Received May 30, 2017, accepted June 24, 2017, date of publication July 11, 2017, date of current

version July 31, 2017.

Digital Object Identifier 10.1109/ACCESS.2017.2724601

Enhanced System Acquisition for NB-IoT

WENJIE YANG1 , MIN HUA1 , JINGJING ZHANG1 , TINGTING XIA1 , JUN ZOU1 ,
CHENGLING JIANG2 , AND MAO WANG3
1 Wireless Networking and Mobile Communications Group, School of Electronic and Optical Engineering, Nanjing University of Science and Technology,
Nanjing 210094, China
2 Information and Telecommunication Branch, State Grid Jiangsu Electric Power Company, Nanjing 210024, China
3 National Mobile Communications Research Laboratory, Southeast University, Nanjing 211111, China

Corresponding author: Wenjie Yang (wj_yang@njust.edu.cn)

This work was supported in part by the Open Research Fund (2017B01) of the National Mobile Communications Research Laboratory,
Southeast University, and in part by the Fundamental Research Fund for the Central Universities under Grant 30920140122005 and
Grant 30917011318.

ABSTRACT Machine-type communication (MTC) is the key technology to support data transfer among
devices (sensors and actuators) in Internet of Things (IoT). Although cellular communication technologies
are developed mainly for ‘‘human-type’’ communications, enabling MTC with cellular networks not only
improves the connectivity, accessibility, and availability of an MTC network but also has the potential to
further drive down the operation cost. However, cellular MTC, especially when applied to low-power massive
IoT (mIoT), poses some unique challenges due to the low-cost and low-power nature of an mIoT device.
One of the most challenging issues is providing a robust way for an mIoT device to acquire the network
under a large frequency offset due to the use of low-cost crystal oscillators and under extended coverage.
Although differentiation is a well-known technique for removing impairments caused by frequency offset,
its ‘‘noise amplification’’ effect limits its applications in cellular communications due to the fact that
cellular communication is typically interference limited. Matched-filter-based detection is, therefore, almost
unexceptionally used. We show that the differential technique can actually benefit system acquisition in
mIoT, where the use of low-cost crystals is a default. Although the existing system acquisition design in a
cellular mIoT system, i.e., NB-IoT, facilitates both techniques, there still remain issues that need to be solved
in order to take full advantage of the design. We provide a comprehensive analysis on the performance of two
most common techniques when applied in a typical NB-IoT environment based on two factors, the geometry
factor and the frequency offset factor. Finally, we derive the operating regions for matched-filter-based
detection and differentiation using these two factors, in which the system acquisition performance of the
two types of techniques is maximized for NB-IoT.

INDEX TERMS Cellular machine-type communications, system acquisition, NB-IoT, low-power massive
Internet-of-Things.

I. INTRODUCTION low-power massive IoT (mIoT) market and the pressure

Internet-of-Things (IoT) enables the objects to exchange from the cellular providers, recently (September 2015), the
information on a massive and global scale [1]. There are a 3rd Generation Partnership Project (3GPP) started a work
myriad of applications with low requirements on data rate item to standardize a new air-interface, known as Narrow-
and time delay, such as smart cities, metering, and remote Band IoT or NB-IoT [5]–[8] for short, operating on a 180-kHz
sensing, whose connectivity is powered by machine-type bandwidth (the bandwidth of an LTE resource block (RB)
communications (MTC) that are characterized by low power, for both downlink and uplink), to enable the power and
long battery life (e.g., 10 years or more), wide coverage, and spectrally more efficient operation of MTC in the bands
support for devices on a massive scale (hence they must be that are narrower than the minimum LTE system operating
low cost to manufacture) [2]. The competition in technologies bandwidth (1.08 MHz), particularly in the 200-kHz GSM
of this category is becoming intense. bands. This design achieves great coexistence performance
In response to the challenges from other technologies with GSM and LTE system. NB-IoT also establishes the
(e.g., the proprietary systems like Ultra Narrow Band foundation of a new solution for mIoT, which will be an
by SIGFOX [3], LoRa by Semtach [4]) in this growing important integral part of the fifth generation (5G) systems

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 5, 2017 Personal use is also permitted, but republication/redistribution requires IEEE permission. 13179
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 1. Typical operations involved in device-originated (power up or in response to higher layer data transfer request) and
network-originated (paging) calls. (a) Device-originated calls. (b) Network-originated calls.

in the future [9]–[11]. The core specifications of NB-IoT factor. The contribution of this paper is thus to show that
were finished in June 2016, and the commercial launch is the matched-filter detector precedes the differential detector
expected to be initiated in 2017. Due to the extreme short in performance when geometry factor prevails and the dif-
time for development, NB-IoT reused most part of the LTE ferential detector becomes superior when frequency offset
air interface. factor dominates. According to the actual operating condi-
Nevertheless, the greatest challenge that cellular MTC tion, i.e., the dominance of the two factors, the device can
faces is the large discrepancy (in transceiver properties and thus select the right detector so that the immunity to both
applications) between the NB-IoT device and the LTE user frequency offset and low signal quality can be maximized.
equipment (UE) for human-type communications (HTC). Section II gives a brief description of the system acquisi-
Among those, low cost of production, deployment, and main- tion procedure and points out the challenges of the synchro-
tenance is one of the most important aspects of NB-IoT nization signal detection in NB-IoT. In Section III, we ana-
devices so that they can be deployed on a mass scale and even lyze the detection performance of two detection techniques,
in a disposable manner, contrasting the high-cost nature of the i.e., a traditional matched-filter based detection and a dif-
LTE UE. Consequently, the transceiver performance may be ferential detection. Section III also makes a comparison of
impaired, and certain cellular operations, like initial system these two techniques and addresses their limitations respec-
acquisition, designed for a regular LTE UE, may not function tively when applied to NB-IoT. Based on the analysis results,
well under certain circumstances. Section IV derives the operating regions for these two tech-
The purpose of this paper is to analyze the impact of niques, thereby enabling a device to select a detector accord-
the frequency offset on the initial system acquisition caused ingly to increase the frequency offset immunity of a device in
by an inaccurate crystal oscillator with a focus on the low- an NB-IoT environment. Section V concludes this paper.
power mIoT system, in particular NB-IoT. We investigate
the behavior of different acquisition techniques, i.e., the II. SYSTEM ACQUISITION AND FREQUENCY
matched-filter technique and the differentiation technique, OFFSET IN NB-IoT
under NB-IoT operating conditions, and we show that the Fig. 1 shows typical operations performed by an NB-IoT
detection performance is governed by two factors introduced device during power-up or wake-up in response to a data
in this paper: the geometry factor and the frequency offset transfer request (e.g., meter reporting) or paging timer

13180 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

of the synchronization signal may trigger a full frequency

re-scan, thereby incurring excessive power consumption and,
consequently, a shortened battery life.
Most mIoT applications are characterized by bursty low
rate transmissions. The size of data involved in each
transaction is typically small. For example, a few trans-
missions a day and several kilobytes per transmission is
typical for an automated water meter. Assuming 1 kilobyte
per report, the data transmission takes ∼32 subframes
(i.e., 32 ms). The NPSS detection thus may account for at
least 10 ms/(10 + 32) ms =24% of the total power consump-
tion considering that the NPSS detection takes at least 10 ms
to accomplish one search. System acquisition thus becomes a
significant part of the total power consumption per wakeup.
Therefore, it is a key factor of an mIoT system, and cannot be
overlooked in the overall system design.
The challenge of the detection of the synchronization sig-
nal lies in the fact that (1) the device has no knowledge about
the system timing, and (2) the local frequency of the device
is not yet synchronized to the network, after waking up from
a sleep, especially, during initial power-on or after a long
sleep, and a large offset/error may exist between the device
and the network. These large time and frequency uncertainties
seriously impair the acquisition performance.
The frequency offset between the device and the network
is determined by the accuracy and stability of the local ‘‘free-
running’’ crystal oscillator of an NB-IoT device. The offset
FIGURE 2. Flowchart of the NB-IoT system acquisition.
is thus the deviation of the oscillator frequency from the
specified target frequency (i.e., the cellular network carrier
frequency). Factors such as temperature, crystal aging, and
vibration affect the frequency accuracy and stability of the
expiration. Upon waking up, the device obtains the system crystal. Among these, temperature is the primary factor that
acquisition information from the cell selection protocol that affects the accuracy, and hence needs to be compensated in
maintains a list of most recently used (MRU) cells. In the case practice.
that none of the cells in the list can be found or in the case of The offset (without compensation) typically ranges from
the first time power-up, the device performs a full frequency ±5 ppm to ±100 ppm (or ±5 kHz to ±100 kHz at 1- GHz
scan looking for a new cell or system at a frequency raster of, carrier frequency) depending on the quality of the crystal. For
e.g., 100 kHz. To simplify the search, the cell selection pro- a low-cost NB-IoT device, the initial frequency error could
tocol also maintains a list of the candidate systems and their be well above ±5 ppm (e.g., ±20 ppm or ±20 kHz at 1-GHz
associated acquisition information provided by the operator, carrier frequency), especially for the very first initial system
commonly referred to as a preferred roaming list or PRL [12]. acquisition when the local crystal oscillator is never synchro-
In either case, whether the device is looking for a cell in the nized to the network (i.e., no compensation/correction has
current system or an entirely new system, it must first perform ever been made). Since the frequency accuracy of a crys-
the system acquisition procedure as shown in Fig. 2. The tal depends heavily on temperature, a cellular HTC device
device searches for the downlink synchronization signals, is thus typically equipped with a temperature-compensated
e.g., a narrowband primary synchronization signal (NPSS) crystal (TCXO). Not only is a TCXO costly but also power
and a narrowband secondary synchronization signal (NSSS) consuming to maintain, which is a luxury that a typical
in NB-IoT system, at the raster frequency to detect the pres- NB-IoT device cannot afford. Clearly, the system acquisition
ence of a cell at this frequency and acquire the accurate timing for an NB-IoT device is more challenging than that for an
and frequency of the system, e.g., the symbol and frame HTC device (e.g., LTE UE). Therefore, one of the key issues
timing, carrier frequency, and sampling clock. The device in NB-IoT is efficient detection of the synchronization signal
can then decode the system information block transmitted on under potentially large frequency offset.
the downlink broadcast channels (e.g., narrowband physical Besides, there is a substantial market for the mIoT use
broadcast channel (NPBCH) in NB-IoT) and obtain the sys- cases in which devices are deployed deep inside a build-
tem information necessary for establishing a communication ing (e.g., in the basement), which requires coverage enhance-
link with the system [13], [14]. A failed detection (i.e., a miss) ment over the current LTE cell footprint (maximum coupling

VOLUME 5, 2017 13181

 W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 3. Ilustration of the narrowband primary synchronization signal (NPSS) resource allocation in an NB-IoT system.

loss of 144 dB). A coverage extension of 20 dB for NB-IoT where

is hence targeted to increase the coverage to areas where
NB-IoT devices are potentially deployed [2]. This extension xm = cm s
= xm,i = cm si , 0≤i≤K −1 .

provides the cellular network with the ability to support (2)
devices in locations with excessive penetration losses, where
the traditional cellular system has difficulty to reach. Thus, Here 0 ≤ m ≤ M − 1 is the OFDM symbol index, 0 ≤ i ≤
another unique challenge for synchronization signal detection K − 1 is the time-domain sample index of an OFDM symbol,
lies in the reduced signal quality due to the coverage exten- and
sion requirement in NB-IoT. ( K −1
)
1 X j2π Kk i
The downlink synchronization signals in NB-IoT consist of s = si = √ uk e ,0 ≤ i ≤ K − 1 (3)
two periodically-broadcast signals: an NPSS and an NSSS. K k=0
NPSS serves the purpose of providing initial timing and
is the NPSS base waveform in time domain, generated from
frequency information; and NSSS is for timing and fre-
a Zadoff-Chu (ZC) sequence in frequency domain (via an
quency refinement (among other things) under much less
inverse Fourier transform as shown in (3)), occupying the
time and frequency ambiguities. Clearly, the detection of
K = 11 subcarriers of an OFDM symbol,
NPSS is performed under the largest time and frequency
ambiguities, and is hence the most challenging part of πµk(k+1)
n o
uk = e−j K , 0 ≤ k ≤ K − 1 , (4)
the overall system acquisition, and ergo the focus of this
paper. with µ = 5. Since a prime number (K = 11) is required by a
ZC sequence [15], [16], the 12th subcarrier is left unused. This
III. NPSS DETECTION WITH FREQUENCY OFFSET base waveform, s, which has a constant amplitude of 1 [17],
The generation of an NPSS waveform is from the fre- occupies the duration of one OFDM symbol and is repeated
quency domain sequence using the K = 11 subcarri- M = 11 times covering the last 11 OFDM symbols of
ers (out of 12) of the last M = 11 OFDM symbols Subframe 5 to form the NPSS. On top, an 11-bit sequence,
of Subframe 5, as shown in Fig. 3. It is then trans- i.e., c = [1111 − 1 − 1111 − 11], is applied as a cover
formed into a time-domain waveform (via inverse Fourier sequence for the 11 repeated base waveforms to avoid the
transform), and transmitted in each radio frame. Specifi- timing ambiguity as a result of the repetitions. As such, cm
cally, the time domain waveform can be mathematically in (2), i.e., the mth element of c, is used as the cover code
expressed as for the mth repeated base waveform on the mth NPSS OFDM
symbol(m = 0, 1, · · · , 10).
The received ith sample of the mth base signal of the
x = x0 x1 · · · xm · · · xM −1 , baseband NPSS by the device with a frequency offset 1f
 
(1)

13182 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 4. A simplified baseband model of a NPSS transceiver, where the transmitter is the base station, and the detector is the NB-IoT device.

between this device and the network (i.e., the base station) NPSS signal is corrupted by noise as shown in (5), the output
can be expressed as of the correlator becomes
1f
(i+mK ) M −1 K −1
ym,i = hxm,i ej2π B + vm,i 1 XX
γ MF = ∗


1f ym,i · xm,i
(i+mK ) MK
= hcm si ej2π B + vm,i , 0 ≤ m ≤ M − 1, m=0 i=0
0 ≤ i ≤ K − 1, (5) M −1 K −1 
1 X X 1f
(i+mK )

∗
= hxm,i ej2π B + vm,i · xm,i
where B is the NPSS signal bandwidth, vm,i is the independent MK
m=0 i=0
zero-mean complex Gaussian noise with variance σ 2 , and = hγ (1f ) + υ, (7)
h is the complex channel gain, unknown but assumed to be
−1 KP
MP −1
constant over the NPSS transmission duration.
 
where υ , 1 ∗ ∼ CN 0, σ 2 , and
vm,i xm,i
The SNR of the received sample ym,i in (5), henceforth MK MK
m=0 i=0
referred to as the geometry of the device, is defined as
1f
 
jπ 1f
B (MK −1)
hxm,i 2
 
|h|2 |h|2 γ (1f ) = e sinc , (8)
η, B/MK
2 o = n  2 o = 2 . (6)
σ
n
E vm,i  E vm,i  where
Geometry is commonly used in the cellular community to 1 sin (π x)
sinc (x) , · , (9)
indicate the position of a device in a cell. A device located MK sin (π x/MK )
at the center of the cell, i.e., at a high geometry, typically
enjoys a high SNR and a device at the edge of the cell, i.e., indicating that the received
 NPSS
 strength is attenuated by
1f

a factor of |γ (1f )| = sinc B/MK thereby entailing a
 
at a low geometry, suffers from a low SNR. Geometry is ,
interchangeable with received signal SNR in the following reduction in the detected NPSS energy by |γ (1f )|2 , due to
discussion. the frequency offset 1f . For instance, a |1f | value about
Based on its unique structure, there are mainly 5 kHz causes a reduction by more than 20 dB (cf. Fig. 5).
two well-known baseline methods that are best suited for In order to keep the degradation manageable, |1f | must be
NPSS detection: the matched-filter detection [18]–[20] and well below 5 kHz (i.e., 1f  5 ppm at 1-GHz carrier
the differential detection [21], [22]. Although there are other frequency), which can be hard to achieve for low-cost NB-IoT
techniques, they can be considered as variants of these two devices. It is clear that the optimal matched-filter technique
fundamental techniques. for the synchronization signal detection in the absence of a
frequency offset/error, suffers signal energy attenuation in the
A. MATCHED-FILTER DETECTION presence of a high frequency offset between the transmitter
To detect the presence and timing of a system, a device and a receiver. Thus, the frequency offset has a profound
employs a detector that is essentially a correlator that per- effect on the performance of NPSS detection for a low-cost
forms correlation of the received signal at every sample point NB-IoT device that may have a large frequency offset, espe-
within a duration whose length equals the period of the NPSS cially during the initial system acquisition or after wakeup
(e.g., 10 ms) against a local copy/template of the NPSS wave- from a long sleep.
form. This correlator is the traditional matched-filter detector, It is worth mentioning that there is an important variant
as shown in Fig. 4. In the absence of a frequency error/offset, of the baseline matched-filter detector that, instead of using
the largest correlation happens when the local waveform is a single local template x [cf. (1)], this technique consists of
aligned with the received NPSS signal. When the received H template waveforms, each of which matches to the NPSS

VOLUME 5, 2017 13183

 W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 5. Plot of |γ (1f )|2 in (8) as a function of frequency offset 1f ∈ (−20kHz, 20kHz) , i.e., ±20 ppm at 1-GHz carrier frequency,
to show how frequency offset changes the detection energy.

waveform with a hypothesized frequency offset [23], i.e., signal is then

K −1
xq = xqm , 0 ≤ m ≤ M − 1 ,

(10) 1 X
∗
φm =


cm−1 ym−1,i · cm ym,i
K
where i=0
K −1

1fq 1fq 1 1f

cm−1 hcm−1 si ej2π B (i+(m−1)K ) + cm−1 vm−1,i
X
xqm = xm · diag ej2π B mK ej2π B (1+mK ) ··· =
K
1fq
  i=0 1f

e j2π B (K −1+mK ) , (11) · c∗m h∗ c∗m s∗i e−j2π B (i+mK ) + c∗m v∗m,i
1f
= |h|2 e−j2π B K
+ wm , 1 ≤ m ≤ M − 1, (12)
and 1fq is the qth (0 ≤ q ≤ H − 1) hypothesis frequency
offset. This variant of matched-filter detector effectively where
reduces the frequency error range by a factor of H , which K −1
1 X
helps improve the frequency error immunity of the detector. wm , wm,i , (13)
K
The only downside of this detector is the linear increase in i=0
detection complexity by a factor of H . and
1f
(i+(m−1)K ) ∗ ∗
B. DIFFERENTIAL DETECTION wm,i = hsi ej2π B cm vm,i
1f
Differentiation is a well-known technique for removing the (i+mK )
+ h∗ s∗i e−j2π B cm−1 vm−1,i
phase component from the received signal, making it a viable
+ cm−1 c∗m vm−1,i v∗m,i . (14)
technique for mIoT systems. The design of the repeated NPSS
structure facilitates the use of such a differential detector to  v m−1,i and vm,i in (14) are mutually independent,
Since
mitigate the effect of large frequency offset [21]. Instead of E wm,i = 0, and
performing correlation between the received signal and the

Var wm,i
local copy of the NPSS waveform, the receiver performs n 1f
o
piecewise correlation between two adjacent received sig- = |h|2 Var si ej2π B (i+(m−1)K ) c∗m v∗m,i
nals (after the cover code is removed), each of K samples (the 1f
n o
+ |h|2 Var s∗i e−j2π B (i+mK ) cm−1 vm−1,i
length of the base waveform). The output of the correlator
+ Var cm−1 c∗m vm−1,i v∗m,i

reaches a maximum when the correlation is time-aligned
with the incoming NPSS, at which the correlation becomes = |h|2 Var v∗m,i + |h|2 Var vm−1,i + Var vm−1,i
  
between two received consecutive base waveforms or base
· Var v∗m,i

signals. The output of the differential correlator between the
(m − 1)th received base signal and the mth received base = 2 |h|2 σ 2 + σ 4 . (15)

13184 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

The total output can be expressed as For sufficiently large

 K (M − 1), each sum isGaussian with
distribution CN 0, 2K (M1 −1) 2 |h|2 σ 2 + σ 4 , according to


M −1
1 X
the central limit theorem. Since ω is the sum of two Gaus-
γ DIF = φm
M −1 sians, it is also Gaussian, with zero mean and variance of
m=1
2 −j2π 1f 1
B K + ω,
 
= |h| e (16) σDIF
2
= 2 |h|2 σ 2 + σ 4 . (22)
K (M − 1)
where The detection SNR of the differential detector (16) is thus
M −1
2
1 X
|h|2 K (M − 1)
ω= wm . (17) ζDIF (η) =
= 2 , (23)
M −1
K (M −1) 2 |h| σ + σ
1 2 2 4 1
m=1 η + η2
The differential operation which is the correlation of the
 adja- which is only a function of the device geometry (not a func-
cent two base signals in (12) results in an output γ DIF  = |h|2 tion of the frequency offset).
that is not a function of 1f , making the detection free of The ‘‘noise amplification’’ factor as a result of the differ-
the effect of frequency offset, whereas the matched-filter ential operation is then
detector in (7) suffers from a detection energy loss due to the  −1
frequency offset. M −1 1
ζDIF (η) M 2 + η
It seems that the differential detector solves the frequency = 
ζMF (η, 1f )
2
1f

offset problem. However, it has a new issue as is discussed in 
sinc B/MK 

detail in the following subsection.
ϕ(η)
=
ρ(1f )
C. DIFFERENTIAL DETECTION VS MATCHED-FILTER
= ϕ(η) − ρ(1f ) (dB), (24)
DETECTION
Now we look into the issue of the differential detector, i.e., where
the impact of the differentiation on the detection SNR of the M −1

1 −1

differential correlator. ϕ(η) , 2+ (25)
M η
First, the corresponding detection SNR of the matched-
filter detector whose correlator is matched to the original is a factor related to the device geometry η, referred to as the
transmitted signal, as shown in Fig. 4, can be represented as geometry factor, and
2
1f
2  
1f
 
ρ(1f ) , sinc
 
|h|2 sinc B/MK (26)
  
B/MK 

ζMF (η, 1f ) =
σ /MK
2
is the frequency offset factor, a function of the frequency
2
1f
 
offset. They are plotted in Fig. 6 (a) and Fig. 6 (b). The
= MK η sinc  ,
 
(18)
B/MK  geometry factor represents the degradation in detection SNR
due to the noise amplification effect of differentiation whose
when the synchronization signal is present, and in
degree of attenuation is determined by the device’s geometry
alignment with the local NPSS sequence, i.e., x =
η, whereas the frequency offset factor also represents the
xm,i , 0 ≤ m ≤ M − 1, 0 ≤ i ≤ K − 1 .

detection SNR degradation, but due to the reduction of the
Next, we look at the effect of differentiation on the output
detection energy as a result of the frequency offset between
SNR of the correlator of the differential detector. The noise
the device and the network.
at the output of the detector (16) can be further expressed as
Clearly, for the frequency offset factor, we have
M −1 M−1
1 2 K
X X −1
1 2 K
X X −1
ρ(1f ) ≤ 1 (0 dB), (27)
ω= w2a−1,i + w2a,i .
K (M − 1) K (M −1) where equality holds when 1f = 0, and for the geometry
a=1 i=0 a=1 i=0
(19) factor
1 M −1 1
The terms in each nof these twoosummations of (19) are ϕ(η) < · < (−3dB). (28)
2 M 2
independent since E wm1 ,i1 w∗m2 ,i2 = 0 (m1 6 = m2 ), for
That is, the output SNR of the differential detector is at least
 
M −1
 3 dB lower than the matched-filter detector at low frequency
m1 , m2 ∈ N1 , 2a a = 1, 2, · · · ,

(20) offset (ρ(1f ) = 0dB), reflecting the noise amplification
2
effect of differentiation. In the high geometry scenario, i.e.,
or η  5dB,
  
M −1 1 M −1 1
m1 , m2 ∈ N2 , 2a − 1 a = 1, 2, · · · , . (21)

 lim ϕ (η) = · < (−3 dB), (29)
2 η→∞ 2 M 2
VOLUME 5, 2017 13185
 W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 6. (a) Plot of geometry factor ϕ (η) in (25); and (b) plot of frequency offset factor ρ (1f ) in (26).

whereas in the low geometry scenario, i.e., for the NB-IoT

device with excessive penetration loss (e.g., deep inside a
building, i.e., the basement) where η  5 dB, we have
M −1
lim ϕ (η) = · η ≈ η, (30)
η→0 M
which is close to the device’s geometry. Since ϕ (η) is a
monotonically-increasing function of geometry η, the lower
the geometry of the device is, the larger the attenuation to
the differential detector SNR. For a device at η = −19 dB,
the noise is boosted by a factor of ∼19 dB (comparing to the
matched-filter detector in the absence of frequency offset).
It becomes clear that, in general, the differential detector
has no benefit when the frequency offset is small such that
the geometry factor prevails over the frequency offset factor,
ϕ (η) < ρ(1f ), (31)
i.e., the SNR attenuation caused by the noise amplification
at geometry η is larger than the attenuation caused by the
frequency offset (noting that a smaller value means a larger
attenuation).
ϕ (η)
From (28), (31) holds as long as FIGURE 7. Plot of ρ (1f ) in (24) as a function of the geometry η and
frequency offset 1f .
1 M −1
ρ(1f ) > · , (32)
2 M
i.e., the degradation due to frequency offset is less than pre-encoding requires the receiver to perform a differential
3 dB or equivalently |1f | < 0.65kHz (cf. Fig. 6 (b)), decoding in order to preserve the same autocorrelation prop-
i.e., 0.65 ppm for 1-GHz carrier frequency, meaning that erty of the original waveform, which leaves the matched-filter
the geometry factor or the ‘‘noise amplification’’ effect is detection out of the option.
the dominant factor as long as |1f | < 0.65kHz, as shown Finally, as a final note of this section, it is worth noting
in Fig. 6 as well as Fig. 7. In other words, differential detector that in the practical detection, due to the limited sampling
is not beneficial for frequency offsets less than 0.65 kHz, resolution, there exists a timing offset between the sampling
regardless of the device geometry η. However, when |1f | > point and the maximum correlation output time, which incurs
0.65kHz, differentiation can be beneficial depending on the additional loss to the detection energy. Fig. 8 plots the com-
device geometry, η. bined detection SNR loss due to the timing and frequency
There is a variant of the differential detection technique offset. It is observed that the matched-filter detector is more
described in [24], in which the original waveform is pre- sensitive to not only the frequency offset but also the timing
differentially-encoded sample-wise at the transmitter. This offset than the differential detector.

13186 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 8. Plot of the combined detection SNR loss due to the timing and
frequency offset, where Ts is the Nyquist sampling interval.

IV. DIFFERENTIAL AND MATCHED-FILTER

DETECTION REGIONS FIGURE 9. Plot of η, 1f ϕ (η) = ρ 1f



in (36) that divides the NB-IoT
From the analysis in Section III, we conclude that (1) The device operating status into two regions according to the dominance of
the two factors: the region <MF where ϕ (η) < ρ (1f ), i.e., the geometry
performance of the matched-filter detector and differential factor prevails, so does the matched-filter detector ; and the region <DIF
detector depends on the device operating status, i.e., (η, 1f ), where ϕ (η) > ρ (1f ), i.e., the frequency-offset factor dominates, and the
differential detector is superior.
the oscillator accuracy and the geometry of the device in the
cell; and
(2) The NPSS has the repetition structure which is facili-
tated to perform differential detection at the receiver, thereby However, there is a substantial NB-IoT use cases in
providing the freedom in choice of a matched-filter detec- which devices are deployed deep inside a building (e.g.,
tor or a differential detector according to the device operating the basement), which incurs an additional penetration loss
status. of 20 dB.
A natural question is then: What is the switching To get an idea of a device’s geometry in a typi-
point or operating region that a differential detector outper- cal mIoT system, assuming the transmit power allocated
forms a matched-filter detector? In this section, we use the to the signal is p (dBm), and the maximum coupling
analysis results from the previous section to derive the differ- loss is 1 (dB), the required receiver sensitivity is then
ential detection region and matched-filter detection region. p − 1 (dBm). If the bandwidth that the signal occupies is
Specifically, we look for B (dB-Hz) and the noise figure of the receiver is δ (dB),
the noise power at the receiver is thus B + σ 2 + δ (dBm),
<DIF , {(η, 1f ) |ϕ (η) > ρ (1f ) } (33) where σ 2 (dBm/Hz) is the noise power spectral den-
sity (i.e., −174 dBm/Hz). The corresponding geometry at the
and
receiver is
<MF , {(η, 1f ) |ϕ (η) < ρ (1f ) } (34)  
η = (p − 1) − B + σ 2 + δ (dB) . (35)
where <DIF is the region where the frequency offset factor
dominates (i.e., ϕ (η) > ρ (1f )), and hence the operating For an NB-IoT system, the transmitter (base station) power
region for a differential detector, and <MF where the geome- p is assumed to be 35 dBm (for in-band and guard-band
try factor prevails (i.e., ϕ (η) < ρ (1f )), and hence the region mode) and 43 dBm (stand-alone mode on the re-farmed GSM
for a matched-filter detector. But first let us take a look at the spectrum) [7]. We also assume that the maximum coupling
geometry (η) properties in NB-IoT. loss is 144 dB (for a regular LTE deployment) plus 20 dB
additional penetration loss in the mIoT environment (i.e.,
A. ANALYSIS OF DEVICE GEOMETRY IN NB-IoT 1 = 144dB + 20dB = 164dB), the geometry η defined
As earlier mentioned, the coupling loss of a device depends in (35) is −13dB for in-band and guard-band mode, and
on the position in the cell, where the position farther away −5dB for standalone mode.
from the center (i.e., the low geometry) experiences higher Since 20 dB is typically the highest geometry in a cellular
coupling loss, and ergo low SNR. The maximum cou- macro cell, it is reasonable to assume that the geometry η for
pling loss for a traditional cellular system is ∼140 dB. NB-IoT falls into the range from −13dB to 20 dB.

VOLUME 5, 2017 13187

 W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 10. Plot of detection SNR ζ against geometry η at different frequency offset 1f : (a) for both matched-filter and differential detectors; and (b) for
a selective detector which performs a selection between the two detectors according to the operating regions in Fig. 9.

B. OPERATING REGIONS larger than 0.9 kHz and geometry −5dB or higher. Similarly,
Based on the above analysis result, particularly the geom- (3 dB, 0.7 kHz) indicates a frequency offset of at least 0.7 kHz
etry factor and the frequency-offset factor that govern the for devices at geometry 3 dB or higher to benefit. In general,
detection performance, we are finally ready to derive the high geometry and/or high frequency offset favor differential
operation regions for a differential detector and a matched- detection.
filter detector. Fig. 10 plots the detection SNR ζ against geometry η at
Fig. 9 plots the ‘‘boundary’’ that separates the opera- different frequency offset 1f , for (a) a matched-filter detector
tion regions of the differential and matched-filter detectors, and a differential detector, and (b) a selective detector which
i.e., the set of device geometry and frequency offset pairs at applies a selection between the above two detectors according
which these two schemes share the same output SNRs, to the operating regions in Fig. 9.
{(η, 1f ) |ϕ (η) = ρ (1f ) } , (36)
C. IMPLEMENTATION EXAMPLE
with η in the range from −20 dB to 20 dB.
A straightforward implication from the above results is
From the previous analysis, we know for a fact that
that for a high-cost device (like the LTE UE for human-
1 M −1 type communications or HTC) equipped with a temperature-
ϕ (η) < · , ∀η. (37)
2 M compensated high-accuracy crystal, the geometry factor is
As such, the matched-filter detector is guaranteed to outper- typically a dominant factor, and a matched-filter detector is
form the differential detector as long as ρ (1f ) > 12 · MM−1 . undoubtedly a first choice (which is in fact the case in HTC).
It is thus expected to see that, For an NB-IoT device, neither the geometry factor nor the
frequency-offset factor dominates in all scenarios, so nei-
ϕ (η) < ρ (1f ) , ∀η, |1f | < 0.65kHz, (38)
ther matched-filter nor differential detector alone is the best
meaning that the matched-filter detector is superior at any solution. An NB-IoT device therefore needs the freedom in
device geometry for frequency offset less than 0.65 kHz. selecting the detector according to the frequency uncertainty
Hence, for a device with a quality oscillator or after frequency and the geometry that it experiences, as shown in Fig. 11 (a).
error has been compensated, the conventional matched-filter For devices with low crystal quality (large frequency uncer-
detector is more advantageous. tainty), the differential detector is a safer choice in the initial
For higher frequency offsets (i.e., |1f |  0.65kHz), how- system acquisition.
ever, which factor prevails or which detector is advantageous Indeed, including both detectors at the receiver will
depends on the device geometry. For example, the switch increase the complexity. However, this increase can be min-
point at (−5dB, 0.9 kHz) in Fig. 9 indicates that the differen- imized as both the detectors share one correlator, the local
tial detector is beneficial for devices with a frequency offset waveform/signal of which depends on the choice of the

13188 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

FIGURE 12. NPSS detection probabilities of a matched-filter detector,

a differential detector, and a selective detector at various geometries,
assuming the frequency offset of the device is uniformly distributed
between −1.2 kHz and 1.2 kHz.

then select the best detection scheme based on this statistic

and its current geometry. Fig. 12 shows the performance of
the selective detector under various geometries in comparison
with the matched-filter and differential detectors.

V. CONCLUSION
In this paper, we study the detection performance of two
detection techniques, i.e., a traditional matched-filter based
detection and a differential detection. We introduce two fac-
tors: the geometry factor and the frequency offset factor.
We show that, in general, the NPSS detection performance
is governed by these two factors. That is, the performance of
these two types of detectors depends on (1) the device geom-
FIGURE 11. Schematic diagram to illustrate the implementation example
of the selective NPSS detection based on the system acquisition etry; and (2) its frequency offset from the network. When the
procedure in Fig. 2. geometry factor prevails, the matched-filter detector outper-
forms the differential detector; and when the frequency-offset
factor dominates, the differential detector becomes advanta-
geous. This indicates that the matched-filter technique (with a
detector. For the matched-filter detector, it is the NPSS wave-
single template waveform) is no longer the optimal technique
form, whereas for the differential detector, it is part of the
in the presence of a frequency offset between the transmitter
received signal.
and a receiver. To this end, a device requires having both
Although the frequency offset is difficult, if not impossi-
techniques at its disposal, and the flexibility of selecting the
ble, to be completely corrected since different crystals have
right one depending on the actual operating conditions, i.e.,
different stabilities and sensitivities to ambient temperature
the dominance of the two factors. This has the obvious advan-
and other factors, the frequency uncertainty is converging
tage of enabling the receiving device to change the detector
to a certain degree (e.g., 1 ppm or 1 kHz at 1-GHz carrier
as needed to increase its tolerance to frequency offset (due
frequency) each time when a device successfully acquires
to cost) and degraded SNR (due to coverage extension in
the system. In addition to making a frequency correction to
NB-IoT). Specifically, based on the geometry and frequency
the local crystal based on the acquired system frequency,
offset factors, the operating conditions are divided into sep-
the device collects and monitors the statistics on the crys-
arate regions in each of which one detector works better
tal, e.g., the deviation of the crystal frequency, σ1f ,
q than the other, such that the device can choose the detector
E |1f |2 . It is updated after each successful acquisition:


accordingly and the immunity to both frequency offset and
σ1f = α · 1f + (1 − α) · σ1f 0 , where 1f is the frequency
low signal quality is maximized, a challenging feat with
offset/error detected from the current acquisition, σ1f 0 is the low-cost crystals and extended coverage in NB-IoT. Finally,
deviation estimated from the previous acquisition, and 0 < in the current study, the rectangular pulse is assumed for
α < 1 is a constant, as shown in Fig. 11 (b). A device can OFDM symbol. Different pulse shapes may affect detection

VOLUME 5, 2017 13189

 W. Yang et al.: Enhanced System Acquisition for NB-IoT

performance but is expected to be secondary to the effect of [22] C. R. Sheu and C. C. Huang, ‘‘A differential sliding correlation scheme for
frequency offset. However, it is an interesting and valid topic symbol timing detection in time domain synchronous OFDM systems,’’ in
Proc. IEEE 69th Veh. Technol. Conf. VTC Spring, Jun. 2009, pp. 1–5.
for future study. [23] H. Kroll, M. Korb, B. Weber, S. Willi, and Q. Huang, ‘‘Maximum-
likelihhood detection for energy-efficient timing acquisition in NB-IoT,’’
in Proc. IEEE Wireless Commun. Netw. Conf. Workshops (WCNCW),
ACKNOWLEDGEMENT
Mar. 2017, pp. 1–5.
The authors would like to thank the editor and the reviewers [24] 3rd Generation Partnership Project Technical Specification Group
for their excellent comments that have greatly improved the GSM/EDGE Radio Access Network Cellular System Support for Ultra Low
Complexity and Low Throughput Internet of Things (Release 13) ver.1.0.1,
presentation and the technicality of this paper. document TR 45.820, Mar. 2015.

REFERENCES
[1] H. Chaouchi, The Internet of Things: Connecting Objects to the Web.
Hoboken, NJ, USA: Wiley, 2010.
[2] W. Yang et al., ‘‘Narrowband wireless access for low-power massive
Internet of Things: A bandwidth perspective,’’ IEEE Wireless Commun., WENJIE YANG is currently pursuing the
vol. 24, no. 3, pp. 138–145, Jun. 2017. Ph.D. degree with the Nanjing University of Sci-
[3] ‘‘Sigfox—One network a billion dreams,’’ Sigfox, Labège, San Ramon, ence and Technology, Nanjing, China. He is with
France, Whitepaper. the Wireless Networking and Mobile Communi-
[4] N. Sornin, LoRaWAN Specfication Ver. 1. 0, LoRa Alliance, Tech. Rep., cations Group, School of Electronic and Optical
Jan. 2015. Engineering. His current research interests are in
[5] 3rd Generation Partnership Project Technical Specification Group Radio the areas of wireless communications and signal
Access Network Evolved Universal Terrestrial Radio Access (E-UTRA) processing.
Physical channels and modulation (Release 13) Ver. 13.2.0, document TS
36.211, Jun. 2016.
[6] Y. P. E. Wang et al., ‘‘A primer on 3GPP narrowband Internet of Things,’’
IEEE Commun. Mag., vol. 55, no. 3, pp. 117–123, Mar. 2017.
[7] R. Ratasuk, B. Vejlgaard, N. Mangalvedhe, and A. Ghosh, ‘‘NB-IoT
system for M2M communication,’’ in Proc. IEEE Wireless Commun. Netw.
Conf. (WCNC), Doha, Qatar, Apr. 2016, pp. 1–5.
MIN HUA is currently pursuing the Ph.D. degree
[8] N. Mangalvedhe, R. Ratasuk, and A. Ghosh, ‘‘NB-IoT deployment study
for low power wide area cellular IoT,’’ Proc. IEEE 27th Annu. Int. Symp. with the Nanjing University of Science and Tech-
Pers., Indoor, Mobile Radio Commun. (PIMRC), Sep. 2016, pp. 1–6. nology, Nanjing, China. She is with the Wireless
[9] C. Bockelmann et al., ‘‘Massive machine-type communications in 5G: Networking and Mobile Communications Group,
Physical and MAC-layer solutions,’’ IEEE Commun. Mag., vol. 54, no. 9, School of Electronic and Optical Engineering. Her
pp. 59–65, Sep. 2016. current research interests are in the areas of wire-
[10] A. A. Zaidi et al., ‘‘Waveform and numerology to support 5G services less communications and signal processing.
and requirements,’’ IEEE Commun. Mag., vol. 54, no. 41, pp. 90–98,
Nov. 2016.
[11] Z. Dawy, W. Saad, A. Ghosh, J. G. Andrews, and E. Yaacoub, ‘‘Toward
Massive Machine Type Cellular Communications,’’ IEEE Wireless
Commun., vol. 24, no. 1, pp. 120–128, Feb. 2017.
[12] C. Johnson, Long Term Evolution in Bullets. Northampton, U.K.: Creates-
pace Press, 2012.
[13] M. Wang, A. Agrawal, A. Khandeka, and S. Aedududla, ‘‘Preamble JINGJING ZHANG is currently pursuing the
design, system acquisition, and determination in modern OFDMA cellular Ph.D. degree with the Nanjing University of Sci-
communications: An overview,’’ IEEE Commun. Mag., vol. 49, no. 7,
ence and Technology, Nanjing, China. She is with
pp. 164–175, Jul. 2011.
the Wireless Networking and Mobile Communi-
[14] S. Ahmadi, LTE-Advanced: A Practical Systems Approach to Under-
cations Group, School of Electronic and Optical
standing 3GPP LE Releases 10 and 11 Radio Access Technologies.
San Francisco, CA, USA: Academic, 2014.
Engineering. Her current research interests are in
the areas of wireless communications and signal
[15] M. Hua et al., ‘‘Analysis of the frequency offset effect on random
access signals,’’ IEEE Trans. Commun., vol. 61, no. 11, pp. 4728–4740, processing.
Nov. 2013.
[16] M. Hua, M. Wang, K. W. Yang, and K. J. Zou, ‘‘Analysis of frequency
offset effect on Zadoff-Chu sequence timing performance,’’ IEEE Trans.
Commun., vol. 62, no. 11, pp. 4024–4039, Nov. 2014.
[17] S. Beyme and C. Leung, ‘‘Efficient computation of DFT of Zadoff-Chu
sequences,’’ Electron. Lett., vol. 45, no. 9, pp. 461–463, Apr. 2009.
[18] F. Tufvsson, O. Edfors, and M. Faulkner, ‘‘Time and frequency synchro-
nization for OFDM using PN-sequence preambles,’’ in Proc. IEEE Veh. TINGTING XIA is currently pursuing the
Technol. Conf., vol. 4. Sep. 1999, pp. 2203–2207. Ph.D. degree with the Nanjing University of Sci-
[19] J. Wang, Z. X. Yang, C. Y. Pan, M. Han, and L. Yang, ‘‘A combined code ence and Technology, Nanjing, China. She is with
acquisition and symbol timing recovery method for TDSOFDM,’’ IEEE the Wireless Networking and Mobile Communi-
Trans. Broadcast., vol. 49, no. 3, pp. 304–308, Sep. 2003. cations Group, School of Electronic and Optical
[20] H. Ni, G. Ren, and Y. Chang, ‘‘Complexity effective cell search scheme Engineering. Her current research interests are in
for OFDM cellular system,’’ in Proc. IEEE Int. Conf. Commun. Syst.,
the areas of wireless communications and signal
Singapore, Nov. 2010, pp. 700–704.
processing.
[21] ‘‘NB-PSS and NB-SSS design 3GPP TSG RAN WG1 Ad-Hoc meeting,’’
Qualcomm Incorporated, San Diego, CA, USA, Tech. Rep. R1-161981,
Mar. 2016.

13190 VOLUME 5, 2017

W. Yang et al.: Enhanced System Acquisition for NB-IoT

JUN ZOU received the Ph.D. degree with the MAO WANG received the Ph.D. degree in elec-
Wireless Networking and Mobile Communica- trical engineering and computer science from the
tions Group, School of Electronic and Optical University of Kentucky, Lexington, KY, USA. He
Engineering, Nanjing University of Science and is currently a Professor with the National Mobile
Technology, Nanjing, China. Since 2016, he has Communications Research Laboratory, Southeast
been with the Wireless Networking and Mobile University, Nanjing, China. He is also an Adjunct
Communications Group, Nanjing University of Professor with the School of Electronic and Opti-
Science and Technology, where he is currently a cal Engineering, Nanjing University of Science
Lecturer at the School of Electronic and Optical and Technology, where he is the Director of the
Engineering. He has over ten IEEE journal publi- Wireless Networking and Mobile Communica-
cations. His research interests are in the areas of wireless communications, tions Group. He holds over 90 U.S. patents and has over 30 IEEE journal
signal processing, and Internet of Things. publications. His research interests include communication theory and wire-
less networking.

CHENGLING JIANG received the Ph.D. degree

from the School of Electronic and Optical Engi-
neering, Nanjing University of Science and Tech-
nology, Nanjing, China. Since 2013, he has
been with the Information and Telecommunication
Branch, State Grid Jiangsu Electric Power Com-
pany. His research interests are in the areas of
private wireless communications and networking
for power networks.

VOLUME 5, 2017 13191