710

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

Weighted Energy Detection for Noncoherent Ultra-Wideband Receiver Design

Feng Wang, Zhi Tian, Senior Member, IEEE, and Brian M. Sadler, Fellow, IEEE
AbstractFor ultra-wideband (UWB) impulse radios, noncoherent energy detectors are motivated for their simple circuitry and effective capture of multipath energy. A major performancedegrading factor in energy detection is the noise oor, which is aggravated for low-duty-cycle UWB signals with a large timebandwidth product. In this paper, weighted energy detection (WED) techniques are developed for effective noise suppression. The received signal is processed by a set of parallel integrators, each corresponding to a different integration time-window within a symbol period. The outputs of these integrators are weighted and linearly combined to generate decision statistics, while the weights are determined by the signal power collected from the corresponding integrators to improve the effective signal to noise ratio. The WED principle is applied to all phases of receiver processing, including signal detection, timing synchronization and data demodulation. For each phase, the optimal linear detector parameters, including decision thresholds and weighting coefcients, are derived analytically. Simulations show that the proposed noncoherent WED receiver enhances the bit-error-rate performance compared to conventional energy detectors. Index TermsUltra-wideband (UWB), weighted energy detection, channel estimation, synchronization, decision direction.

I. I NTRODUCTION

N recent years, ultra-wideband (UWB) communication has become a competitive technology for short-range, lowpower wireless applications [1]. UWB impulse radios convey information over repeated ultra-short pulses, which occupy large bandwidth and endow signals with ne resolution [2]. Since UWB radios operate at extremely low transmission power density under FCC spectral regulations, they open up a host of new wireless services capable of overlay with legacy narrowband systems. Interest in UWB is further motivated by several attractive features including: low probability of interception and detection, enhanced capability to penetrate through obstacles, potentially high user capacity, and robustness against multipath fading [1], [2]. Realizing the potential of UWB radio technology calls for cost-effective receiver design under practical implementation constraints. Conventional coherent UWB receivers, e.g. RAKE, have to sample and operate at or above the Nyquist

Manuscript received August 3, 2010; accepted September 30, 2010. The associate editor coordinating the review of this paper and approving it for publication was X. Wang. Part of this work was presented at the IEEE SPAWC Conf. in 2005. Z. Tian was supported by U.S. NSF Grant No. CCR-0238174. F. Wang was with Michigan Technological University. He is now with Arizona State University (e-mail: fwang28@asu.edu). Z. Tian is with the Department of Electrical and Computer Engineering, Michigan Technological University, Houghton, MI, USA (e-mail: ztian@mtu.edu). B. Sadler is with the Army Research Laboratory. Digital Object Identier 10.1109/TWC.2010.120310.101390

rate that can be prohibitively high in the UWB regime, reaching a minimum of hundreds of MHz or even multi-GHz. Further, they need to employ a large number of RAKE ngers and accurately estimate the dense multipath channel in order to capture a considerable portion of the signal energy [3]. To circumvent these drawbacks, suboptimal receivers have been proposed for low-cost, low-data-rate applications, using either auto-correlation based receivers, e.g. transmit reference [4] and differential detection [5], [6], or noncoherent receivers, e.g. energy detectors [7]. In general, auto-correlation based receivers hinge on correlating two different segments of the received signal, one of which serves as the template waveform for the other. Such receivers do not require channel estimation, but face hardware challenges in implementing accurate waveform delay. Noncoherent energy detection, on the other hand, stands out as a simple radio architecture for hardware implementation in the UWB regime. Despite its simplicity, conventional energy detection may not work well for UWB communications due to detrimental noise effects. Because UWB impulse radios typically operate at low duty cycle, the symbol period can be much larger than the time-width of the transmitted short pulses, giving rise to a large time-bandwidth product. Integrating the squared received waveform over the entire symbol period, a conventional energy detector may unwittingly capture the unwanted energy of the noise-only portion of the received signal, which causes a high noise oor proportional to the time-bandwidth product [8]. When the symbol epoch time and channel delay spread are coarsely known, noise reduction is possible by appropriately selecting a small integration time-window [9], [10]. However, this strategy relies on timing information, and hence is not applicable during the signal detection and timing synchronization stages. In order to mitigate the noise effects, this paper develops a weighted energy detection (WED) technique that employs a set of parallel integrators, each collecting the portion of the received signal energy in one of the consecutive non-overlapping time intervals per symbol period. The output of each integrator is weighted and linearly combined to generate the decision statistic, and then compared with a decision threshold for decision making. Our general design principle is to assign large weights to those integrator outputs with strong signal components, and very small weights to integrators collecting mostly noise within their integration time-windows. By doing so, the effective signal to noise ratio (SNR) at the receiver output can be improved. Noticeable SNR improvement has been demonstrated in a conference version of this work for signal detection and symbol demodulation purposes [11]. This paper extends the development to provide

c 1536-1276/11$25.00 2011 IEEE

WANG et al.: WEIGHTED ENERGY DETECTION FOR NONCOHERENT ULTRA-WIDEBAND RECEIVER DESIGN

711

a complete noncoherent UWB receiver design with four main components: signal detection, timing synchronization, symbol demodulation and channel estimation (c.f. the receiver block diagram in Fig. 1). Based on the WED structure, we derive the optimal decision thresholds and weighting coefcients for the signal detection problem that aims at maximizing the probability of detection, the synchronization problem that aims at maximizing the output SNR, and the symbol demodulation problem that aims at minimizing the bit error rate (BER). Further, efcient parameter estimation methods are developed to acquire the channel-related information needed for computing the weighting coefcients and decision thresholds. Together, a full-scale noncoherent UWB receiver arises, with good tradeoff in performance and implementation complexity. There are some related works on weighted processing for noise suppression. A weighted autocorrelation UWB receiver was developed for data demodulation in transmit reference systems in [12]. Several variants of the WED principle in [11] have appeared for different modulation schemes and/or decision rules, e.g., [13], [14], but they are all conned to the data demodulation module only. In particular, [14] derives the optimal maximum likelihood demodulator based on the WED structure, which minimizes the bit error rate but calls for a nonlinear demodulation rule. In this paper we focus on a simple linear receiver, and we show that this approach has little performance loss when compared to the optimal nonlinear rule. A related line of work is the noncoherent RAKE-based receiver developed in [15]. Therein, multiple RAKE ngers collect the fractional energy of (strong) channel paths in a noncoherent manner, and then maximum ratio combining is performed by weighting different RAKE ngers based on the channel amplitude values. The RAKE ngers are steered toward selected strong channel paths at different path delays; in contrast, the integrators in our WED receiver have xed timewindows that are placed consecutively, which considerably simplify implementation and channel estimation compared with selective RAKE. The noncoherent RAKE can outperform the WED when enough RAKE ngers are employed to collect adequate multipath energy, and vice versa. Finally, a multi-leg structure is developed in [16] which decomposes the UWB signal in the frequency domain using a chain of parallel analog processors, with an integrator in each leg. It provides a general parallel processing approach for implementing transmitted reference UWB receivers, but is more complex than the WED. The rest of the paper is organized as follows. The signal model of UWB impulse radio is described in Section II. The WED schemes for signal detection, synchronization and data demodulation are presented in Sections III, IV and V, respectively. Section VI provides parameter estimation solutions that acquire the channel-related and noise-related information used for WED design. Corroborating simulations are provided in Section VII, followed by concluding remarks in Section VIII. II. S IGNAL M ODEL In impulse radio UWB communication, data are conveyed by a stream of ultra-short pulses. For exposition clarity, this paper focuses on baseband UWB signals with binary on-off keying (OOK) modulation, while extension to other noncoherent modulation such as pulse position modulation (PPM)

Fig. 1.

Block diagram of the weighted energy detector.

is straightforward [13]. In OOK UWB signaling, symbols [] { 0, 1} are transmitted over a train of ultra-short pulses () of pulse width . There are frames of duration per symbol period, with one pulse per frame. Setting , a low-duty-cycle transmission format arises. The transmitted 1 symbol-waveform is given by () = =0 ( ), which has symbol duration = and is assumed to have normalized unit-energy 2 () = 1. With energy per symbol, the transmitted OOK UWB signal is given by () =
=0

[] ( ).

(1)

A pulsed UWB signal typically experiences a frequency selective channel [2], whose impulse response () can be described by a quasi-static tapped delay line as in IEEE 802.15.4a models [17]. That is, () =
=0

( ),

(2)

where + 1 is the total number of resolvable multipath taps, and are the channel gain and delay of the -th tap respectively, and s are real-valued with phases 0 or . Hence, the received signal is () =
=0

[] ( 0 ) + (),

(3)

where () is the channel-inicted received symbol waveform in the form of () = =0 ( ( 0 )), and 0 is the time delay of the rst arriving channel path. The ambient noise () is approximated as a zero-mean white Gaussian process with two-sided power spectral density (PSD) 0 /2 and bandwidth dictated by the ideal lowpass frontend lters cutoff frequency. In this paper, we set the frame duration to satisfy > in order to avoid inter-frame and inter-symbol interferences, where = + is the nonzero support of () determined by the channel delay spread = 0 . III. WED R ECEIVER S TRUCTURE The overall structure of the proposed WED receiver processing is depicted in Fig. 1. In terms of hardware implementation, the receiver is made of branches following a signal squarer, each of which collects a portion of the received signal energy within each symbol period. As shown in Fig. 2, this is done by setting the integration window of the -th branch to + + [( 1) , ) for = 0, . . . , and = / , where is set differently in different

712

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

TM 0

y0

decision statistic:
a0

r (t )

[ ] = =

1 1

( + )

+ 1 1 +(+1) = =0 +

2 () 2 (). (4)

Ts (M 1)TM

yM

aM
Fig. 2. Structure of the weighting procedure.

stages of the receiver processing. Samples from the parallel integrators are weighted and linearly combined to generate the decision statistic for energy detection, where is typically set to be reasonably small for good performance-complexity tradeoff. Depending on the time scale, let us decompose the timing offset as 0 = + 0 , where is an integer and 0 is the residual timing offset within a symbol period, i.e., 0 [0, ). Then, the -th symbol spans the time window 0 + [ , ( + 1) ). The rst stage of receiver processing is to detect the presence of the received signal (), which is jointly done with coarse symbol-level timing acquisition of . During this stage, there is no prior timing information, and = 0 is used in the integration windows of the WED detector to collect digital samples. After () is detected and the symbol-level timing offset is acquired, the receiver enters the ne-scale timing synchronization phase in order to estimate the residual 0 . Here, = + is used in the WED, where [0, ) is the candidate time shift used to search for 0 . Finally, during data demodulation, samples are collected from the branches of the WED for detecting each symbol [], with = + 0 + . These samples are linearly weighted and combined to perform optimal linear demodulation. Throughout the process, the channel estimation module operates in parallel with the above modules in order to acquire channel parameters needed for computing the optimal weighting coefcients and decision threshold. Details of each module are presented in ensuing sections, emphasizing the optimization of weighting coefcients and decision threshold. IV. WED FOR S IGNAL D ETECTION The very rst step in receiver processing is to detect the received UWB signal (). To facilitate a binary hypothesis test for energy detection, a sequence of training symbols containing ones are utilized, from which the detection statistic and decision threshold can be derived. Suppose that the length of an observation window for signal detection is , which contains all-one training symbols if the UWB signal is present. Let denote the candidate value for the symbol-level timing offset , which is integervalued. Given , a conventional square-law energy detector conducts a binary hypothesis test based on the following

Through a line search on , the offset estimate is decided to be the one maximizing [ ]. Meanwhile, the presence of () is declared when [ ] , where is a threshold determined by the desired probability of false alarm [7]. For WED, every received signal frame is partitioned into segments, each of time length = / . A set of integrators are employed to collect energy from these segments. Specically, given a candidate offset , the -th integrator generates its output [ ] as + 1 1 + +(+1) 1 [ ] = 2 (), (5) + +
= =0

= 0, . . . , 1.

The decision statistic in WED is then produced by a linear combination of all { [ ]}1 as follows: =0 [ ] =
1 =0

[ ] = a y[ ]

(6)

where y[ ] = [0 [ ], . . . , 1 [ ]] collects all the integrator outputs and a = [0 , . . . , 1 ] contains weighting coefcients. Now, the coarse timing offset can be estimated as = arg
[0,..., 1]

max

[ ].

(7)

The peak value is denoted as = max [ ] = [ ], which ], = 0, . . . , 1. is a linear combination of = [ A threshold operator follows based on . If exceeds a chosen threshold , the presence of () is declared. The remaining task of WED design is to optimally calculate the weighting coefcients a and the threshold . A. Analysis of Detection Statistics For a binary hypothesis test, calculation of the decision threshold requires knowledge of the probability density function (pdf) of the decision statistic. Conventional energy detectors have been analyzed [7], [8], whose outputs obey the 2 distribution. To facilitate receiver design and analysis, we adopt the Gaussian approximation on the pdfs of the WED outputs. To justify the Gaussian approximation, we note that the ltered noise () at the output of an ideal lowpass lter has a rectangular spectral density over . Hence, the autocorrelation function of () is ( ) = 0 sinc(2 ), and () can be expressed by the sampling theorem as [7], ( ) () = 2 sinc (8) 2 = where are .. Gaussian variables with zero mean and variance 0 . As a result, each integrator output can be viewed as the sum of (2 + 1) i.i.d. random variables. For UWB systems with multiple frames per symbol

WANG et al.: WEIGHTED ENERGY DETECTION FOR NONCOHERENT ULTRA-WIDEBAND RECEIVER DESIGN

713

and a large time-bandwidth product, the central limit theorem applies, and hence Gaussian approximations on the WED outputs are accurate [19]. For analysis, the joint problem of symbol-level timing acquisition and signal detection can be cast as a single binary hypothesis test problem. Under the signal-present hypothesis 1 , is acquired and all the contributing symbols to { } are the all-one training symbols. Under the noise-only hypothesis 0 , { } only collects the noise energy. The estimate aims to separate these two hypotheses as much as possible, while the decision threshold is designed based on the statistics assuming the symbol-level timing offset is known. Dening , () = ( + + ), we summarize the decision statistic as

and (15) as follows: 1 =

2 1,det = 1 =0 1 =0

E{ 1 } = 1 = 0 ( 1) + , (16)
2 2 var{ 1 } = 1 = 1 / 2 1 = 0 ( ) + 20 ( ). 2

where

(17)

B. Optimal Parameters for Signal Detection

Based on the approximate pdfs obtained in Section IV-A, the optimal decision threshold and weighting coefcients for signal detection can be derived to maximize the probability of detection for a xed probability of false alarm . 1) Optimal Weighting Coefcients: Since the pdfs of the = (9) WED output are approximately Gaussian distributed under 1 1 1 (+1) 1 and 0 , given the probability of false alarm , it is 2 ( () + , ()) , 1 ; straightforward to nd the probability of detection as =0 =0 ( ) 1 1 (+1) 1 1 0 0,det 2 , (18) = (1 ; 1 ) = 1 , (), 0 . 1,det
=0 =0

Under 0 , is the average of noise-square terms, which can be well approximated to be Gaussian distributed. Using (8), the mean and variance of under 0 are [7]: 0,
2 0,

= =

E{ 0 } = 0 , var{ 0 } =
2 0 / .

(10) (11)

where = 1 ( ) is a scalar, and () = 2 (1/2) /2 is the well-known complementary error function. To maximize the probability of detection , it is 0 equivalent to maximizing the function det () = 1 1,det 0,det . Noting that 0 , 1 , 0,det and 1,det are functions of , the optimal weighting vector det is obtained by det = arg max det ().

(19)

Let us collect the means and variances into 0 = 2 2 [0,0 , . . . , 0,1 ] and 0 = diag{0,0 , . . . , 0,1 }. 2 2 Dening 0 = 0 , 0 = 0 and 1 = [1, . . . , 1] of length , it holds that 0 = 0 1 and 0 = (0 / )I. Based on (6), (10) and (11), the mean and 2 variance of the decision statistic under 0 can be expressed as: 0
2 0,det

For WED, det () can be deduced from (12)-(17) as 2 (0 / ) , det () = 2 ( + (2 / 2 )) (0 / ) 0 0

(20)

where is the identical matrix. Setting det () = 0 and imposing the constraint 1 = to avoid trivial solutions, the optimal det is found to obey:

= =

E{0 } = 0 = 0 ( 1), var{0 } = 0 = where

2 0 = 0 ( ). 2

(12) (13)

det =

2 0 /

( + )1 . 1 ( + )1

(21)

In (21), is viewed as a non-negative diagonal loading term, which can be optimized by

= arg max det (det ( ))

Under 1 , consists of both noise energy and re ceived symbol energy = 0 2 () that is col lected by integrators, each of which captures a portion (+1) 2 (). Obviously, it holds that , = 1 =0 , = . The mean and variance of under 1 are: 1, 2 1, = = E{ 1 } = 0 + , , var{ 1 } = (0 + 20 , )/ . 2 (14) (15)

(22)

where det (det ( )) =

2 ( + )1 (0 / ) ( + )2 . (0 / ) ( + )2 ( + (20 /0 )) 2 2

Dene 1 and 1 in the same manner as in the 0 case: 1 = 2 2 diag{1,0 , . . . , 1,1 } and 1 = [1,0 , . . . , 1,1 ] . With denitions = [,0 , . . . , ,1 ] and = diag{}, the mean and variance of under 1 can be derived from (14)

A closed-form solution to (22) is difcult to obtain, but a simple line search can be performed numerically to nd the optimal . The objective function det () can also be equivalently expressed as a function of 0 , 1 , 0 and 1 . To show this, (19) can be equivalently written as (1 0 ) 0 det () = (23) 1

714

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

which leads to ( + 1 1 )1 (1 0 ). 0 1 ( + 1 1 )1 0 (24) Computation of det based on (21) or (24) requires the knowledge of channel-related information and noise-related information 0 . Both parameters can be obtained by channel estimation, which will be described in detail in Section VII. 2) Detection Threshold: Having acquired det , the detection threshold det can be found from the desired probability of false alarm . Specically, a false alarm occurs when det under 0 , that is: ( ) det 0 . (25) = (1 ; 0 ) = 0,det det = Solving the inverse problem yields the desired threshold for signal detection, given by det = 0,det 1 ( ) + 0 = 1 ( ) 0 det + 0 . det det (26) V. WED FOR T IMING S YNCHRONIZATION After the signal is detected at the receiver, an accurate estimation of the timing offset is needed in order to reliably demodulate information symbols. In this section, we develop a data-aided timing synchronization scheme using pilot symbols. Utilizing the WED principle, the scheme offers accurate estimation of the symbol epoch offset at low complexity. A. Timing Offset Estimation via WED Recall the signal model in (3), where the timing offset is 0 = + 0 . Having acquired , we can simply assume = 0 without loss of generality, such that 0 = 0 is the residual timing offset to be estimated. Let be a candidate value of the timing offset. The goal here is to recover the residual 0 within a symbol period, that is, = 0 [0, ). For timing synchronization under the general principle of energy detection, we adopt a special pattern of = 2 pilot symbols that consist of pairs of alternating symbols (1, 0). Correspondingly, we employ an 2 + + 1 objective function sync ( ) = 1 2+ 2 (), =0 which collects the average signal energy within a symbollong integration window starting at . As shown in Fig. 3, at an arbitrary = 0 , sync ( ) collects the signal energy from two consecutive segments of the received waveform: one segment is from the pilot symbol 1 with positive energy, and the other is from the neighboring symbol 0 with zero energy contribution. At correct timing = 0 , the -long integration windows [2 + , (2 + 1) + ), , collect signal energy entirely from the symbols of 1s, with no contribution from 0s. Hence, 0 = arg max sync ( ) in the noise-free case1 , which yields the maximum value of sync ( ) max as sync ( ) = sync (0 ) = , with = 0 2 ().
1 More precisely, if the frame length is longer than the delay spread, that max is, = > 0, then the maximum sync is reached for any [0 , 0 ]. Such timing ambiguity is quite small since < , and it does not affect the symbol demodulation accuracy because all the multipath energy per symbol is collected for any [0 , 0 ].

Ts

2Ts

3Ts

Ideal received synchronization symbol waveform (

0 )

Ts

2Ts

Ts

r 2 (t ) dt

Fig. 3.

Schematic illustration of energy detection for timing estimation.

Because a nonzero pilot symbol 1 occurs every 2 in this alternating training pattern, to nd its epoch we only need to search over a 2 range, i.e., [0, 2 ). Applying this synchronization idea under the WED structure, the objective function becomes sync ( ; ) = where
1 1 2 + +(+1) + 2 (), =1 =0 2 + + + (28) and { } are weighting coefcients to be designed. As is evident from (28), each frame is partitioned into consecutive segments of length = / , and integrators are used to collect the average energy ( ) from the corresponding segments. Given , the objective function sync ( ; ) reaches its maximum value when perfect timing offset is acquired, yielding the following timing offset estimator for the WED receiver: (29) 0 = arg max sync ( ; ). 1 =0

( )

(27)

( ) =

[0,2 )

B. Optimal Weighting Coefcients for Synchronization We now derive the optimal weights to be used in (29). max min Let us dene sync () = max sync ( ; ) and sync () = that sync ( ; ) in (27) reaches its maximum value when = 0 and all collected symbols are 1s; similarly, it reaches its minimum when = 0 + and only 0s are collected. Hence, max min sync and sync respectively coincide with the decision statistic under hypotheses 1 and 0 in Section IV-A, except that is replaced by . Under the Gaussian assumption, the pdfs max min () of sync and sync are given by ) ( { max 2 (sync ) = ( 1 , 1,sync) , [] = 1, (30) min 2 (sync ) = 0 , 0,sync , [] = 0, where (, ) denotes the Gaussian pdf, 1 and 0 are the 2 2 same as in (12) and (16), and the variances 1,sync = 1 / =
[0,2 )

min

sync ( ; ). Comparison between (28) and (9) reveals

[0,2 )

WANG et al.: WEIGHTED ENERGY DETECTION FOR NONCOHERENT ULTRA-WIDEBAND RECEIVER DESIGN

715

( 2 ) 2 2 0 (a a) + 20 ( ) / and 0,sync = 0 / = 2 0 (a a)/ are similar to (13) and (17) except that rather than training symbols are used for noise averaging. During the synchronization stage, the goal of choosing max weights is to maximally distinguish sync () at = 0 from other sync ( ; ) values for = 0 , in order to recover 0 max min reliably. This goal boils down to separating sync and sync as much as possible. According to detection theory, to distinguish two Gaussian random variables at a xed false alarm rate, a larger Euclidean distance between the normalized mean values of these variables results in better detection performance. For max min sync and sync , their normalized Euclidean distance is reected in the deection coefcient 2 dened by ( )2 (1 0 )2 2 = . (31) = 2 2 0,sync (0 /) From this expression, the optimal weighting coefcients that maximizes the deection coefcient 2 in (31) shall be proportional to , that is, sync = arg
: 1=1

Utilizing (34), the objective function is equivalent to

2 2 2 = 2 +0 +2 +1 +22 20 (1 +0 )+2 2(1 0 ). 0 1 0 (36) Therefore, the optimal values of and 0 that minimize are given by

= 1 0 which in turn reduce to

and

= 0 0 (37)

2 2 2 = 0 + 1 = (0 + 1 ).

From (37) and imposing the constraint 1 = , the optimal weighting coefcients decode is given by decode = (0 + 1 )1 1. 1 (0 + 1 )1 1

(38)

Subsequently, the MMSE for [] is given by [] [] 0 ( 1) 0 = and the nal decoder is MMSE = [] = sign{MMSE 0.5}. (39)

max {2 } = /( 1).

(32)

(40)

max The above sync maximally separates sync at = 0 from sync at other values of . It is obtained from the channel information regardless of , and is employed in (28) for all possible shifts . VI. WED FOR DATA D EMODULATION

After timing offset 0 has been acquired, the WED enters the data recovery phase. For OOK modulation, the -th integrator generates its output [] at the -th symbol period as 1 (+1) 2 ( + + + 0 ). (33) [] =
=0

Finally, we remark that the optimal weighting vector and decision threshold are derived differently in the three phases of receiver processing. The individual designs reect the different design goals in these phases, which result in different optimality criteria. Nevertheless, the general principle of all three phases is to improve the effective SNR. In this sense, when design complexity is concerned, each of the parameter designs can be adopted for all three phases to expect performance enhancement over conventional energy detection, without much performance degradation from the individually optimal design. VII. C HANNEL PARAMETER E STIMATION In previous sections, design of the WED receiver modules relies on the knowledge of the noise-related information 0 and the channel-related information = [,0 , . . . , ,1 ] that describes the fractional energy of the received symbolwaveform collected by the integrators in the noise-free case. In practice, these quantities are replaced by their estimates acquired during the channel estimation phase. This section provides low-complexity solutions for channel parameter estimation. It is worth emphasizing that the ( + 1) channel parameters {, 0 } are different from the channel state information, which refers to the amplitudes and delays of the ( + 1) channel paths, with for a typical dense multipath UWB channel. Hence, the complexity of the channel estimation task in the WED receiver is dictated by , independent of . In contrast, a selective RAKE receiver requires channel state information [3], [15]. A. Estimation of and 0 In the signal detection phase, an all-one training sequence is transmitted. Consider a signal-present symbol-long window [, + ], for > 0 . Let denote the time residual of within one , that is, = / [0, ) ( denotes integer oor). Similarly, we have dened 0 = 0

[ ] Given [] = 0 [], . . . , 1 [] and assuming equal prior probability for the binary random symbol [], the optimal detector for [] that minimizes the bit error rate is the maximum likelihood (ML) estimator, which turns out to be a nonlinear demodulator [14]. For simplicity, we conne our attention to linear demodulation and decide [] from the linearly weighted decision statistic [] = []. Contrasting [] with in section IV-A, it is straightforward to show that they obey the same distributions and pdfs except replacing by 1. Hence, the conditional pdfs [] () can be deduced as { 2 1 () = (1 , 1 ), [] = 1 (34) 2 0 () = (0 , 0 ), [] = 0
2 2 where 1 , 0 , 1 , 0 are the same as those in (12)-(17) for a single symbol, and are functions of . To avoid the complexity of ML detection, we formulate a simple linear MMSE estimator for decoding [] and optimizing . Suppose that [] is a linear function of [], that is, [] = [] + 0 for some coefcients and 0 . Accordingly, can be solved in the MMSE sense as follows: { { }} decode = arg min = E [] 0 []2 . (35) ,,0

716

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

0 / [0, ). Taking the expectation {} on the noisy signal () in (3) within [, + ] yields (assume 0 without loss of generality) {()} = ( ; 0 ) (41) { (( ) (0 )), 0 = (( ) (0 ) + ), < 0 for [, + ]. Here, (; 0 ) for [0, ] is a wrapped version of () by delaying () by 0 and wrapping the ending 0 -long segment to the begining. It holds that (; 0) = (). In general, , ( ; 0 ) is a wrapped version of ( ) and nevertheless contains all channel paths. Hence, the entire channel energy can be collected from ( ; 0 ) regardless of . The resulting (+1) 2 channel parameters are [] = (; 0 ), = 0, . . . , 1. In practice, the expectation is replaced by sample mean estimate as follows: (; 0 ) =
1 1 ( + + ), =0

the effective SNR for demodulation, or can be re-computed from (38) directly for enhanced performance. B. Performance Enhancement via Decision-Direction and Bootstrapping It is known that the error variance of a parameter estimate is directly affected by the number of samples used for estimation. To accurately estimate channel parameters and 0 using a small number of training symbols, a couple of performanceenhancing strategies are proposed in this subsection. Our rst strategy is to adopt the decision-directed (DD) technique to develop a low-complexity channel tracking algorithm for updating {()} in (41). Channel tracking will take place during demodulation ( = 0 ). The idea is to utilize the information-conveying received waveforms to rene channel estimation, after the impact of information symbols is removed by the symbol estimates. Specically, whenever a symbol [] = 1 is detected during the demodulation phase, its corresponding waveform (), [0 + , 0 +(+1) ] will be used to update the sample mean estimate (; 0) in (42) and noise estimate () in (44), as follows: (; 0) = (; 0) + (1 )( + 0 + ), (46) ] [ () = () + (1 ) ( + 0 + ) (; 0) , (47) for [0, ], where the scalar (0, 1) is a forgetting factor. In DSP implementation, the received waveform is sampled every sam and hence there are = /sam samples per . Let denote the index of digital samples within each symbol, and [; ] denote the -th sample during the -th symbol period. Then, (46) and (47) can be re-written as: [] = [] = [] + (1 )[; ], ) ( [] + (1 ) [; ] [] , for = 0, . . . , 1. Correspondingly, (43) and (45) can be expressed as 1 [] = 2 (50) =(1) []sam . 1 2 [] =0 0 = . (51) Our second strategy is to adopt the bootstrapping technique [21], which is particularly useful when the number of training symbols is small due to the need to reduce communication overhead. During training-based channel estimation, say for signal detection, let us collect all the digital samples received from the pilot symbols into a sample set R = {1 , . . . , }, where = [[1; ], . . . , [ ; ]] corresponds to samples from the -th symbol, . These sample vectors are assumed to be independent and identically distributed (i.i.d.), whose cumulative distribution function is denoted as . Evidently, the channel and noise parameters and 0 have some unknown dependence on . The bootstrapping method aims to nd the distributions of and 0 , the estimators of and 0 , from the sample set R. Because is unknown, an empirical distribution function is constructed from the sample set so that it is close to in some sense [21]. For example, can be constructed (48) (49)

[0, ]. (42)

Accordingly, the channel parameters can be estimated from (42) as (+1) (; 0 ). (43) [] =

This estimator does not need the knowledge of 0 , but just the chosen which is the starting point of the observation time-window for channel estimation. Similarly, the noise () can be ideally extracted from () by () = () {()}. Replacing {()} by its sample mean estimate ( ; 0 ) in (42) for some , the noise parameter 0 can be estimated as follows: (44) () = ( + ) (; 0 ), [0, ]. 2 () 0 = 0 . (45) 2 After the signal is detected, the receiver proceeds to the synchronization phase. Note from Section V-A that the optimal max weighting coefcients are ideally obtained to maximize sync under = 0 . In the absence of 0 , a coarse offset can be used to acquire the channel estimates from (43) and compute the weights in (32) accordingly, and the resulting WED is still expected to offer performance gain over a conventional energy detection. Alternatively, we start with coarse timing offset estimation using equal weights = [1, . . . , 1] . Using a small number of pilot symbol pairs, we can estimate 0 from (27) based on the non-optimal weights. Then, the channel parameters can be updated from (43) by setting = 0 , which in turn yields updated weights and hence improved timing offset estimate. Given enough pilot symbols, this procedure can be iterated for several rounds to produce the optimal estimates of and 0 eventually. In the next data demodulation phase, the channel parameters are the same as those used for timing synchronization, since both phases correspond to (43) for = 0 . The weighting coefcients can remain the same values to improve

WANG et al.: WEIGHTED ENERGY DETECTION FOR NONCOHERENT ULTRA-WIDEBAND RECEIVER DESIGN

717

by assigning equal mass 1/ to each observation , , which is known to approach as under some mild assumptions [21]. By re-sampling from , a new i.i.d. data set R = { , . . . , } can be generated in a bootstrap 1 fashion from the original sample set R, without collecting new samples from the received signal. This procedure can be repeated to keep rening until the cumulative (re-sampled) data set is large enough. At the same time, and 0 can be acquired from each updated , and the accumulated estimates can be used to deduce the probabilistic distributions of and 0 , which in turn produce the improved estimates and 0 . Employing the above bootstrapping principle, the channel estimation steps are summarized below: Step 1) Initialize the estimates and 0 from the original data set R which consists of waveform samples from received pilot symbols. Step 2) Perform bootstrap resampling by uniformly randomly picking elements from the original set R to form a new data set R . Note that any element in R can be picked any number of times, with equal probability 1/ . Step 3) Obtain new estimates and 0 from the new data list R according to (50) and (51), and replace R by R . Step 4) Repeat Step 2) and Step 3) many times (typically 100 to 1000). Step 5) Approximate the distributions of and 0 by the and 0 derived histograms of all the estimates from R , and then compute the mean and variance of each distribution to obtain the nal estimate value and estimation variance of the corresponding channel parameters and 0 . When the decision-directed algorithm and the bootstrap method are employed in the channel estimation phase, only a small number of pilot symbols need to be transmitted to obtain initial coarse channel estimates. The estimates are updated with improved quality when information symbols are fed back reliably and/or when re-sampled bootstrap data are utilized. VIII. S IMULATIONS This section presents simulation results of WED-based receivers for OOK modulated baseband UWB signals. Signaling parameters for the simulations are as follows: the shaping pulse () is the second derivative Gaussian pulse with duration = 1ns. The number of frames per symbol is = 10, and the frame length is = 70ns so that the channel excess delay does not cause inter-frame or inter-symbol interference. The channel is generated from the IEEE CM1 model [20], and is normalized to have unit power gain: E{ 2 } = 1. =0 All performance curves are obtained by averaging over 100 random channel realizations, unless specied otherwise. A. Channel Estimation of and 0 A training sequence with = 50 symbols is transmitted for channel estimation. The performance is evaluated by the mean square error (MSE) criterion and the estimation error is normalized by the actual value. Fig. 4 plots the normalized MSE versus SNR for the estimation of channel parameters

10

normalized MSE of h

10

(43) Bootstrapping DD

10

10

10

4 SNR (dB)

10

Fig. 4. Normalized mean square estimation error of the channel-related information vs. SNR, = 20.

10

(45) Bootstrapping

normalized MSE of N

10

10

10

100

200 300 400 Number of Pilot Symbols

500

Fig. 5. Normalized mean square estimation error of the noise-related information 0 vs. N, SNR = 10dB.

in (43). We compare the estimation performance using three methods: i) simply using (43), ii) using (43) plus the bootstrapping technique, and, iii) with additional decision directed steps implemented. For the decision directed algorithm, another 100 data bits with a BER of 102 are fed back to update the estimation result. The performance gaps among these three methods are evident (about 3 dB). It is clearly shown that the decision directed algorithm improves the estimation accuracy for when SNR is moderately high to ensure reliable feedback. Since SNR values do not affect the estimation accuracy of 0 , we show how the number of training symbols affects the estimation performance of 0 in Fig. 5, for SNR = 10dB. The performance improves as the number of pilot symbols increases but eventually attens out when the number of pilot symbols is above 100. This indicates that 100 training symbols are adequate to provide an accurate estimate of 0 and adding more pilot symbols leads to little improvement in estimation accuracy.

718

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

1 Normalized MSE of Timing offset

0

10

Probability of Detection

0.8

10

0.6

0.4

10

M=1 =T /4
f

M=1 =T M=4 =T

0.2

0 10
Fig. 6.

M=1 M=2 M=4 5 0 5 SNR (dB) 10 15

M=4 =Tf/4
p

10

10

SNR (dB)

15

20

Probability of detection vs. SNR, = 150.

Fig. 8.

Normalized timing estimation error vs. SNR, = 50.

0

1 Normalized MSE of Timing offset

0

10

Probability of Detection

0.8

0.6

10

0.4

M=1 =Tf/4 M=1 =T M=4 =T

p

0.2 M=1 M=4 50 100 Data Size (N) 150 200

M=4 =Tf/4 10
2

0 0
Fig. 7.

50 100 150 Number of Pilot Symbols

200

Probability of detection vs. data size , SNR = 6dB.

Fig. 9.

Normalized timing estimation error vs. N, SNR = 10dB.

B. Signal Detection and Synchronization The effects of the noise variance and the traning sequence length on the signal detection performance are studied in Fig. 6 and 7. In Fig. 6, the probability of false alarm is xed at = 102 where a total number of = 150 pilot symbols are received. WEDs with different numbers of parallel integrators are compared, where = 1 coincides with the conventional energy detector. The gure clearly shows that the WED with > 1 parallel integrators, outperforms the conventional energy detector ( = 1) in terms of the probability of detection , when the SNR is between -5dB and 7dB. Meanwhile, the performance improves as increases, at the expense of increased implementation costs. Fig. 7 depicts vs. for SNR = 6dB. Again, there is evident performance gain of WED with = 4 compared with the conventional ED with = 1. When 100 training symbols are available, the probability of detection reaches almost 1 for = 4 while the conventional one only achieves = 0.5. Figs. 8 and 9 focus on the synchronization phase. Fig. 8 shows the MSE of estimated 0 under different noise levels for = 50. Performances in four different scenarios are

compared: conventional ED ( = 1) where the search step sizes of the candidate shift are = /4 and = respectively, and the WED receiver ( = 4) with = /4 and = , respectively. The pulse-level step size = is used to illustrate the performance of ne-scale search, while the frame-level step size = /4 is adopted to illustrate that our synchronization technique works well for low-complexity coarse time acquisition with exible step sizes. The advantage of the WED with = 4 over = 1 is obvious, and a smaller step size leads to more accurate estimation. In Fig. 9, synchronization performance is delineated for different numbers of pilot symbols, for SNR = 10dB. In all the cases tested, synchronization accuracy improves until the number of training symbols exceeds 50. C. System Error Rates Fig. 10 depicts the BER of OOK modulation for various energy detectors using (40), comparing the case of using estimated channel and timing information with that of using perfect knowledge. As a benchmark, the nonlinear ML

WANG et al.: WEIGHTED ENERGY DETECTION FOR NONCOHERENT ULTRA-WIDEBAND RECEIVER DESIGN

719

10

10

10

10

10

BER

BER

10

10

WEDR = 7ns M = 1 optimal M=1 M = 4 optimal M=4 M = 10 optimal M = 10

5 10 15 20 25

10

10

M = 1 CM1 M = 10 CM1 M = 1 CM4 M = 10 CM4 10 15 SNR (dB) 20 25

10

SNR (dB)

Fig. 10. BER performance vs. SNR for WED with perfect information and estimated information.

Fig. 11. BER performance vs. SNR for WED using different channel models.

weighted energy detector in [14] is also tested, with a subinterval duration = 7ns [14]. This value of is chosen so that the ML detector corresponds to our MMSE-based linear detector with = 10. Clearly, our simple linear MMSEbased WED receiver performs closely to the nonlinear MLbased WED, both of which effectively reduce the noise oor and result in improved BER performance over conventional energy detection. At BER = 102 , the performance gain of = 4 over = 1 is 2.1 dB, while the gain of = 10 over = 4 is 0.7 dB. On the other hand, the hardware cost is determined by the number of integrators, which is ; also, the computational complexity increases quadratically in , due to the matrix-vector operations involved in computing the optimal weights. Overall, the complexity is more than doubled by increasing from 4 to 10, but the performance gain of 0.7 dB is not as signicant as that of the WED with = 4 over the conventional energy detector. This fact indicates that the value of needs to be judiciously chosen to achieve a desired design tradeoff between performance and implementation complexity. When the delay of multiple paths becomes more random, i.e., an indoor environment described in the CM4 channel model, the advantage of WED is even more obvious than the previous CM1 case, since the signal energy is spread into a longer time interval. The result for the CM4 channel is given in Fig. 11. IX. C ONCLUSIONS In this paper, weighted energy detection techniques are developed for the design of a suite of receiver modules for pulsed UWB communication systems. Optimal weighting coefcients and decision thresholds are derived for signal detection, timing synchronization and data demodulation. Overall, weighted processing is effective in alleviating the noise effect in energy detection and thus enhances the output signal-tonoise ratio. It not only offers performance benets to the detection of low-duty-cycle (UWB) waveforms containing noise-only segments with large time-bandwidth product , but also can improve the detection (with unknown timing

offset 0 ) and demodulation (with known 0 ) of ISI-free fullduty-cycle signals. These performance advantages come at the expense of extra complexity compared with conventional energy detection, since multiple integrators and a DSP unit for calculating weighting coefcients are required. Nevertheless, the complexity cost of such a noncoherent receiver is quite affordable, which is only linear in the number of integrators employed. In return, performance gains are evident even for a small value of < 10, as validated by the simulation results. R EFERENCES
[1] J. Foerster, E. Green, V. S. Somayazulu, and D. Leeper, Ultra-wideband technology for short- or medium-range wireless communications, Intel Technology Journal, Q2, 2001. [2] M. Z. Win and R. A. Scholtz, Characterization of ultra-wide bandwidth wireless communications channels: a communication theoretic view, IEEE J. Sel. Areas Commun., vol. 20, no. 12, pp. 1613-1627, Dec. 2002. [3] M. Z. Win and R. A. Scholtz, Energy capture vs. correlator resources in ultra-wide bandwidth indoor wireless communications channels, in Proc. IEEE MILCOM Conf., vol. 3, pp. 1277-1281, 1997. [4] R. T. Hoctor and H. W. Tomlinson, Delay hopped transmitted reference experimental results, in Proc. IEEE Conf. on UWBST, Baltimore, MD, pp. 105-110, May 2002. [5] Y.-L. Chao and R. Scholtz, Optimal and suboptimal receivers for ultrawideband transmitted reference systems, in Proc. IEEE Globecom,, vol. 6, pp. 759-763, Dec. 2003. [6] G. Durisi and S. Benedetto, Peformance of coherent and noncoherent receivers for UWB communications, in Proc. IEEE Int. Conf. Commun., vol. 2, pp. 3429-3433, May 2004. [7] N. Urkowitz, Energy detection of unknowm deterministic signals, Proc. IEEE,, vol. 55, pp. 523-531, Apr. 1967. [8] S. Paguelet and L. M. Aubert, An energy adaptive demodulation for high data rates with impulse radio, in Proc. IEEE Radio and Wireless Conf., pp. 323-326, Sep. 2004. [9] M. Weisenhorn and W. Hirt, Robust noncoherent receiver exploiting UWB channel properties, in Proc. IEEE Conf. on UWBST, pp. 156160, Kyoto, Japan, May 2004. [10] M. E. Sahin, I. Guvenc, and H. Arslan, Optimization of energy detector receivers for UWB systems, in Proc. IEEE Vehic. Technol. Conf., Spring 2005. [11] Z. Tian and B. M. Sadler, Weighted energy detection of ultra-wideband signals, in Proc. IEEE Signal Processing Workshop on Advances in Wireless Communications (SPAWC), pp. 158-162, June 2005. [12] G. Leus and A.-J. van der Veen, A weighted autocorrelation receiver for transmitted reference UWB communication, in Proc. IEEE SPAWC Conf., 2005.

720

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 10, NO. 2, FEBRUARY 2011

[13] J. Wu, H. Xiang, and Z. Tian, Weighted noncoherent receivers for UWB PPM signals, IEEE Commun. Lett., vol. 10, no. 9, pp. 655-657, Sep. 2006. [14] D. Mu and Z. Qiu, Weighted non-coherent energy detection receiver for UWB OOK systems, in Proc. ICSP Conf., 2008. [15] M. Schmidt, D. Simic, and R. Moorfeld, Low complexity low data rate UWB devicesarchitecture and performance comparison, in Proc. IST Mobile Summit, Dresden, June 2005. [16] S. Hoyos and B. M. Sadler, Frequency-domain implementation of the transmitted-reference ultra-wideband receiver, IEEE Trans. Microwave Theory and Techniques, vol. 54, no. 4, part 2, pp. 1745-1753, Apr. 2006. [17] A. F. Molisch et al., IEEE 802.15.4a channel model-anl report, Tech. Rep. Document, IEEE 802.15-04-0662-02-004a, 2005. [18] S. Kay, Fundamentals of Statistical Signal Processing, vol. I: Estimation Theory. Prentice Hall, 1993. [19] J. R. Brown, Error analysis of some normal approximations to the chisquare distribution, J. Academy of Marketing Science, vol. 2, no. 3, pp. 447-454, June 1974. [20] IEEE P802.15 Working Group for WPANs, Channel Modeling SubCommittee Report Final, Nov. 2002 [21] Efron, Bootstrap methods: another look at the Jacknife, Ann. Statist., vol. 7, pp. 1-26, 1979. Feng Wang received the B.S. degree from Southeast University, Nanjing, in 2005 and the M.S. degree from Michigan Technological University, Houghton, in 2007, respectively, all in electrical engineering. He is currently working towards the Ph.D. degree in Electrical Engineering at Arizona State University. His research interests lie in the areas of digital communications, particularly coding and detection techniques, with application to digital storage systems and wireless communication systems.

Zhi Tian (M98, SM06) received the B.E. degree in Electrical Engineering from the University of Science and Technology of China, Hefei, China, in 1994, the M. S. and Ph.D. degrees from George Mason University, Fairfax, VA, in 1998 and 2000. Since August 2000, she has been on the faculty of Michigan Technological University, where she is currently an Associate Professor. Dr. Tians general interests are in the areas of signal processing for wireless communications, estimation and detection theory. Current research focuses on cognitive radio networks and distributed wireless sensor networks. She served as Associate Editor for IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS and IEEE T RANSACTIONS ON S IGNAL P ROCESSING. She received a CAREER award in 2003 from the US National Science Foundation. Brian M. Sadler (Fellow, IEEE; Fellow, Army Research Laboratory) received the B.S. and M.S. degrees from the University of Maryland, College Park, and the PhD degree from the University of Virginia, Charlottesville, all in electrical engineering. He is a senior research engineer at the Army Research Laboratory (ARL) in Adelphi, MD. Dr. Sadler is an associate editor for EURASIP Signal Processing, was an associate editor for the IEEE T RANSACTIONS ON S IGNAL P ROCESSING and IEEE S IGNAL P ROCESSING L ETTERS , and has been a guest editor for several journals including IEEE JSTSP, IEEE J OURNAL ON S ELECTED A REAS IN C OMMUNICATIONS, and the IEEE Signal Processing Magazine. He is a member of the IEEE Signal Processing Society Sensor Array and Multi-channel Technical Committee, and has been active in IEEE conference planning and organization for many years. He received a Best Paper Award (with R. Kozick) from the Signal Processing Society in 2006. He has received several ARL and Army R&D awards, as well as a 2008 Outstanding Invention of the Year Award from the University of Maryland (with J. Baras and P. Yu). His research interests include information science, networked and autonomous systems, acoustics, optics, and mixedsignal integrated circuit architectures.