar is as follows: −50
2
3vgr −60
var {r} ≥ (6)
(2π)2 ηN B 2
It can be seen that a large bandwidth, a high SNR and −70
−10 0 10 20 30 40 50
the absolute number of samples increase the maximum pos- signal to noise ratio η in dB
sible accuracy of an FMCW radar. However improving these
properties might be difficult. If instead of the beat frequency Fig. 1. CRLB for accuracy estimation
the phase of the beat frequency is evaluated a much higher
accuracy can be reached, but the 2π ambiguity problem has III. I MPLEMENTED ALGORITHMS AND SIMULATION
to be taken into account. The phase is related to the distance
Section II shows that the phase evaluation can reach a much
via the following equation
higher accuracy than the frequency evaluation at the same
2r SNR. But as described by [3] the frequency estimation has to
ϕb = 2πfmin (7) be better than λmin to avoid problems with the unambiguity
vgr 2
range of the phase. Therefore the frequency algorithm needs
with the ramp start frequency fmin [3]. Furthermore the
to be resource efficient, fast and sufficiently accurate. In the
CRLB for the phase of a sine wave is given by [4]
following three frequency estimation algorithms are compared.
2(2N − 1) The first frequency method is the Fast Fourier Transform
var {ϕb } ≥ (8) (FFT) with zero padding. In this method the frequency reso-
ηN (N + 1)
lution and therefore the accuracy is increased by adding zeros
The variance of the beat frequency phase (eq. 7) is to the sampled signal. However this method proves to be
4(2π)2 fmin
2 inefficient, since it tends to get biased for very high SNR
var {ϕb } = 2
var {r} (9) values and needs large amounts of memory for very high zero
vgr
padding values.
Combining eq. (8) and eq. (9) leads to the simplified CRLB A more efficient method is the chirp z-transform [8]. With
(N >> 1) for phase evaluation in an FMCW radar. the chirp-z transform it is possible to increase the number of
bins in a selected area instead of increasing the number of bins
(vgr )2
var {r} ≥ 2 ηN (10) in the whole sample vector. The proposed algorithm first calls
(2π)2 fmin a FFT for a coarse search of the beat frequency. In the next
Eq. (10) shows that the accuracy can be improved if the step it uses the chirp z-transform to increase the resolution
number of samples or the ramp start frequency is increased. in the area around the beat frequency. The accuracy can be
To get more concrete values the CRLB is estimated for further increased by recursively calling the chirp z-transfrom.
the used FMCW radar (see table I and section IV). As the The esprit algorithm is a so called high resolution method
CRLB is calculated assuming a normal data distribution, 99.7 and the third algorithm used to estimate the frequency [9].
percent of the values lie within three standard deviations 3σ. This algorithm is able to separate two targets that are too
Due to the large range of values the logarithm of the 3σ close together to be resolved by Fourier-based methods. The
deviation is taken and as the standard deviation has the unit drawback of this method is that the number of expected
meter, the 3σ deviation in the plot has the unit dBmeter. targets must be known, which is normally the case in high
The CRLBs are shown in Fig. 1. The bounds for the phase accuracy frequency estimation scenarios. This algorithm works
estimation algorithm are about 20 dB lower compared to the as follows: It estimates a special structured covariance matrix
frequency estimation at the same SNR. Hence improving the from the sampling vector. This matrix is splitted into a

247
signal and a noise subspace. The eigenvalues of the signal tends to have a value, which cannot be adapted to the sampling
subspace are the frequency estimates. The implemented esprit frequency by an integer value. In this case a non matching
algorithm only needs a small amount of computational power, integer value would lead to a biased estimated beat frequency.
as the covariance matrix is built several times over a part of Therefore it is necessary to use resampling algorithms in-
the measured samples and the result of each eigenvalue is stead of pure downsampling. The multi-rate or also called
averaged. polyphase algorithms proposed by the authors in [11] fulfill
The algorithms were tested in a simulation with parameters the aforementioned requirement of a fractional downsampling.
according to the measurement system. The chosen parameters But the necessary filtering in the algorithm leads to a slight
are depicted in table I. error at the beginning and the end of the sampling vector,
resulting in a biased beat frequency at high SNR values.
beat frequency fb sampling frequency fs samples N
Fig. 3 shows the simulation result of a resampled signal with
25 kHz 1 MSa/s 500
bandwidth B sweep time T group velocity vgr the same simulation parameters as in the previous simulation
1 GHz 500 µs 2.4727 · 108 m/s (see table I) but at a different beat frequency of 24 kHz. The
TABLE I biasing can be clearly seen. Due to the estimated SNR of the
CRLB SIMULATION PARAMETERS used FMCW radar of 30 dB (see section IV), the biasing at
high SNR values can be neglected, as an accuracy of around
2.3 mm is sufficient for a phase evaluation.
The ratio between the beat frequency and the sampling
frequency ffsb is at 0.025. As the ratio between the beat
frequency and sampling frequency should be around 0.25 to −10
FFT
avoid biasing effects from the mirror frequency and to avoid CZT
aliasing [4], [10], the signal must be downsampled by a factor 3σ deviation in dBmeter ESPRIT
of 10. −20 CRLB frequency
The chosen standard deviation of the beat frequency results
in a range standard deviation of around 3.11 mm. If the
standard deviation of the beat frequency is set too low, it is −30
estimated by the frequency algorithm and no comparison with
the CRLB is possible. Fig. 2 shows the simulation results.
The performance of the different algorithms is similiar. If the −40
SNR is too low, the algorithms are not able to determine the
frequency and the 3σ deviation is much higher. This effect is
called threshold effect [5]. −50
−10 0 10 20 30 40 50
Only 50 out of 500 samples are used for the frequency signal to noise ratio η in dB
estimation, because the signal was downsampled. Nevertheless
it is possible to reach the CRLB by appropriate downsampling Fig. 3. resampled beat frequency signal (simulation results)
as long as the downsampled signal consists of at least four
points.
IV. M EASUREMENT S ETUP AND R ESULTS
10 The measurement setup is depicted in Fig. 4. The FMCW
FFT radar consists of a K-band VCO with a center frequency of
CZT
0 ESPRIT 24 GHz and the same parameters as shown in table I. To
3σ deviation in dBmeter

Range CRLB minimize multi-path propagation problems and to achieve a


−10 comparable one target scenario a waveguide with a sliding
short is used.
−20
measurement line
−30

directional
−40
coupler

−50
−10 0 10 20 30 40 50
signal to noise ratio η in dB FM modulator

Fig. 2. simulation results comparable to the used FMCW radar mixer

However in practical measurement setups the beat frequency Fig. 4. FMCW radar measurement setup (block diagram)

248
The waveguide has a cross section of 11 times 11 mm. V. C ONCLUSION
A reference is necessary to verify the FMCW radar mea- In this paper the CRLB was calculated for a K-band
surements. For this purpose a laser with its target connected FMCW radar and compared with measurement results. Several
to the radar target is used. Currently it is only possible to algorithms were presented that are able to reach the FMCW
carry out relative distance measurements with this setup. The CRLB for frequency estimation. The measurement and the
measurement itself is taken over a distance of 100 mm with simulation results show that the derived CRLBs hold for
1 mm steps, measuring 60 times at each position. frequency and phase evaluation and can be used to design
Fig. 5 shows the cumulative distribution function at one an FMCW radar for high accuracy applications. To further
position of the waveguide for the combinated phase and improve the accuracy of the FMCW radar, a W-band system
frequency evaluation. 99.97 percent of the position errors are with a much larger bandwidth is currently developed. As the
smaller than 35 µm. The SNR of the FMCW radar is at about FMCW CRLB shows that the bandwidth and the start ramp
30 dB for this measurement position, so the CRLB would frequency are the main modifiable parameters that influence
allow a higher accuracy, but does not take errors into account the accuracy, an improvement can be expected.
from nonlinearities in the ramp and multiple reflections.
R EFERENCES
[1] R. Anderson, C. Huang, and N. Buholz, “Sensing the position and
1 vibration of spacecraft structures,” J. Guidance and Contr., vol. 4, no. 2,
number of measurement values in %

pp. 141–147, Mar. 1981.


[2] G. Woods, D. Maskell, and M. Mahoney, “A High Accuracy Microwave
0.8 Ranging System for Industrial Applications,” IEEE Transactions on
Instrumentation and Measurement, vol. 42, no. 4, pp. 812–816, Aug.
1993.
0.6 [3] S. Ayhan, M. Pauli, T. Kayser, S. Scherr, and T. Zwick, “FMCW
radar system with additional phase evaluation for high accuracy range
detection,” in European Radar Conference (EuRAD) 2011, Manchester,
0.4 Dec. 2011, pp. 117–120.
[4] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory, 1st ed. Prentice Hall, 1993.
[5] M. Braun, T. Shu, C. Sturm, and F. Jondral, “On the Single-Target Ac-
0.2 curacy of OFDM Radar Algorithms,” in IEEE International Symposium
on Personal, Indoor and Mobile Radio Communications, Canada, Sept.
2011, pp. 794–798.
0 [6] S. Schuster, S. Scheiblhofer, L. Reindl, and A. Stelzer, “Performance
0 1 2 3 4
Evaluation of Algorithms for SAW-Based Temperature Measurement,”
position error in m −5
x 10 IEEE Transactions on ultrasonics, ferroelectrics, and frequency control,
vol. 53, no. 6, pp. 1177–1185, Jun. 2011.
Fig. 5. cumulative distribution function for a measurement at a single position [7] Q. Guoqing, L. Xueling, and T. Yanyan, “Range Estimation Accuracy
Analysis of the FMCW Level Radar,” in 8th International Conference
on Signal Processing (ICSP), Beijing, Nov. 2006.
Fig. 6 shows the 3σ deviation over 100 measurement [8] L. Rabiner, R. Schafer, and C. Rader, “The Chirp z-Transform algo-
positions for the combinated frequency and phase evaluation. rithm,” IEEE Transactions on Audio and Electroacoustics, vol. 17, no. 2,
pp. 86–92, Jun. 1969.
Due to the low losses in the waveguide, the average of the 3σ [9] A. Paulraj, R. Roy, and T. Kailath, “Estimation Of Signal Parameters
deviation only slowly increases. Via Rotational Invariance Techniques - Esprit,” in nineteenthe Asilomar
Conference on Circuits, Systems and Computers, Pacific Grove, Nov.
−5
1985, pp. 83–89.
x 10 [10] C. Candan, “A Method For Fine Resolution Frequency Estimation From
7 Three DFT Samples,” IEEE Processing Letters, vol. 18, no. 6, pp. 351–
354, Jun. 2011.
[11] R. Crochiere and L. Rabiner, Multirate Digital Signal Processing.
3σ standard deviation in m

6 Prentice Hall, 1983.

2
0 20 40 60 80 100
relative measurement position in mm

Fig. 6. 3σ deviation over 100 measurement positions

249

You might also like