Cognitive node selection and assignment algorithms for weighted cooperative sensing in radar systems

Physical Communication 2 (2009) 58–72

Contents lists available at ScienceDirect

Physical Communication journal homepage: www.elsevier.com/locate/phycom

Full length article

Cognitive node selection and assignment algorithms for weighted cooperative sensing in radar systems Lingfeng Stephen Wang a,∗ , Angela Doufexi a , Chris Williams b , Joe McGeehan a a

Centre for Communications Research, University of Bristol, Woodland Road, Bristol, BS8 1UB, United Kingdom

b

Centre for Communications Research, University of Bristol, United Kingdom1

article

info

Keywords: Cognitive node teaming Opportunistic radar spectrum access Spectrum sensing

a b s t r a c t Given the favourable transmission characteristics of the radio spectrum allocated for radar and the possibility for its temporal and spatial reuse, opportunistic access is very desirable. This opportunistic access can be implemented in a secondary cognitive communication system, which must have an accurate and reliable sensing capability in order to ensure safe and efficient operation. In this paper, the swept radar’s rotation mechanism is exploited to allow spectrum re-use; cooperative sensing is implemented in the secondary system in order to ensure accuracy. Several node teaming algorithms are proposed for cooperative sensing, along with the use of weighted sensing algorithms. The function of these teaming algorithms is to select the member nodes of the teams and to assign sensing task assignments within the team. Performance results show that selection of appropriate sensing nodes to join the sensing-active team for each sensing cycle and exploring their frequency diversity (to perform the sensing task at the most suitable frequency sub-channels), yields a substantial improvement in performance. In addition, it is shown that proper node teaming algorithms should be chosen based on several key factors, including the characteristics of the primary signal and the sensing team node’s computational capabilities. © 2009 Elsevier B.V. All rights reserved.

1. Introduction In order to improve spectrum efficiency, the concept of Cognitive Radio was proposed in [1], which enables an increase in spectrum utilisation by exploiting white space or bandsharing through a more dynamic, flexible spectrum allocation strategy. An effective cognitive spectrum access algorithm requires efficient and reliable spectrum sensing functionality within the physical layer. The International Telecommunication Union (ITU) has assigned 2300 MHz–2500 MHz and 2700 MHz–3700 MHz as the specific frequency ranges for radar applications in the S-Band for ground-based commercial aeronautical radio navigation, meteorological radio location and military

∗

Corresponding author. Tel.: +44 7796911188. E-mail address: [email protected] (L.S. Wang).

1 Former address. 1874-4907/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.phycom.2009.02.006

usage [2]. These frequency ranges are attractive due to their long range transmission characteristics and non lineof-sight capabilities. However, these frequency allocations were made before spectrum efficiency became such an important issue in spectrum management due to the significant demand for appropriate spectrum resource for wireless communication services. This trend is likely to continue over the next several decades as reported in [3] and has led to considerable interest in exploiting spectrum opportunities for wireless communications. The potential of cognitive radio benefits is straightforward: expensive spectrum resources can be dynamically managed to relieve the resource scarcity for wireless communication applications. However, the potential benefits in this application depend heavily on efficient and effective opportunistic spectrum sensing and access mechanisms. As interference to the radar system must be avoided, reliable spectrum sensing and access algorithms are key for the opening

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

of the radar spectrum to secondary services. Therefore, the development of related cognitive algorithms is important. Previous research work [4,5] has investigated bandsharing between radar and communication systems in the S-band. However, this work did not consider the possibility of cognitive spectrum opportunistic access. Cooperative sensing for bandsharing has been investigated in [6,7]. A recent in-depth analysis is also provided in [8,9] on the design of cooperative spectrum sensing in a radar scenario for improving spectrum usage efficiency. In addition, [10,11] provide references for sensing node selection and deployment issues. For node deployment, node density control is an important issue for saving energy and signalling cost. Ghasemi and Sousa [10] indicate that it is not necessary to use many users in sensing as the sensing sensitivity threshold will asymptotically rise to a limit when increasing the number of cooperative users under a certain distance spread. Furthermore, Peh and Liang [12] show that instead of grouping all the available nodes into teams for sensing work, the optimum sensing performance is achieved by grouping those nodes having the highest SNR (received power on the primary spectrum to noise ratio). In [11], a coverage-preserving self-scheduling protocol is proposed based on a cooperative sensing model. In each sensing round, sensing nodes that have minimum contribution to coverage and also satisfy the coverage preserving rule, will be scheduled to enter hibernation state. This protocol is shown to maintain the original network coverage with fewer on-duty (sensing-active) nodes. Consequently, higher energy efficiency and lower signalling demand are achieved. This paper will focus on addressing the issue of team node selection and sensing task assignment to mobile candidate nodes in a swept radar scenario. Previous work [9] has proposed a weighted cooperative sensing algorithm to improve the sensing performance, in which it is assumed that the secondary cognitive sensing nodes are in a fixed scenario, in other words, these nodes have time-invariant SNRs. In this paper, this work is extended to a scenario where mobile nodes explore their opportunistic access in a radar system with a rotating antenna. Consequently, different sensing nodes have different SNRs at different times and possibly in different frequency sub-channels. Therefore, it is worthwhile to design a node teaming algorithm to select effectively the appropriate nodes to join the cooperative sensing team and to assign optimally the team nodes to perform sensing for different times and frequency sub-channels (if in a frequency selective channel) to keep achieving the target sensing performance in a mobile scenario. The rest of this paper is organised as follows. In Section 2, the sensing scenario is presented and the features of both the primary system and secondary cognitive system are described. In Section 3, a cooperative sensing team node selection algorithm and several team node assignment algorithms are proposed, along with the weighted sensing algorithms by considering nodes’ sensing credibility. Section 4 discusses the performance results of the proposed algorithms and Section 5 concludes the paper.

59

2. Scenario description 2.1. S-band swept pulsed radar The main parameters of the pulsed radars considered in this paper, based on several types of commercial (aeronautical radio-navigation and meteorological) and military radars [13], are shown in Table 1. The pulse waveform is of interest due to its wide deployment in S-band radar systems. Its waveform is given by, spw (t ) =

∞ X

sspw (t − nT )

(1)

n=−∞

where T is the radar Pulse Repetition Time (PRT), sspw (t ) is the single pulse waveform and is given as, sspw (t ) = rect

(p

  t

τ

Pt

= 0

−

τ

≤t≤

τ

2 2 otherwise

(2)

rect τt is defined as a rectangular pulse of pulsewidth τ . In a pulsed radar, average transmitted power Pav g =




Pt × Tτ is expected to be high for a good radar receiver SNR. Therefore the pulsewidth τ should be longer for a limited peak transmitted power Pt and PRT T . However, for an unmodulated pulsed radar, the radar range resolution c ∆R = 2B (c = 3 × 108 m/s is the light speed) and the time-




bandwidth product ξ= Bτ is unity (radar bandwidth B = τ1 ). Thus, a trade-off exists between a fine range resolution (which needs a high bandwidth) and a good average transmitted power (which requires a long pulsewidth). This issue can be solved by introducing a pulse compression technique and corresponding modulation, by which the range resolution is substantially improved by increasing the bandwidth while keeping the pulsewidth unchanged (therefore the average transmitted power and corresponding detection range is maintained). Pulse compression is commonly used in modern radar systems, and it can be realised by modulating the radar pulse, therefore a Linear Frequency Modulation (LFM) radar waveform is adopted as a wideband radar signal in this paper. A LFM waveform is expressed by, sLFM (t ) = rect

  t

τ

u 2 ej(2π f0 t + 2 t )

(3)

where f0 is the radar’s operating frequency, u = 2πτ B denotes as the LFM coefficient. Thus the time-bandwidth product ξ is non-unity and only depends on the modulation frequency spread (bandwidth) B = f2 − f1 . As the pulse modulation does not affect the radar pulsewidth while widening the radar bandwidth, it effectively solves the trade-off between range resolution and receiver SNR. From spectrum sensing perspective, a wider radar bandwidth might lead to more frequency diversity. As most radar systems have a cosecant-squared elevation pattern (e.g. aeronautical radio-navigation radar), which radiates almost the same power along different elevation angles, exploring the opportunistic radar spectrum access along the elevation dimension is of no interest.

60

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Table 1 Swept radar parameters. Parameters

Unit

Radar operating frequency Pulse Repetition Time (PRT) Pulsewidth 3 dB mainlobe width Pulse moduation

2.7 1000 10 1.4 (Cosine aperture distribution) Radar 1: Unmodulated; Radar 2: Linear Frequency Modulation (LFM) 4.8

Radar rotation period

Consequently a one-dimensional (horizontal) swept radar aperture distribution is considered here, the antenna pattern is given by [2], E (φ) =

Z

a 2

− 2a



A(z ) exp j2π

z

λ



sin φ dz

πz a

|z | <

E (φ) =

4

"

a

(5)

2

  sin ψ + π2 ψ+

π

2

+

  sin ψ − π2 ψ−

π

of the characteristics of the radar and, by enhancing the sensing performance, increase spectrum efficiency through reliable spatial opportunistic access to the radar spectrum. 2.2. Cognitive secondary communication system

Thus, its corresponding field-intensity pattern is expressed by [2],

π

second

(4)

where a is the width of the aperture in z-dimension. λ is the radar wavelength. φ is the radar swept rotation angle and A(z ) is the radar aperture distribution. A Cosine aperture distribution is chosen as a typical parameter and expressed by [2], A(z ) = cos

GHz µs µs degree

# (6)

2

where ψ = π ( λa ) sin φ . Unlike a generic communication system’s transmitter that transmits its primary signal omnidirectionally at all times, the swept radar transceiver rotates its antenna 360 degrees horizontally in a preset rotation period. e.g. 4.8 s for air traffic control radars at airports; 20 s for meteorological radars [13]. In addition, the transmission power of the first sidelobe of the radar is usually over 23 dB below the mainlobe. This rotation mechanism makes the radar protection range/area (in which, secondary devices are forbidden to access the radar spectrum or only permitted to have very low transmission power due to the radar’s protection criteria) not an omnidirectional circle but a mainlobe-led area which keeps rotating based on the radar rotation period. Therefore the radar spectrum spatial opportunistic access/sharing is worth investigating for spectrum efficiency purposes. From a spectrum sensing perspective, obtaining a higher primary radar signal power by tracking the radar mainlobe can effectively increase the detection probability while reducing the false alarm probability in a limited sensing cycle time. From an opportunistic radar spectrum access viewpoint, avoiding accessing the radar spectrum when being swept by the mainlobe can dramatically reduce the radar protection range and/or increase the transmission power of the secondary cognitive device. In the following sections, several sensing teaming algorithms in terms of a node selection algorithm and node assignment algorithms, along with the weighted cooperative sensing algorithms will be proposed. These algorithms will make full use

A generic communication system with a spectrum sensing function is adopted as a cognitive secondary system in the radar spectrum, e.g. 3GPP Long Term Evolution (LTE) system with sensing capability or a sensing-enabled WiMAX system. Such an OFDM-based system is suitable for opportunistic access as it allows a very flexible spectrum access on a sub-carrier-by-subcarrier basis. The secondary system could operate on its licensed spectrum as a primary system while sensing the radar spectrum for opportunistic access as a secondary cognitive system. All the sensing information signalling is transmitted on its primary licensed frequency and will not interfere with the radar operation. In this paper, the communication terminals (the Mobile Stations) are called sensing nodes with a radar spectrum sensing function. Under the assumption of perfect noise estimation, energy detection is used in the sensing node as the local sensing processing technique. This assumption can be relaxed if some information about the structure of the primary signal, e.g. cyclic frequency (built-in periodicities features), is available. Better sensing performance may be achieved by using a more sophisticated sensing technique such as cyclic feature detection [14] at the cost of increased complexity of the sensing nodes. The cognitive system in a swept radar scenario is presented in Fig. 1. Cognitive sensing nodes are mobile with speeds of 3 km/h and they experience a Rayleigh multipath fading ITU pedestrian B channel [15]. All the nodes in the coverage of a central controller (e.g. base station) can act as candidate sensing nodes. The central controller has some basic knowledge of the radar (primary system), e.g. opportunistic access-permitted spectrum ranges (e.g. S-band), but it has no knowledge of detailed radar operation information. As shown in Fig. 1, those nodes being selected as part of the sensing team will be classified as either in a:

• Sensing-active state (nodes performing sensing tasks) • Sensing-hibernated state (nodes not performing sensing tasks) A sensing and opportunistic access process flow is given in Fig. 2. Each sensing-active team node will perform the sensing by sampling the observations in its assigned

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

61

sweeping on a sensing node (therefore a much higher received radar signal power), and is calculated as 0.0186 s (which is equivalent to four sensing cycles) based on our simulation parameters. It is assumed that for each team node, its received SNR frequency response is unchanged during one mainlobe sweeping period. In order to make sure the sensing results are timely, i.e. unexpired, the sensing process should be performed periodically. The term ‘sensing period’ is introduced in [16], and defined as the time interval between two continuing sensing processes. The sensing period determines the maximum time during which the secondary cognitive device will be unaware of the changes of the radar protection range and hence may harmfully interfere with it. In our scenario, shown in Fig. 2, the sensing period is set as the updated period of the sensing-active team. It mainly depends on the primary system’s characteristic, e.g. radar rotation period, and has to be set for the safe operation of radar systems by the regulator. In our simulator, the sensing period is set as one mainlobe sweeping period. Fig. 1. Mobile cognitive communication nodes’ spectrum sensing in a swept radar scenario.

radar spectrum. One sampling duration is regarded as one sensing cycle. Only one team is ‘sensing-active’ in one sensing cycle. Sensing-active team nodes’ local decisions are made at the end of each sensing cycle. Considering the radar’s parameters used in our MATLAB simulator (shown in Table 1), one sensing cycle is set as the duration of four cumulated radar pulses. The term mainlobe sweeping period is defined as the time duration of the mainlobe

3. Cooperative sensing team node selection, assignment and weighted sensing algorithms When employing cooperative sensing, relatively heavy signalling traffic in the transmission of sensing information might occupy a large proportion in the Media Access Control (MAC) frame and leave less room for data transmission. It may also increase the operational costs for network operators. However, employing single-node sensing might require much more time (a longer sensing

Fig. 2. Sensing and opportunistic access process flow.

62

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Fig. 3. Sensing-active team node request and selection process.

cycle) to carry out the required sensing work, compared to cooperative sensing, and this might not be possible for real time opportunistic access data and relatively simple signalling transmission. Therefore, there is a trade-off between cooperative sensing gain, cooperation signalling cost and single node sensing capabilities (its complexity). By exploring the time/spatial diversity in a swept radar scenario as shown in Fig. 1, a sensing team selection algorithm is presented in Section 3.1. Several sensing assignment algorithms for team nodes are proposed in Section 3.2. In Section 3.3, a brief description of the weighted sensing algorithms will be provided to further increase the team sensing performance. 3.1. Selection algorithm of the sensing team A selection procedure of the sensing-active team and its team nodes is as follows: (1) A central controller will select a number of distributed nodes as sensing team coordinators, depending on the state of the nodes (e.g. battery state, signalling channel state, complexity capability) and their positioning information (available by positioning system, such as an in-built GPS gadget). Team coordinators will take responsibility for selecting the team nodes; assigning team nodes to perform sensing at different frequency sub-channels for enhancing the sensing performance; collecting their local decisions and making the global decision on whether the radar spectrum is accessible. (2) The selected team coordinators will transmit a local low-power teaming-request signal to their neighbouring nodes as shown in Fig. 3. Only the nodes which are capable of receiving this request are eligible to join the sensing team of the team coordinator. This low power teamingrequest mechanism can guarantee the effective distance spread among the team nodes, thereby leading to a highly effective and efficient sensing process. (3) All team coordinators send their preliminary SNRs to the central controller, which then decides which team is to switch to the sensing-active state with the remaining teams staying in the sensing–hibernated state. These preliminary SNRs are measured in a very limited time duration, which is insufficient for making a good global sensing decision. However, these measurements are helpful in providing a reference to determine the teams’

sensing state (sensing-active or sensing-hibernated) for initialising the following team sensing process. (4) A sensing-active team starts to perform cooperative sensing to achieve the maximum sensing team SNR (average SNR of the team nodes) and required sensing performance. When the global sensing decision is made at the end of each sensing cycle shown in Fig. 2, the sensing-active team coordinator shares the updated radar spectrum state with the sensing-hibernated team nodes through their coordinators, whose teams perform the sensing tasks in the preceding/following sensing cycles. (5) At the end of the each sensing cycle, the sensingactive team coordinator informs its neighbouring team coordinator to switch into sensing-active state for the next sensing process. Considering the radar antenna is clockwise-rotating, it is reasonable to assume that in the next sensing cycle the neighbouring team has a better team SNR and therefore switch into sensing-active state. However, when the constraint of the signalling traffic is released, this assumption can be relaxed by requesting all team coordinators to make a limited-time sensing at the end of each sensing period and send their preliminary SNRs to the central controller, which then determines which team turns into sensing-active state for the next sensing cycle. As shown in Fig. 2, from the perspective of the secondary system, it is desired to maintain the sensing cycle time well below the sensing period in order to maximise the time available for opportunistic radar spectrum access transmission. Wang et al. [9] have shown that for the same sensing requirement, compared to sensing with a single node, a ‘team cooperative’ sensing approach can offer the same performance in a much short sensing cycle time. In addition, the sensing cycle time can be prolonged for a better sensing performance or reduce the sensing team nodes complexity. However, the opportunistic access time is shrunk accordingly in the case of a fixed sensing period as illustrated in sensing period X -1 (Fig. 2). As shown in Fig. 4, while team nodes are being swept by the radar mainlobe, it is assumed that they have the same average SNR, e.g. 0 dB. The actual received SNR is subject to a Rayleigh distribution with a Doppler shift due to the mobility of the nodes and the fading channel experienced. However, because of the spatial opportunities, arising from the rotation of the radar mainlobe, each candidate node has a substantially different SNR at different times during one radar rotation period. To achieve the best sensing performance at all times, each sensing-active team coordinator will select the highest-n SNR nodes (n = 6 in our simulator) from the teaming-request responsive candidate nodes to join the sensing team in each sensing cycle. In the case where the number of responsive candidate nodes increases dramatically, the preliminary state reporting mechanism (by which candidate nodes transmit their states, such as preliminary SNRs, to the team coordinator) will create heavy signalling traffic. This issue may be addressed by inserting a credibility threshold into the teaming-request signal for the candidate nodes. Only those nodes achieving a higher credibility (e.g. high preliminary SNR, low movement speed) than the threshold,

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

63

Fig. 4. The selection of the team nodes from candidate nodes in each sensing cycle.

may be allowed to respond to the team coordinator. As illustrated in Fig. 3, the team coordinator may specify a tunable threshold through the broadcast channel in order to satisfy the different priorities of the sensing requirements, such as a strict signalling traffic limitation or a guarantee on a certain amount of sensing candidate nodes supply. Furthermore, instead of creating new MAC frames to exchange the teaming request-respond signalling messages, signalling in the physical layer can be employed as suggested by Weiss et al. [17]. 3.2. Sensing assignment algorithms for the sensing-active team The selected team nodes can further explore the frequency diversity of the received SNR to make local sensing decisions on the most suitable frequency subchannels. By considering different criteria, in terms of whether to maximise the sensing performance or to consider the node’s computational complexity, four assignment algorithms are proposed in this section to achieve a better sensing team SNR for a required cooperative sensing performance. In the case of a frequency-flat scenario, where the radar signal is considered as a narrowband signal (e.g. unmodulated radar signal), the highest-SNR can be obtained in the centre of the radar spectrum. In the case of a frequencyselective scenario, e.g. a multi-path channel and a wideband LFM radar signal, the SNRs of the sensing nodes vary both in the time domain and the frequency domain. However, it should be noted that the received SNR is assumed to be time-invariant during one mainlobe sweeping period. Fig. 5 shows the frequency diversity of the received SNR (noise power is normalised to unity). Since the LFM radar signal bandwidth is much wider than the channel coherence bandwidth, the highest SNR possible is located at a frequency sub-channel randomly within the radar bandwidth. The team node with a wider sensing bandwidth can benefit from increased frequency diversity, which consequently leads to a higher SNR opportunity.

Fig. 5. Frequency response of the LFM radar (bandwidth = 10 MHz) after transmitted through the ITU Pedestrian B channel (node average received SNR = 0 dB).

• MaxSense algorithm In this algorithm, each team node selects the frequency sub-channel that has the highest SNR along its frequency domain as expressed by, SCHi = arg max SNRi,j

∀i ∈ I

(7)

j∈J

where I = {1, 2, . . . , n} for indices of nodes, J = {1, 2, . . . , m + p} for indices of sub-channels (n ≤ m + p), SNRi,j denotes node i’s SNR value of sub-channel j, and SCHi is the index of the selected sensing sub-channel for node i. Thus its local sensing decision is made at the chosen frequency sub-channel, then the local decision is sent to the team coordinator which will be making the final sensing decision. This algorithm can offer the highest average team SNR at the end of each sensing cycle. However, if only energy detection is used and a certain frequency sub-channel suffers interference from a nonprimary signal (e.g. malicious signal as described in [18]), the team nodes may incorrectly choose this frequency sub-channel to do the sensing. Consequently, this action will result in an unacceptable false alarm probability repeatedly. We term this issue as the double-false alarm problem. Therefore, it is desirable to make the team nodes make their local sensing decisions at different frequency sub-channels whilst still keeping a high average team SNR for achieving a good sensing performance.

• EasySense algorithm This algorithm provides an alternative to configuring the team nodes’ sensing frequency sub-channels which makes each team node make its local sensing decision at different sub-channels while still maintaining a good sensing performance. The SNR matrix in the frequencyselective case for a certain sensing cycle k is shown in Table 2. To significantly reduce the signalling cost, instead of transmitting the SNRs of all frequency sub-channels from

64

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Table 2 The SNR matrix in the frequency-selective case for one sensing cycle. Sensing cycle k

Frequency sub-channel 1

Frequency sub-channel 2

...

Frequency sub-channel j

...

Frequency sub-channel m + p

Node 1 Node 2 ... Node i ... Node n

SNR1,1 SNR2,1 ... SNRi,1 ... SNRn,1

SNR1,2 SNR2,2 ... SNRi,2 ... SNRn,2

... ... ... ... ... ...

SNR1,j SNR2,j ... SNRi,j ... SNRn,j

... ... ... ... ... ...

SNR1,m+p SNR2,m+p ... SNRi,m+p ... SNRn,m+p

all n team nodes, each team node chooses only to report the sub-channels of the highest-m SNR (m > n) to the team coordinator. As the m sub-channels reported by team nodes are less likely to be fully coincidental, there exist (m + p) columns in the SNR matrix, and p ‘null’s are filled in the corresponding matrix elements. The team coordinator then executes the EasySense assignment algorithm as described below, EasySense algorithm (Performed by sensing team coordinator) (1) Collect and build up the team nodes SNR matrix. (n Rows: team nodes, (m + p) Columns: frequency subchannels, m > n). Thereby, the elements of this matrix are SNRi,j i ∈ I, j ∈ J. (2) For each frequency sub-channel j, select the index of the node which has the maximum SNR on this subchannel, SNDj = arg max SNRi,j

∀j ∈ J

(8)

i∈I

where SNDj is the index of the node with the highest SNR on sub-channel j. (3) For each team node i, set up a frequency candidate pool 5i by collecting the sub-channels to which this node is indexed (in step 2),

5i = {j ∈ J : SNDj = i} ∀i ∈ I.

(9)

(4) For each node, select the sub-channel of the highest SNR in its candidate pool as the sensing sub-channel for the node (If no candidate exists, skip to Idle-node 2nd round assignment process) SCHi = arg max SNRi,j j∈5i

5i 6= ∅,

∀ i ∈ I.

(10)

Idle-node 2nd round assignment process: For the sensing-idle node after the above assignment process, choose the sub-channel of the highest-SNR in the available frequency pool, based on a non frequency sub-channel overlapped principle. After receiving the assigned sensing sub-channels from the team coordinator, team nodes make local decisions and send them back to the coordinator, which then makes the final sensing decision. This algorithm allows the team nodes to perform the sensing task at different frequency sub-channels, however it cannot offer the optimum team SNR due to its frequencyassignment procedure. e.g. node 1 cannot perform the sensing task at its highest available SNR frequency subchannel j, if sub-channel j has been already selected into node 2’s frequency candidate pool but not being finally selected as node 2’s sensing sub-channel.

• EasySense2 algorithm The EasySense2 algorithm is proposed as a modified version of the EasySense algorithm. The EasySense2 algorithm keeps the main structure of the EasySense algorithm while introducing both 1st and 2nd frequency candidate pools (51i , 52i ) for each team node. The 1st frequency candidate pool performs the same function as the EasySense algorithm’s frequency candidate pool, which is rewritten (9) as,

51i = {j ∈ J : SND1j = i} ∀i ∈ I

(11)

where SND1j is the index of the node with the 1st highest SNR on sub-channel j and is expressed as, SND1j = arg max SNRi,j i∈I

∀j ∈ J.

(12)

Furthermore, for each node, it sets up a 2nd frequency candidate pool by collecting the frequency sub-channels which the node having the 2nd highest SNR is indexed to,

52i = {j ∈ J : SND2j = i} ∀i ∈ I

(13)

with SND2j is the index of the node of the 2nd highest SNR on sub-channel j and is given by, SND2j = arg max SNRi,j i∈I\SND1j

∀j ∈ J.

(14)

The EasySense2 algorithm effectively addresses the issue mentioned in the EasySense algorithm by collecting the 2nd highest-SNR node index at each frequency subchannel, which provides all frequency sub-channels with a second chance to be the sensing sub-channel of the team nodes, and consequently improves the sensing performance. EasySense2 algorithm (Performed by sensing team coordinator) (1) Collect and build up the team nodes SNR matrix. (n Rows: team nodes, (m + p) Columns: frequency subchannels, m > n). (2) For each sub-channel j ∀j ∈ J, select the indices of the nodes (SND1j , SND2j ) which have the 1st and 2nd highest SNR on this sub-channel. (3) For each team node i ∀i ∈ I, set up the 1st and 2nd frequency candidate pools (51i , 52i ) by collecting the sub-channels to which this node is indexed (in step 2). (4) For each node i, select the sub-channel of the highest SNR in its 1st and 2nd candidate pools respectively as

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

the 1st and 2nd sub-channels (SCH1i , SCH2i ) of the node (If no candidate, set as null). SCH1i = arg max SNRi,j

∀i ∈ I

(15)

SCH2i = arg max SNRi,j

∀i ∈ I

(16)

j∈51i

j∈52i

Now, define I1 , {i ∈ I : SNR(i, SCH1i ) ≥ SNR(i, SCH2i )} I , I \ I = {i ∈ I : SNR(i, 2

1

SCH1i

) < SNR(i,

(17) SCH2i

)} (18)

where SNR(.) is node i’s SNR value of 1st/2nd subchannel If i ∈ I1 If SCH1i 6= ∅ (1st sub-channel is non-null) Do Select 1st sub-channel as the sensing sub-channel for node i, SCHi = SCH1i

∀ i ∈ I1

(19)

Build the sense-idle frequency pool Ψ = J and update it by removing SCHi from Ψ . Else Skip to Idle-node 2nd round assignment process End Else Do Team coordinator sorts nodes in I2 according to values of SNR(i, SCH2i ). Node with the highest SNR of 2nd subchannel has the priority to select its sensing sub-channel in the sense-idle frequency pool If SCH2i ∈ Ψ (2nd sub-channel of the node i has not been assigned to other node) Do Select 2nd sub-channel as the sensing sub-channel SCHi = SCH2i

∀ i ∈ I2

(20)

Elseif 6= ∅ ∧ ∈ Ψ (1st sub-channel is nonzero and has not been assigned yet) Do Select 1st sub-channel as the node’s sensing subchannel SCH1i

SCHi = SCH1i

SCH1i

∀ i ∈ I2

(21)

Else Skip to Idle-node 2nd round assignment process End End Idle-node 2nd round assignment process: For the sensing-idle node after the above assignment process, choose the sub-channel of the highest-SNR in the available frequency pool, based on a non frequency sub-channel overlapped principle.

• Hungarian Algorithm (HA) algorithm Hungarian Algorithm [19] is a combinatorial optimisation algorithm which solves assignment problems. This algorithm models an assignment problem as a n × (m + p) cost matrix and performs minimisation on the elements of the matrix. The element of the matrix, Ci,j ∀i ∈ I ∀j ∈ J denotes the cost of assigning the ith worker to the jth job. By introducing the Hungarian Algorithm into the node sensing assignment algorithm, the node assignment procedure in the sensing team converts to a maximised overall combinatorial SNR problem, where the ith sensing node

65

is assigned to perform the sensing task on the jth subchannel at a maximal team SNR basis. As shown in Table 2, to be able to use the Hungarian Algorithm, this SNR matrix needs to be modified as a cost matrix by subtracting all the elements from the maximum value of the whole matrix (‘null’ is ignored as they are not real-measured value). As a result, the minimisation of its elements of this modified matrix leads to maximising the original combinatorial SNR values, which then results in a good sensing performance. The algorithm works as follows. (1) Build up the cost matrix C by subtracting all the elements of the SNR matrix from the maximum value of the SNR matrix as, Ci,j = SNRmax − SNRi,j

∀i ∈ I ∀j ∈ J = max max SNRi,j ∀i ∈ I ∀j ∈ J.

where SNRmax

i∈I

(22)

j∈J

(2) For each sensing node i, calculate the minimal cost mi associated with this node as, mi = min Ci,j

∀i ∈ I

j∈J

(23)

and then subtract the minimal cost from the costs associated with node i as given by, Ci,j = Ci,j − mi

∀j ∈ J ∀i ∈ I.

(24)

(3) Likewise, for each frequency sub-channel, calculate the minimal cost ej associated with the sub-channel as, ej = min Ci,j i∈I

∀j ∈ J

(25)

and then subtract the minimal cost from the costs associated with the sub-channel j as shown by, Ci,j = Ci,j − ej

∀i ∈ I ∀j ∈ J.

(26)

(4) Use as few lines as possible to cover all the zeros in the cost matrix. Let Γ present the minimal number of lines needed to cover all zeros elements. If Γ < (m + p), skip to step 5. Otherwise skip to step 6. (5) For Γ < (m + p), the elements of the cost matrix needs to be partitioned into three sets: those elements covered by one line only are categorised in set L; those covered by two lines are categorised in set X ; and those are uncovered are categorised in set Q . And then calculate the minimum element α in set Q ,

α = min Q . (27) Subtract α from all elements in Q and add α to all elements in X as expressed by,

 Ci,j =

Ci,j − α Ci,j + α

if Ci,j ∈ Q if Ci,j ∈ X .

(28)

Then return to step 4. (6) Starting with the top row in the cost matrix, an assignment now can be made when there is exactly one zero in a row. Remove the corresponding row and column once an assignment is made. If any of n nodes cannot be assigned after working through all the rows (This is possible in the case that the rows have more than one zero), switch to make the assignment column by column. Iterate between row assignments and column assignments until as many unique assignments are made as possible. Arbitrary assignment might be needed if there is still more than one zero in rows or/and columns after going through all the rows and columns.

66

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

3.3. Weighted cooperative sensing algorithm It was shown in [8] that cooperative sensing can relax the requirement on sensing sensitivity thresholds due to its mitigation of multipath and shadowing effects. Nevertheless, when the Sensing Information Signalling (SIS) channel is constrained by limited bandwidth or is not transmission error-free, the channel quality will directly influence the global sensing performance. Cheng and Varshney show in [20] that it is helpful to relax the requirements on channel quality when increasing the number of sensing nodes or the decision quantisation levels in the nodes. Furthermore, to reduce the signalling cost of a constrained SIS channel, instead of transmitting large amounts of the original sensing observation data of the team nodes to the team coordinator, Ghasemi suggests in [7] that nodes can only share final 1-bit decisions in sensing information signalling. However the optimal sensing performance is not obtained since the team coordinator has no knowledge of the sensing credibility information associated with those decisions (local probability of detection, local probability of false alarm). In this section, the credibility of local sensing information is taken into account. Two weighted cooperative sensing algorithms, analytical and SNRdirect weighted algorithms, are discussed to further improve the sensing performance in terms of the global detection probability Pd_G and the global false alarm probability Pfa_G . For simplicity, in this section, we consider the case that the sensing-active team only samples the sensing data in the time domain. The performance of the teaming algorithms (exploiting frequency diversity) in combination with the weighted sensing algorithms will be presented in Section 4. In a cognitive radio environment, the observations sensed by the secondary nodes have two hypotheses. Hypothesis H1 , indicates that the primary system is active and Hypothesis H0 , shows that the state of the spectrum of the primary system is vacant. H1 : y(n) = s(n) + u(n)

(29)

H0 : y(n) = u(n)

(30)

where s(n) is the primary system’s signal and here is assumed to be a pulsed radar signal with normalised power level σs2 . The noise u(n), is assumed to be a complex Gaussian Independent and Identically Distributed (IID) random signal with zero mean and variance 2σn2 . s(n) and u(n) are assumed to be independent. The detection probability Pd and the false alarm probability Pfa are two key measurement criteria in the sensing performance, as given by, Pfa = P (T (y) > Vthreshold |H0 )

(31)

Pd = P (T (y) > Vthreshold |H1 )

(32)

where T (y) is the test statistic for a sensing node, and Vthreshold is the detection threshold. The primary system is claimed to be detected when T (y) > Vthreshold in the case of H1 . However it can be a false alarm if T (y) > Vthreshold in H0 . (31) and (32) show that a lower detection threshold leads to a higher detection probability if H1 is true

whereas a higher false alarm occurs if H0 is the case. Therefore, there always exists a trade-off between Pd and Pfa for a certain detection threshold in a team sensing node. In a cognitive environment, when the primary system is detected, secondary nodes will reconfigure their transmission behaviour (either suspend their transmission or decrease the transmission power) to avoid interfering with the primary system. Therefore, a higher detection probability can lead to a higher protection level to primary system while a lower false alarm probability offers better opportunistic access to secondary cognitive nodes. Consequently, a sensing algorithm with higher Pd and lower Pfa is desired. A better sensing performance can be achieved by grouping nodes in a team for a cooperative sensing. The sensing performance can be further improved by introducing the sensing credibility information to each team node. In addition, by introducing two sensing-priority criteria as described below, the sensing performance can be flexibly enhanced according to a specific sensing requirement, e.g. an aggressive cognitive rule (low Pfa_G -priority) which allows more opportunistic cognitive access, or a conservative rule (high Pd_G -priority) giving a higher protection level to the primary system. We first describe the sensing performance of a local team sensing node algorithm for a pulsed radar system in two sensing-priority criteria. The first criterion, Local Constant False Alarm Rate (LCFAR), is introduced from the perspective of maximising the opportunistic access for secondary cognitive nodes. The sensing-active team nodes try to achieve a target global detection probability, constrained with an assigned local constant false alarm probability (Pfa_i ) for each team node. For the second criterion, Local Constant Detection Rate (LCDR), which offers a better protection level to the primary system (radar), each team nodes are assigned a local constant detection probability (Pd_i ). Thus a global detection probability is guaranteed when team attempts to achieve a target global false alarm probability. The test statistic for an energy detector is given by [21], T (y) =

N 1 X

N n=1

|y(n)|2

(33)

where N is the number of sampled observations received by the sensing node. Under the H0 case, for large N, and using the central limit theorem, the PDF of T (y) is approximated by a Gaussian distribution with mean µ0 = 2σn2 , and variance σ02 = probability is given by,

4σn4 . Therefore the local false alarm N

Pfa_i = P (T (y) > Vthreshold |H0 ) = Q

 =Q

Vthreshold 2σn2

 −1 ×

√ N



Vthreshold − µ0



σ0 

.

(34)

For a given local constant false alarm probability Pfa_i (LCFAR criterion), the detection threshold is derived from (34). Vthreshold = 2σn2 ×



 1 √ Q −1 (Pfa_i ) + 1 . N

(35)

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

67

Under the H1 case, the local detection probability related with decision threshold Vthreshold is also approximated by a Gaussian distribution with mean µ1 = 2σn2 + σs2 and variance σ12 =

(2σn2 +σs2 )2 N

, and given by,

Pd_i = P (T (y) > Vthreshold |H1 )

 =Q

Vthreshold 2σn2 + σs2

  √ −1 × N .

(36)

For a given local constant detection probability Pd_i (LCDR criterion), the decision threshold is derived from (36) as, Vthreshold = (2σn2 + σs2 ) ×



 1 √ Q −1 (Pd_i ) + 1 . N

(37)

By substituting (37) into (34), the local false alarm probability in the LCDR criterion is calculated accordingly. In the following, a standard cooperative sensing architecture is first presented, followed by two proposed weighted cooperative sensing algorithms to further improve the sensing performance in terms of the global detection probability Pd_G and the global false alarm probability Pfa_G . As to the architecture of a standard cooperative sensing system, the cooperative sensing team consists of n sensing nodes, and each node is in a different SNR (Radar signal to WGN power ratio) environment. Sensing node i collects the observations yNi , yNi −1 , . . . , y1i i = 1, 2, . . . , n, during the sampling time t (one sensing cycle) and makes its local decision ui based on its local node decision rule which is given by,



+1 ui = −1

yi ≥ Vthreshold_i yi < Vthreshold_i .

(38)

Then these local decisions are sent to a team coordinator (fusion node), based on a certain fusion rule to make the global decision u0 .

u0 =

    1

n X

   0

X

ui ≥ K

i=1 n

(39) ui < K .

i=1

Here a MAJORITY fusion is chosen, which declares that Prule n the radar exists when i=1 ui ≥ K for K = 0. (If over half of the local decisions declare the existence of a radar signal, then the global decision declares that a radar signal exists and being swept by the radar mainlobe). Other fusion rules, e.g. AND/OR rules are discussed in [22]. A performance comparison between the standard cooperative sensing and local single sensing is shown in Fig. 6. In this case, for the same SNR environment, the standard cooperative sensing (6-node sensing team) provides much better sensing performance compared to the local sensing. When all the nodes in the cooperative sensing team are located in a similar SNR environment, the standard cooperative sensing architecture is a suitable solution for sensing the primary system. However, when nodes are in different SNR environments, e.g. in the antenna-rotating radar systems, the global sensing performance can degrade as indicated in Fig. 6.

Fig. 6. Performance comparison among standard cooperative sensing and local single sensing (radar parameters: PRT = 555 µs, pulsewith = 10 µs).

By means of adequately weighting the sensing credibility of the local nodes considering their SNRs, a weighted cooperative sensing architecture is proposed as shown in Fig. 7. Where the weighted fusion rule is given by,

u0 =

    1

n X

   0

X

Wi × ui ≥ K

i=1 n

(40) Wi × ui < K .

i=1

Pn

That declares a radar exists if i=1 Wi × ui ≥ K , K = 0 for the MAJORITY fusion rule. Firstly, an analytical weighted algorithm is proposed as follows, which is based on Pd_i -related weights and Pfa_i related weights. Different weighting factors for the nodes are calculated according to the local sensing performance of these nodes. In the criterion of LCDR with a local constant detection probability Pd_i , the weighting factor Wi for each local sensing node is defined as, Wi = 1 − Pfa_i .

(41)

Since each node is assigned the same fixed target local detection probability, their local false alarm probability Pfa_i is different corresponding to the different SNRs experienced. A node in a higher SNR location has a lower Pfa_i , in other words, it has a more reliable local sensing performance, therefore its contribution to the global sensing performance is expected to be more significant than those nodes with higher Pfa_i . In (41), this is considered by increasing the influence of the more creditable nodes with the better sensing performance on the global sensing decision making. In the criterion of LCFAR with a local constant false alarm probability Pfa_i , the weighting factor Wi is defined as, Wi = Pd_i .

(42)

Similar to the case of LCDR, those nodes that have a higher local detection probability Pd_i will have a more significant contribution to the global sensing decision.

68

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Fig. 7. Weighted cooperative sensing architecture.

In addition, a SNRdirect weighted algorithm is proposed as, Wi = Normalised_SNRi .

(43)

Where the weighting factor Normalised_SNRi is the normalised SNR value for each node, the basic concept of this algorithm is that the nodes having better SNRs contribute more in the global decision making, and these weighting factors have no relations to the preset target Pd_i or Pfa_i for each node. In our simulator, a team coordinator will group a 6-node sensing-active team by selecting the most suitable team nodes, which have different SNRs because of their different distance spreads. The team SNR is defined as the median of the SNR range in this team. e.g. team 1 with a SNR of −20 dB indicates that the nodes in team 1 have uniformly distributed SNR values in a range from −25 dB to −15 dB. A performance comparison between a standard cooperative sensing algorithm and the weighted cooperative sensing algorithms is presented as follows.

Fig. 8. Performance comparison between the SNRdirect weighted, analytical weighted, standard algorithms in terms of the global false alarm probability (LCDR criterion).

• Local Constant Detection Rate (LCDR) When nodes are assigned the same target local constant detection probability Pd_i in one team, teams with different SNR values can achieve the same global detection probability Pd_G , regardless of the global false alarm probability performance. Thus a protection level to the radar system can be guaranteed. As can be seen in Fig. 8, the weighted cooperative sensing algorithms lead to a substantially lower global false alarm probability Pfa_G than that of the standard cooperative sensing algorithm. As expected, teams which are in a better SNR region, have a lower Pfa_G than those teams in poorer environments. It can be seen in Fig. 9 that the global detection probability of the standard cooperative sensing algorithm is slightly better than those of the weighted cooperative sensing algorithms, but all algorithms outperform the assigned target local detection probability. It also shows that the analytical weighted algorithm has slightly better detection probability than the SNRdirect weighted algorithm. Nevertheless, as indicated in Fig. 8, the SNRdirect algorithm offers a better performance in terms of the global false alarm probability. The choice of the cooperative sensing algorithm is based on the requirements of the target global detection

Fig. 9. Performance comparison for different local constant detection probabilities (LCDR criterion).

probability and target global false alarm probability. In the case when a higher detection probability is required for a higher protection level to primary systems, the analytical

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

algorithm that provides a higher Pd_G is preferred. When more opportunistic access for secondary cognitive nodes is the sensing priority, the SNRdirect algorithm is likely to be chosen, which leads to a lower Pfa_G .

• Local Constant False Alarm Rate (LCFAR) By assigned the same local false alarm probability Pfa_i for each team node, it shows that the weighted algorithms provide a better global detection performance Pd_G than that of the standard one. Fig. 10(a) shows the performance where Pfa_i = 0.1. The SNRdirect weighted algorithm has similar detection performance with the analytical weighted algorithm, and both of them offer much better detection performance than using the standard cooperative sensing algorithm. When we consider the performance of the global false alarm probability, all algorithms provide a similar false alarm performance as shown in Fig. 10(b). For a higher requirement on the global detection probability, e.g. Pd_G > 99% (sensing teams with SNR >−18 dB), the analytical weighted algorithm offers the best false alarm performance, followed by the standard algorithm and the SNRdirect weighted algorithm. Note that the Pfa_G of all the algorithms outperform the assigned target local constant false alarm probability of 0.1. The overall performance of the different sensing algorithms on sensing a pulsed radar signal is shown in Fig. 11. The Receiver Operating Characteristic (ROC) curves show the weighted algorithms provide better sensing performance than the standard sensing algorithm in terms of both the global detection probability and the global false alarm probability. In this section, two weighted cooperative sensing algorithms are proposed using two sensing-priority criteria to improve the performance of cooperative sensing in a pulsed radar system. The Receiver Operating Characteristic curves of all the algorithms conclude that two weighted algorithms offer better global sensing performance than using standard cooperative sensing algorithm. When considering the LCDR criterion, the weighted algorithms provide a better global false alarm performance than that of

(a) Global detection probability.

69

the standard algorithm. Therefore better opportunistic access to the radar spectrum for secondary communication is provided due to the lower global false alarm probability, while at the same time maintaining the interference into the radar at the required levels. When considering the LCFAR criterion, the weighted sensing algorithms provide a better global detection performance than that of the standard one, and similar global false alarm performance to that of a standard algorithm. Therefore it is critical to choose the most suitable sensing criterion and sensing algorithm according to the different cognitive environments and sensing priorities, which can be determined by regulators. 4. Node teaming performance evaluation In this section, the performance of the node teaming algorithms described in previous sections will be presented. Section 4.1 shows the benefits of the sensing-active team selection algorithm. The sensing performance of the sensing assignment algorithms in combination with the weighted sensing algorithms will be shown in Section 4.2. 4.1. Selection of the sensing-active team This section presents the performance results obtained in the simulator by using the team node selection algorithm described previously. It is assumed that each node receives the same SNR, e.g. 0 dB, when being swept by the radar mainlobe. The distance spread between adjacent team nodes ranges from zero radar mainlobe angle width (0 degrees) to one radar mainlobe angle width (1.4 degrees). As can be seen in Fig. 12, n highest-SNR nodes are chosen to join a sensing-active team in each sensing period (e.g. n = 6 in our scenario). A suitable number of team nodes should be determined by considering the availability of the responsive candidate nodes, as well as their sensing signalling cost. Although the team SNR decreases when the distance spread between adjacent team nodes increases, 6-node sensing teams can still

(b) Global false alarm probability.

Fig. 10. Performance comparison between SNRdirect weighted, analytical weighted, standard algorithms (LCFAR criterion).

70

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Fig. 11. Performance comparison among weighted cooperative sensing algorithms, standard cooperative algorithm and local single sensing (radar parameters: PRT = 555 µs, pulsewith = 10 µs).

Fig. 13. The ROC performance of different sensing node assignment algorithms combined with SNRdirect weighted sensing algorithm and standard sensing algorithm.

at all times although they only need to perform the sensing task at a certain sensing cycle. Therefore, it is quite fair for all nodes to do the sensing contribution and share the spectrum access opportunities at different sensing periods as long as their cognitive transmission will not cause unacceptable interference to the radar. 4.2. Sensing assignment for the selected team nodes

Fig. 12. Achievable Team SNR by using team node selection algorithm in a swept radar scenario.

achieve much higher team SNRs for the required sensing performance, compared to the SNR received by a single node which only has a peak SNR during the mainlobe sweeping period. In a swept radar scenario, by updating the sensingactive team for each sensing period, the optimised sensing performance can always be reached. Sensing fairness is considered by introducing a sensing information-sharing mechanism. At the end of each sensing cycle, the sensingactive team coordinator will share the updated radar spectrum state with the sensing-hibernated team nodes. From a cognitive opportunistic access perspective, it means that these sensing-hibernated nodes can choose to access the radar spectrum at a higher transmission power because they are not being swept by the radar mainlobe (in other words, not being directly pointed by radar receiver antenna) and therefore cause less interference to the radar. By using the sensing information-sharing mechanism, all the sensing participants have knowledge of the radar state

Team nodes with different sensing bandwidths can experience different highest-SNR at different frequency sub-channels. If more sensing bandwidth and therefore frequency diversity is available, then by sensing at the highest SNR frequency sub-channel we can obtain a better sensing performance. The following analysis is based on the assumption that the team nodes have 5 MHz single side sensing bandwidth, which can cover the entire radar signal bandwidth. Fig. 13 shows the ROC performance comparison among five different node assignment algorithms considering two sensing algorithms namely: (i) SNRdirect weighted sensing algorithm and (ii) standard [non-weighted] sensing algorithm). For the same average team SNR, the HA Algorithm, which assigns each team node with its best sub-channel at an overall optimal team SNR basis, offers the optimal overall sensing performance by exploiting the frequency diversity. The HA algorithm offers almost the same sensing performance as the MaxSense algorithm besides guaranteeing that there are no same frequency sub-channels used by two different nodes at the same sensing cycle, thereby solving the double-false alarm problem. However, as shown in Fig. 14, the HA algorithm requires high computational complexity in the sensing team coordinator which needs to calculate and assign the sensing frequency sub-channels. The EasySense algorithm has a slightly worse sensing performance compared to the above two algorithms but significantly better than the performance of the TimedomainSense which performs the sensing without exploiting the frequency diversity. More importantly, the EasySense

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

Fig. 14. Comparison of the computational complexity among sensing node assignment algorithms.

algorithm provides a much lower computational complexity than the HA algorithm as shown in Fig. 14. As expected, the EasySense2 algorithm offers a very competitive sensing performance compared to the HA and MaxSense algorithms while only requiring a slight increase in computational complexity over the EasySense algorithm. As shown in Fig. 11 and further confirmed in Fig. 13, the SNRdirect weighted cooperative sensing algorithm provides a superior sensing performance to the standard cooperative sensing algorithm regardless of which node assignment algorithm is chosen. Furthermore, by using the SNRdirect weighted sensing algorithm, the introduction of the nodes’ sensing credibility (weighting factors) for the sensing algorithm effectively compensates for the loss in the sensing performance resulting from those lowcomplexity node assignment algorithms, e.g. EasySense algorithm, therefore making the performance of the four frequency node assignment algorithms close to each other. Consequently, under the limitation of the node computational complexity, the EasySense algorithm can be chosen to offer a competitive sensing performance in combination with the weighted sensing algorithm.

71

overall combinatorial SNR of the available nodes offers almost the same sensing performance as the MaxSense algorithm. In addition, it effectively solves the double-false alarm problem that may happen in the MaxSense algorithm by assigning nodes to perform the sensing task at different frequency sub-channels in one sensing cycle. To reduce the computational complexity and the corresponding power consumption of the sensing nodes, the EasySense algorithm is proposed to achieve a good sensing performance, which is close to the performance of the MaxSense and the HA algorithms. Meanwhile, the EasySense algorithm also solves the double-false alarm problem and significantly decreases the requirement of the computational complexity on the sensing-active team coordinator. By tracking the 2nd highest-SNR node index for each frequency sub-channel, the EasySense2 algorithm offers a good solution to the trade-off between high sensing performance and low computational complexity. In addition, the introduction of the weighted sensing algorithm that considers each team node’s sensing credibility leads to a satisfactory sensing performance even using a lowcomplexity assignment algorithm which relaxes the team coordinator’s complexity requirement. Hence the key issue is to select the most suitable cognitive node teaming algorithm along with the weighted sensing algorithm by considering the sensing-priority requirement, the characteristics of the primary signal and the cognitive secondary team coordinators’ computation capabilities.

Acknowledgements This project is funded as part of the Core 4 Research Programme of the Virtual Centre of Excellence in Mobile & Personal Communications, Mobile VCE, (www.mobilevce.com), whose funding support, including that of EPSRC is gratefully acknowledged.

References [1] J. Mitola III, G.Q. Maguire Jr., Cognitive radio: Making software radios more personal, IEEE Personal Communications 6 (4) (1999) 13–18.

5. Conclusion

[2] M.I. Skolnik, Introduction to Radar Systems, 3rd ed, McGraw-Hill Higher Education, New York, Dec. 2000.

In this paper, opportunistic spectrum access in a swept radar system is investigated, several node teaming and sensing algorithms have been considered for cooperative sensing in a mobile scenario. These algorithms were then presented in terms of the team node selection in the time domain and the sensing assignments of the team nodes in the frequency domain. For the team node selection, the proposed algorithm effectively updates the sensingactive team and selects the team nodes in a swept radar scenario. Consequently, the maximum sensing team SNR can be obtained. When it comes to the team node assignment performance, the MaxSense algorithm has the lowest complexity and the highest average team SNR, which consequently leads to a high sensing performance. The Hungarian Algorithm (HA) algorithm which maximises the

[3] Analysys Consulting Limited: Spectrum demand for nongovernment services 2005–2025, Final Report for the Independent Audit of Spectrum Holdings, 2005, Available at http://www.spectrumaudit.org.uk/pdf/spectrum_demand.pdf. [4] QinetiQ: An Investigation into the characteristics, operation and protection requirements of civil aeronautical and civil maritime radar systems, Radio Communications Agency Contract AY4051, Oct. 2002. [5] BAE Systems: Study into spectrally efficient radar systems in the L and S bands, Ofcom spectral efficiency algorithm 2004–2005 (SES2004-2), 2006. [6] S.M. Mishra, A. Sahai, R.W. Brodersen, Cooperative sensing among cognitive radios, in: IEEE International Conference on Communications, 4, 2006, pp. 1658–1663. [7] A. Ghasemi, E.S. Sousa, Collaborative spectrum sensing for opportunistic access in fading environments, in: IEEE International Symposium on New Frontiers in Dynamic Spectrum Access Networks, DySPAN 2005, 2005, pp. 131–136.

72

L.S. Wang et al. / Physical Communication 2 (2009) 58–72

[8] L.S. Wang, J.P. McGeehan, C. Williams, A. Doufexi, Application of cooperative sensing in radar-communications coexistence, IET Communications 2 (6) (2008) 856–868. Cognitive Spectrum Access (special issue). [9] L.S. Wang, C. Williams, A. Doufexi, J.P. McGeehan, Weighted cooperative sensing for pulsed radar signals, in: IET International Seminar on ‘Wideband and Ultrawideband systems and technologies’, London, 2008. [10] A. Ghasemi, E.S. Sousa, Asymptotic performance of collaborative spectrum sensing under correlated log-normal shadowing, IEEE Communication Letters 11 (2007) 34–36. [11] B. Yang, H. Yu, H. Li, H. Hou, A coverage-preserving density control algorithm based-on cooperation in wireless sensor networks, in: International Conference on Wireless Communications Networking and Mobile Computing, WiCOM 2006, 2006, pp. 1–4. [12] E. Peh, Y.C. Liang, Optimization for cooperative sensing in cognitive radio network, in: IEEE Wireless Communications and Networking Conference, WCNC 2007, 2007, pp. 27–32. [13] Rec.ITU-R.M.1464-1: Characteristics of radiolocation radars, and characteristics and protection criteria for sharing studies for aeronautical radionavigation and meteorological radars in the radiodetermination service operating in the frequency band 27002900 MHz, 2003. [14] W.A. Gardner, Signal interception: A unifying theoretical framework for feature detection, IEEE Transactions on Communications 36 (8) (1988) 897–906. [15] TR 101 112 V3 2 0 (1998-04): Selection procedures for the choice of radio transmission technologies of the UMTS, (UMTS 30.03 version 3.2.0). [16] A. Ghasemi, E.S. Sousa, Spectrum sensing in cognitive radio networks: Requirements, challenges and design trade-offs, IEEE Communications Magazine 46 (4) (2008) 32–39. [17] T.A. Weiss, J. Hillenbrand, A. Krohn, F.K. Jondral, Efficient signalling of spectral resources in spectrum pooling systems, in: Proc. 10th Symp. Commun. and Vehic. Tech., SCVT, 2003, pp. 1–6. [18] R. Chen, J. Park, Y. Hou, J.H. Reed, Toward secure distributed spectrum sensing in cognitive radio networks, IEEE Communications Magazine 46 (4) (2008) 50–55. [19] H.W. Kuhn, The Hungarian method for the assignment problem, Naval Research Logistics Quarterly 2 (1955) 83–97. [20] Q. Cheng, B. Chen, P.K. Varshney, Detection performance limits for distributed sensor networks in the presence of nonideal channels, IEEE Transactions on Wireless Communications 5 (2006) 3034–3038. [21] S. Kay, Fundamentals of Statistical Signal Processing-Volume II Detection Theory, Prentice-Hall PRR, New Jersey, 1998. [22] P.K. Varshney, Distributed detection and data fusion, Springer, New York, USA, 1997.

Lingfeng Stephen Wang received the B.Eng. degree in Electronic Engineering from the University of Shanghai for Science and Technology in 2003, graduated with the best graduate award. He received his M.Sc. degree (with Distinction) in Advanced Photonics and Communications from the University of Warwick in 2005. He served as a wireless communication engineer in China Telecom and Nokia respectively before pursuing his Ph.D. in the Centre for Communications Research at the University of Bristol as a Mobile VCE researcher. His current research interests include cognitive radio systems, spectrum sensing and allocation in wireless communication systems and radar systems.

Angela Doufexi is currently Senior Lecturer in Wireless Networks at the University of Bristol. She graduated from the University of Athens with a B.Sc. in Physics in 1996. She received her M.Sc. in Electronic Engineering from Cardiff University in 1998. She then joined the Centre for Communications Research at the University of Bristol, where she received her Ph.D. in 2002. She has worked in a number of projects including the European IST projects SATURN, ROMANTIK, WCAM and ASTRALS. Her research interests include OFDM and OFDMA systems, multiuser diversity and resource allocation, wireless LANs and WiMAX (PHY and MAC), space time coding and MIMO, LTE and 4th generation communications systems, cognitive radio and multimedia transmission. She has published over 80 journal and conference papers in these areas.

Chris Williams received an M.A. in Engineering Science from Oxford University. After graduating he joined the Defence Research Agency where he was involved in RF ECM research and development. He then did a Ph.D. in chaotic communication systems at Bristol University. He joined the Defence Evaluation Research Agency Communications Department in October 1997, and worked as a research team leader on physical layer research and chaotic systems. He then joined the Centre for Communications Research in the University of Bristol in 2002 and worked on synchronisation for OFDM and multiple antenna systems. Between 2004 and 2006 he was an RCUK Academic Research Fellow, carrying out research into robust detection of chaotic signals for communications and active sensing systems. In 2007 he joined the Fujitsu Laboratories of Europe, where he led research and development on WiMAX systems. Since 2008 he has been the communications infrastructure team leader with the Defence Science Technology Labs. His other interests include cognitive radio, UWB systems and interference mitigation.

Joe McGeehan received the Ph.D. and D.Eng. degrees in Electrical and Electronic Engineering from the University of Liverpool in 1971 and 2003 respectively. He is currently Director of the Centre for Communications Research, University of Bristol and Managing Director of Toshiba’s Telecommunications Research Laboratory. Since 1973 he has been researching spectrum-efficient mobile-radio communication systems and has pioneered work in many areas including linearised power amplifiers, WCDMA(3G) and smart antennas He was joint recipient of the IEEE Vehicular Technology Transactions ‘Neal Shepherd Memorial Award’ for work on SMART Antennas, and the IEE Proceedings ‘Mountbatten Premium’ for work on satellite-tracking He is a Fellow of the Royal Academy of Engineering and was awarded a CBE in 2004 for services to the Communications Industry. In 2004 he was listed as one of the world’s top technology agenda setters by silicon.com (USA). He has served on numerous international committees and was advisor to the UK’s first DTI/MOD ‘‘Defence Spectrum Review Committee’’ in the late 70s.