Vous êtes sur la page 1sur 4

Analysis and Implementation of a Novel Single-Channel Direction-Finding Method

Nathan Harter, John J. Keaveny, Swaroop Venkatesh and R. Michael Buehrer


Mobile and Portable Radio Research Group (MPRG), Virginia Tech, Blacksburg, VA 24061 Email: {nharter, jkeaveny, vswaroop, rbuehrer}@vt.edu
Abstract In this paper, we propose and analyze a novel single channel direction nding (DF) algorithm based on phase-locked loops (PLL). The algorithm is evaluated via Matlab simulation to determine the accuracy of the estimated angle-of-arrival (AOA) of a target signal in an AWGN channel. The proposed algorithm is shown to provide exceptional accuracy. Additionally, we present results from a hardware implementation using a software radio platform and an eight-element circular array which validate the simulation results.

relatively benign environment for proof-of-concept testing. This paper will show that the basic version of the algorithm can result in a signicant computational burden, thus we investigate a low-complexity approach and demonstrate its performance. It will be shown that a signicant computational reduction can be achieved with minimal performance penalty. II. P HASE -L OCKED L OOP A LOGORITHM FOR S INGLE -C HANNEL D IRECTION F INDING Consider a uniform circular array of Na elements that are sampled successively in a manner similar to the PseudoDoppler technique [3], [4]. However, after sampling, the signal samples are partitioned into one of Na logical data streams. Each data stream is then processed with its own independent digital PLL. The PLLs track the common phase offset 0 plus the offset that is specic to each antenna to within an ambiguity related to the data modulation. This ambiguity can be removed by tting the resulting Na values to a sinusoid as we will discuss. The architecture is illustrated in Figure 1. The received signal entering the ith PLL is given by 2i 2r cos + n(t), r(t) = m(t) cos c t + 0 + Na where m(t) represents the modulating signal, c is the carrier frequency, is the carrier wavelength, r is the radius of the circular antenna array, is the AOA that needs to be estimated and n(t) represents noise and other interference effects.
DOA
Commutating Switch

I. I NTRODUCTION Direction nding (DF) techniques have classically been based on multiple-antenna systems employing multiple receivers. Classic techniques such as MUSIC [1], [2] and ESPRIT use simultaneous phase information from receivers connected to each antenna to estimate the angle-of-arrival (AOA) of the signal of interest. In many scenarios (e.g., handheld systems), however, multiple receivers are impractical. Thus, single channel techniques are of interest, particularly in mobile scenarios. Although the amount of existing research is considerably less than multi-channel direction nding, single channel direction nding techniques have been previously investigated. Specically, the Watson-Watt technique [4] and the Pseudo-Doppler technique [3], [4] are among the most prominent and are classic analog techniques. In this work we investigate a new single channel direction nding technique that takes specic advantage of digital capabilities. Specically, we propose a method that uses a software bank of Phase-Locked Loops (PLLs) in combination with a circular array. Our method is similar to a digital version of the Pseudo-Doppler method in that it samples antennas in a circular array. In the proposed approach the sampled data is fed to a bank of PLLs which track the phase on each element. The parallel PLLs are implemented in software and their outputs are fed to a signal processing block that estimates the AOA. This paper presents the details of the new algorithm and assesses its performance through simulations. We also describe the implementation of the algorithm on a single channel software radio and present measured performance results. The implementation of the algorithm is on a single-channel DSPbased software radio with a homemade eight element circular antenna array. The implementation was tested using a CW signal at 1.57 GHz in a low multipath environment. The performance of the prototype is compared to results provided by simulation. Implementation results focus on CW measurements in a

Target

Bank of PLLs

DOA Estimator

Fig. 1.

System schematic for PLL-based AOA estimation

The linearly modulated data m(t) is detected by mapping (minimum-distance) the phase to either or pi (in the case
0-7803-8966-2/05/$20.00 2005 IEEE

IEEE Communications Society / WCNC 2005

2530

of BPSK) and the PLL outputs a modulo 2 value of the phase that lies between and . The phase of the signal input to the ith PLL is given by 2i 2r cos . (1) i = 0 + Na Given the values of i and the common phase offset 0 , we could estimate the AOA using = arccos i 0 2i . A Na where A = given by
2r .

PLLs is given by i

2i = 2A sin sin Na Na Na + (ni ni1 )

However, the phase detected by the PLL is

2i 2r cos i = 0 + + ni , Na

i = 0, 1, , Na 1.

where ni 0, 1, 2, chosen such that i lies in [, ] 2r due to phase-wrapping. Now since A = > based on the design parameters, and the phase offset 0 lies between [, ], in general i 6= i . Therefore calculating i based on i would lead to incorrect results. We therefore need to nd a method to retrieve i from i or, in other words, to nd the sequence ni , (i = 0, 1, , Na 1). Once this is found, given that does not change, the same sequence ni , can be added to i to obtain i . A simple algorithm to solve the aforementioned problem would be to store all possible (58 ) sequences {ni } in a table. We then add these sequences to the received i , eliminate the DC offset (i.e., 0 ) and exhaustively search for the correct sequence by minimizing the squared error. However, this is a very computationally demanding algorithm, and if changes quickly, the receiver would need to compute this search process continuously. Hence, we would like a less computationally expensive algorithm with a smaller search-space. We now suggest an algorithm which has a much smaller search-space and is far more computationally efcient. The idea is as follows: the expression corresponding to the difference between two consecutive antenna phases is given by 2i sin i = i i1 = 2A sin Na Na Na (2) Let us call this the target difference curve. From (2), given r the value of A (= 1.28 for Na = 8, = 0.64), we can nd the maximum difference Dmax between any two points of the actual phase i . In particular, for Na = 8, Dmax = 3.078. This means that the magnitude of any point on the target difference curve cannot exceed Dmax . A simple method to nd from the difference curve is take the Na -point FFT k , (k = 0, 1, , Na ) of the sequence i (i = 0, 1, , Na ), from which the can be estimated from the phase of second FFT coefcient: (3) = (6 k ) |k=1 The difference between two consecutive phases output by the

We call this the given difference curve. Our aim is be able to nd the target difference curve from the given difference curve and thereby estimate using (3). We see from the above equations that we have done away with the phase offset 0 but now the difference ni = (ni ni1 ) has a much larger (integer) range than ni . However, since we know that the maximum value of i is Dmax , we can successfully eliminate several of these possibilities. We see than since Dmax < for Na = 8 and the given value of A, we can eliminate all integer values of ni , except two: one point belonging to the given difference curve and the other corresponding to either adding or subtracting in order to constrain the point to the interval [, ]. This means that for every i, there exist only two possible points which lie on the target difference curve. For Na = 8, since we require points corresponding to i = (0, 1, , 7), we have only 2Na = 256 possible curves corresponding to all possible i . This searchspace is much smaller than the previous search-space (58 ). We divide the possible AOA range in radians [, ] into a number of bins, and select the sequence ni and binned angle for which the mean squared error between the created difference curve and the difference curve calculated from (3) is minimized. Once we have the sequence ni we can create the target difference curve from the given difference curve and use (3) to compute the angle . III. A LGORITHM S IMULATION In order to measure the performance of the method described in the previous section, we simulated the algorithm using Matlab. In our initial simulations, the algorithm was simulated in an environment without multipath or modulation at an SNR of 10 dB and r = . The performance will be 2 measured by performing numerous AOA estimate computations for each of several true AOAs and determining the mean and variance of the error. This paper also characterizes the RMS error in the system simulations for varying SNR, array radius and elevation angle. A. PLL-Based Algorithm in AWGN Fig. 2a shows the error (in degrees) of our PLL-based algorithm and Fig. 2b shows the standard deviation of the estimate for each AOA at an SNR value of 10 dB. It is observed that our algorithm produces extremely accurate estimates of the AOA. The RMS value of the estimated AOA error averaged over all AOA values was found to be 0.07 . Further, simulation suggests that the estimator is unbiased. Fig. 3 shows a summary of the RMS errors for SNR values from 0 dB to 10 dB. The algorithm is shown to provide a very accurate estimate down to approximately 4 dB, below which the error begins to increase rapidly. The RMS error is in some ways misleading. The error increases as SNR decreases due

IEEE Communications Society / WCNC 2005

2531

0-7803-8966-2/05/$20.00 2005 IEEE

Average Error ( )

0.1 0 0.1 0.2 0 50 100 150 200 True DOA ( ) 250 300 350

180 120 60 0 60 120 180

Error ( )

45

90

135

180 225 True DOA ( )

270

315

360

0.8 Estimate ( ) 0.75 0.7 0.65 0.6 0 50 100 150 200 True DOA ( ) 250 300 350 Error ( )

180 120 60 0 60 120 180

Error ( )

45

90

135

180 225 True DOA ( )

270

315

360

180 120 60 0 60 120 180

45

90

135

180 225 True DOA ( )

270

315

360

Fig. 2. Figure 2a (top): Mean of Estimated AOA versus true AOA, Figure 2b (bottom): Standard Deviation of estimated AOA versus true AOA.

Fig. 4. Scatter plot of AOA estimate error samples versus true AOA for (a, top) 4 dB SNR, (b, center) 2 dB SNR and (c, bottom) 0 dB SNR.

to the fact that inaccurate values are chosen more frequently. Such values may be very far from the true value. This is demonstrated in Fig. 4 which is a scatter plot of the estimate error samples for SNRs of 4 dB, 2 dB, and 0 dB. It shows that the variance of the estimate increases not because the error of every estimate increases, but because it becomes more likely to choose the wrong curve that represents more error. Moreover, the grouping of the error samples shows that the estimator exhibits denite trends in the curves it selects.
10
1

elevation angle. It shows that beyond 40 the estimate becomes unreliable. The elevation angle reduces the signal strength on the array and thus reduces the effective SNR. The higher the elevation angle, the lower the effective SNR.
10
3

10

RMS Error ( )

10

10

10 RMS Error ( )

10

10

20

30

40 50 60 Elevation angle ( )

70

80

90

10

Fig. 5.
2

AOA Estimate RMS Error versus Elevation Angle ().

10

5 6 SNR ( dB )

10

C. Impact of Array Size


r As can be seen from (1), the value of is an important aspect of the estimator. Simulations were run for various array sizes relative to the wavelength. Fig. 6 summarizes the RMS r error of the estimator vs. the value of . It can be seen r that increasing values in the range of 0.1 to 0.6 improves performance slightly. Outside of that range, the estimate error increases rapidly due to the element spacing exceeding /2. Large array sizes can improve performance in the presence of multipath, but must be coupled with an increase in the number of elements. We are currently investigating these aspects of the system.

Fig. 3.

AOA Estimate RMS Error versus SNR (dB)

where is the elevation angle and 0 is parallel to the plane of the array. Elevation angles were examined in 5 degree steps from 0 (as in previous results) to 90 (fully perpendicular to the plane of the array) with an SNR of 10 dB. The 90 case was not expected to produce any meaningful results as the angle completely zeros out the phase values at each antenna and leaves the algorithm with nothing but noise to compute its estimate. Fig. 5 summarizes the RMS error in the algorithm vs.

B. Impact of Elevation Angle When considering a signal impinging upon the array from an angle other than the horizon, (1) becomes 2i 2r cos() cos . (4) i = 0 + Na

IV. S OFTWARE R ADIO I MPLEMENTATION The implementation of the PLL algorithm was performed on a TI-based software radio provided by DRS Technologies. The radio allows one to develop algorithms in the C programming language. To facilitate the single channel direction nding system, we designed a commutative switch in order to pass

IEEE Communications Society / WCNC 2005

2532

0-7803-8966-2/05/$20.00 2005 IEEE

10

10

10

10

True AOA ( ) 0 22.5 45 67.5 90 112.5 135 157.5 180

Mean AOA Est.( ) 0.4 25.5 49.0 69.3 91.4 113.9 132.2 157.2 178.8

Mean Est. Error( ) 0.4 3.0 4.0 1.8 1.4 1.4 -2.8 -0.3 -1.2

Est. Std. Deviation( ) 0.088 0.1789 0.0775 0.0837 0.0632 0.0837 0.0775 0.1449 0.0775

RMS Error ( )

10

0.1

0.2

0.3

0.4

0.5 0.6 r/ ( unitless )

0.7

0.8

0.9

TABLE I AOA ACCURACY FROM I NDOOR M EASUREMENTS

Fig. 6.

RMS error of OA estimate versus

r .

10

the signals from a homemade circular antenna array (r = /2 at 1.5GHz) into the receiver which implemented the bank of PLLs in software. In order to control the commutative switch, we use an output of the software radio receiver to send a signal that has eight different voltage levels to an Analog-to-Digital Converter (ADC). Each voltage level corresponds to one antenna. The parallel output bits from the ADC are buffered to a Xilinx CPLD (9800XL). The CPLDs main function is to compute the correct switch output sequence given the input from the ADC. The output from the CPLD is buffered and passed to the TTL driven SPDT and SP4T RF switches. These RF switches are connected to the eight-antenna circular array. A set of initial measurements were taken in a laboratory environment using a homemade anechoic chamber to reduce multipath. The transmitter-receiver distance was set at about 4 feet. The transmitting antenna that was used was a WJ directional antenna. The initial transmitter power was set at 10 dBm. The accuracy of the software radio implementation in a sample measurement set is shown in Table IV. These initial results show that we have a maximum error of 4 and a minimum error of 0.3 with estimate variances of less than 1 . The measurements were taken at a transmit power of 10dBm with a Tx-Rx separation of 1.1 meters. This AOA accuracy held until the transmitter power was reduced to less than 70 dBm. Fig. 7 shows the variance in the estimate as the transmitter power was varied. Human error can be a factor in the calibration of these measurements since the transmitter had been lined up by eye with the aid of a laser level. This explains the mean error since the estimator is unbiased. V. C ONCLUSION In this paper we presented a new single channel direction nding technique based on a vector of Phase-Locked Loops (PLLs) in combination with a homemade circular antenna array. The switched antenna signal is sampled and fed to a bank of PLLs which track the phase observed on each antenna. The PLL outputs are then fed to a signal processing block that estimates the AOA.

0 22.5 45 67.5 90 112.5 135 157.5 180 Mean

() 10
1

70

60

50

40 30 20 T power ( dBm )
X

10

10

Fig. 7. Experimental AOA estimate standard deviation versus transmit power for various AOAs (mean of standard deviations: thick solid line)

The algorithms performance was evaluated via simulation and hardware experimentation. Simulations have shown that the PLL-based algorithm can estimate the angle of arrival of a CW signal to within 0.1 . The implementation of the PLLbased algorithm on a software radio platform has estimated the AOA of a CW signal to within 0.3 - 4 at a Tx-Rx distance of 1.1 meters with the transmitter power set as low as -70dBm. Further research will investigate the effects of multipath, interference cancellation, frequency offset compensation and increasing the number of elements in the circular array. ACKNOWLEDGMENT This work was sponsored by DRS Technologies, Gaithersburg, MD. The authors gratefully acknowledge the assistance provided by DRS engineers including Tom Potter, Guy Zaybekian, Jared McGehee, and Jeff Silvey. R EFERENCES
[1] R.O. Schmidt, Multiple Emitter Location and Signal Parameter Estimation, Proc. of RADC Spectrum Est. Workshop, Grifss AFB, NY, pp.243258, 1979. [2] R.O. Schmidt, Multiple Emitter Location and Signal Parameter Estimation, Proc. IEEE Trans. On Antennas and Propagation, Vol. AP-34, No.3, Mar. 1986. [3] D. Peavey and T. Ogunfunmi, The Single Channel Interferometer Using a Pseudo-Doppler Direction Finding System, Proc. of 1997 IEEE Conference on Acoustics, Speech, and Signal Processing, Vol. 5, pp. 4129 4132, April 1997. [4] R. Hammerle, Factors Limiting the Accuracy of Doppler and Adcock Direction Finding Systems, Proc. of 1989 IEE Colloquium on Passive Direction Finding, pp.3/1-3/13, Jan. 1989.

IEEE Communications Society / WCNC 2005

2533

0-7803-8966-2/05/$20.00 2005 IEEE

Vous aimerez peut-être aussi