See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/311980840

Chirp spread spectrum as a modulation technique for long range

communication

Conference Paper · November 2016

DOI: 10.1109/SCVT.2016.7797659

CITATIONS READS

213 5,812

2 authors:

Brecht Reynders S. Pollin

KU Leuven KU Leuven
14 PUBLICATIONS   936 CITATIONS    427 PUBLICATIONS   8,040 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Comité d'experts sur les radiations non ionisantes (IBGE) View project

OmniDrone: 720° perception for safe and autonomous systems View project

All content following this page was uploaded by S. Pollin on 17 June 2022.

The user has requested enhancement of the downloaded file.

 Chirp Spread Spectrum as a Modulation Technique
for Long Range Communication
Brecht Reynders, Sofie Pollin
Department of Electrical Engineering
KU Leuven
Heverlee, B-3001 Belgium
Email: brecht.reynders@esat.kuleuven.be, sofie.pollin@esat.kuleuven.be

Abstract—Long range low power is a family of technologies in SigFox [6]. More specifically, the coverage and the coexis-
promising to connect thousands of sensors to the future internet tence with other competing signals in the same band will be
of things. Within this family of possible technology choices, taken in consideration. Coexistence is an important parameter
two different branches have emerged: the standards based on
spreaded wideband communication and the standards based on because of the unlicensed bands these networks typically use.
narrowband communication, both promising to reach long range The remainder of this paper is structured as follows. The
connectivity at very low power. More specifically, this paper first section summarizes the real world model we used in our
focuses on chirp spread spectrum (CSS). This paper presents a simulations. The next section gives a detailed overview of the
detailed model of CSS, showing that the symbols are not perfectly technology and the model we implemented in Matlab. Section
orthogonal. The results show the potential communication ranges
are close to narrowband networks and the robustness against IV shows the results of our simulations and model. And finally,
interfering signals is 22 dB better for spreading factor 10 this paper ends with a small discussion about higher layers and
than BPSK. Although wideband communication is robust to the conclusion.
interference, in long range communication, this robustness is
insufficient due to the long range and the larger footprint. The II. M ODEL
large propagation losses due to the long range and the larger
footprint make CSS prone to collisions with other noise sources, For our range simulations, some assumption have been
possibly larger than the coding gain. made. First, the noise is assumed to be just thermal noise.
I. I NTRODUCTION The equation for the noise power Pn is shown in Eq. 1
with T the temperature in Kelvin (25o C), k the constant of
Long range communication or low power wide area net- Boltzmann and B the bandwidth. This noise power is the
works (LPWANs) are networks where high data rate require- lower bound, as there will be other ambient noise sources in
ments are traded for range and energy improvements. In these the ISM-band.
networks, data rates are very low, leaving hardly a few bytes
per seconds, but enabling ranges of a few kilometers.
In our previous paper [1], we have compared two such
Pn = kT B (1)
LPWANs: an ultra narrowband, Sigfox-like standard with a
wideband, LoRa-like technology. This paper elaborates more Second, the signal power decays with distance following
on our detailed model of the physical layer of CSS and prove the Hata model [7]. In this comparison, the model for urban
that the spreading factors are not perfectly orthogonal. areas is used as the highest number of devices will occur in
This paper is definitely not the first paper to evaluate the these areas. The formula for losses in urban environment L
LoRa standard. Many papers have been published. However, is shown in Eq. 2. In this formula, d is the distance between
most of them rely on measurements in the field to derive base station and device in meter, f the frequency (868MHz),
physical layer properties [2], [3]. The authors of these papers hB the height of the base station in meter (30m ), hM the
measured the reliability of the link at different distances form height of the mobile device (1m):
the base station for different spreading factors. In [4], the
authors combine the information from the datasheet of a L = 69.55 + 26.16log10 (f ) − 13.82log10 (hB )
LoRa transceiver to identify the capacity and range constraints −3.2(log10 (11.54hM ))2 (2)
in these networks. An analysis of orthogonal chirp spread −4.97 + [44.9 − 6.55log10 (hB )]log10 (d).
spectrum is given in [5], however, where orthogonal symbols
are achieved by relying on both up and down-chirps. A Third, only the uplink is considered, as the majority of the
detailed analysis of the CSS-symbols used LoRa with different data flows in this direction. The uplink data is modeled as
sources of noise and interference is missing. packets with length l and l equals 160 bits or 20 bytes. Finally,
This paper gives an extended insight in the performance the transmitter of the mobile device sends with a power of
of chirp spread spectrum. Whenever possible, CSS will be 25mW, which is the maximal allowed power in most of the
compared to narrowband communications like the BPSK used ISM-band.

978-1-5090-4361-3/16/$31.00 2016 European Union

 To investigate the coexistence, three different noise sources f
were used: additive white Gaussian noise (AWGN), a con-
tinuous wave and a CSS symbol with different spreading fmax B
factor. The continuous wave interferer has, for each simulated ts/2
symbol, a constant random frequency uniformly distributed
B
between fmin and fmax (the minimal and maximal frequency ts
of CSS) with a random phase offset, also uniformly distributed f0 B B
between 0 and 2π. Eq. 3 shows the equation for generating a fmin ts/2 ts

CW C(t). In this equation, a is the amplitude of the signal, t

is the time, f is the constant frequency and φ is the constant 0 ts/2 ts t
phase offset.
C(t) = aei(2πf t+φ) (3) Fig. 1: Graphical representation of the frequency in time for
two CSS symbols with spreading factor s and s/2.
Further, the transmitter and the receiver are assumed to be
perfectly synchronized to enable coherent detection.
III. C HIRP S PREAD S PECTRUM Proposition 1. Two CSS symbols are not guaranteed to be
Chirp spread spectrum is a technique known from the orthogonal.
physical layer of LoRaWAN [8]. The standard uses chirp
spread spectrum (CSS) as modulation to transmit over large Proof. Assume two different CSS symbols s1 (t) and s2 (t) are
distances in the 868MHz band. LoRaWAN includes a scal- orthogonal to each other, then Eq. 7 should hold:
able bandwidth (125kHz, 250kHz or 500kHz). Based on the Z ts
bandwidth and the SNR, a variable spreading factor is chosen 0= s∗1 (t)s2 (t)dt
to spread the signal to make it more robust to noise. The
Z0 ts
resulting bit rate will vary between 300bps and 37.5kbps.
The used parameters will be mentioned in the evaluation = e−i2πf1 (t)t ei2πf2 (t)t dt, (7)
0
section. The rest of this section introduces the CSS symbols
in detail, proves that they are not perfectly orthogonal and where
finally, describes the decoder functionality assumed for the (
B ki
performance evaluation. ts t + SB x < ti
fi (t) = B ki (8)
1) CSS symbols: A CSS symbol has four important param- ts t + SB −B x ≥ ti
eters: the spreading factor S, the minimal frequency fmin , the
maximal frequency fmax and the input bits. To generate a with i = 1, 2, and k1 and k2 the different integer input symbols
CSS symbol, a starting frequency f0 is needed. The starting from 0 to S − 1.
frequency, lying between fmin and fmax , represents the in-
put information of log2 (S) bits. For a spreading factor S 1 , S − ki
ts = ti , i = 1, 2, (9)
log2 (S) bits define f0 . Then, the length of the symbol can be S
calculated. The length of one symbol ts is shown in Eq. 4:
with t1 and t2 the times where respectively s1 (t) and s2 (t)
S S switch from fmax to fmin shown in Fig. 1. Assume that k1 >
ts = = , (4)
B fmax − fmin k2 such that 0 ≤ t1 < t2 ≤ ts , then Eq. 7 becomes:
where B is the bandwidth or the difference between fmin and Z t1 Z t2
fmax . From this formula, the bit rate RB (S) can be calculated: k1 −k2 k1 −k2
0= e−i2π S Bt
dt + e−i2π( S B−B)t dt
log2 (S) 0 t1
(10)
Rb (S) = (5) Z ts
k1 −k2
ts + e−i2π S Bt dt.
When the symbol length is defined, the signal constantly chirps t2
over all frequencies starting at f0 to fmax and continuing from
fmin to f0 . The mathematical representation is given in Eq. 6. Solving the integrals in Eq. 10 results in Eq. 11:
The frequency in function of time f (t) is also shown in Fig. k1 −k2 k1 −k2
1. i[e−i2π S B0
− e−i2π S Bt1
]
0= k1 −k2
S B
s(t) = ei2πf (t)t . (6) k1 −k2 k1 −k2
i[e−i2π( S B−B)t1
− e−i2π( S B−B)t2
]
+ k1 −k2
(11)
S B−B
1 Itshould be noted that the spreading factor is defined here as the amount
of chips per symbol. This is different from the spreading factor used in the k1 −k2 k1 −k2

LoRaWAN specification or used in other papers. The value there equals the i[e−i2π S Bt2
− e−i2π S Bts
]
+ k1 −k2
SpreadingFactor register value. For more information, [9] Sec. 4.1.1.2
S B
Because k1 and k2 are not equal to each other and the of this fit is shown for a CSS symbol with 10 bits in Fig. 3.
difference can never be equal to S, we can rewrite the Also the BER curve for a BPSK symbol with 1 bit per symbol
denominator. Including Eq. 9 in Eq. 11, gives: (Eq. 17) [11] and an FSK symbol with k equal to 10 bits per
k1 − k2 k1 −k2 S−k1
symbol (Eq. 18) [12] are shown. To convert SNR to energy per
0=( B − B)[1 − e−i2π S Bts S bit to noise density (Eb /N0 ), Eq. 20 was used, where Ps /Pn
S
k1 −k2 S−k2 k1 −k2 is the signal-to-noise ratio, B is the bandwidth and Rb (S) is
+ e−i2π S Bts S − e−i2π S Bts ] the bit rate.
(12) r
k1 − k2 k1 −k2 S−k1 Eb
+( B)[e−i2π( S B−B)ts S Pe,BP SK = Q( 2 ) (17)
S N0
k1 −k2 S−k2
− e−i2π( S B−B)ts S ].
Z ∞ Z y
1 1 2 k
Pe,F SK = √ [1 − ( √ exp−x /2 dx)2 −1 ]
Combining Eq. 12 and Eq. 4, we show that 8π −∞ 2π −∞
q
2kEb 2
− 21 (y− )
k1 − k2 k1 −k2 exp N
0 dy
0=( B − B)[1 − e−i2π S (S−k1 )
S (18)
k1 −k2
+ e−i2π S (S−k2 ) − e−i2π(k1 −k2 ) ] This can be simplified as Eq. 19, proposed in [13].
(13)
k1 − k2 k1 −k2
2k
r
+( B)[e−i2π( S −1)(S−k1 ) Pe,F SK = Q( k
Eb
) (19)
S 4 N0
k1 −k2
− e−i2π( S −1)(S−k2 ) ]. Eb Ps B
= (20)
Because ei2πk with k integer is always 1, Eq. 13 yields: N0 Pn Rb (S)
Eb √
r
1
k1 − k2 k1 −k2 k1 −k2 Pe,CSS = Q(1.28 k − k1.28 + 0.4) (21)
0=( B − B)[−ei2π S k1 + ei2π S k2 ] 2 N0
S (14)
k1 − k2 k1 −k2 k1 −k2
It can be seen that both 10-bit FSK and 10-bit CSS have a
+( B)[ei2π S k1 − ei2π S k2 ] BER that decreases faster with increasing SNR compared to
S
BPSK. In FSK or any orthogonal constellation, it is possible
k1 −k2 k1 −k2 to project each received symbol with noise to the signal space
0 = ei2π S k1
− ei2π S k2
(15) spanned by the 2k orthogonal symbols. As it turns out, the
k −k (k −k )2
i2π 1 S 2 k2 i2π 1 S 2 received symbol will align perfectly in one direction (the
=e (e − 1) (16)
direction of the corresponding symbol), while the noise will
Eq. 16 is only true if and only if the difference between k1 align with all the vectors in all directions. To decode, the
and k2 equals a multiple of the square root of the spreading maximal direction is selected. After mapping any 2 different
factor. orthogonal symbols in that 2k dimensional space, we can
2) Decoding CSS symbols: Because the symbols are close observe that their distance only depends on their energy Es .
to orthogonal, a decoder was designed that uses correlation to A higher Es means they are further away and hence a lower
decode the signal. Every received symbol in the decoder is probability of symbol and hence bit errors. In the case of
correlated with the base CSS symbols. The decoder makes a antipodal symbols, it should be noticed that their signal space
decision based on the maximum correlation. has only 2k−1 spanning symbols. So, some of the symbols
Another possible decoder is presented in [10]. The authors lie on the same vector but with opposite sign. The distance
in [10] propose to mix and resample the incoming symbol to between 2 symbols on the same vector is however larger. This
create a clear shifted narrow peak in the frequency domain. is visualized for the case of 1 bit or 2 symbols in Fig. 2.
Both decoder implementations however give the same perfor- From this figure, it should be clear that BFSK performs worse
mance with AWG noise. than BPSK as the distance between symbols is smaller. When
orthogonal symbols are considered with more bits per symbol,
IV. S IMULATION R ESULTS the distance between the symbols determines the symbol error
This section shows the results of the implementation of our rate (SER). The energy per symbol Es depends linearly on the
model to show the potential range and coexistence capabilities. energy per bit Eb and the amount of bits per symbol k, shown
in Eq. 22. So, with a constant energy per bit, the symbol energy
A. Bit error rate will scale linearly with k, resulting in a larger Es and lower
Before estimating the potential range for a CSS based SER when k increases. As a result, the BER is lower too, as
network, a closed-form equation for the bit error rate (BER) the average BER is 50% of the SER. This reasoning is only
of CSS symbols with varying spreading factors is needed. valid if the SNR is high enough and the symbol error rate is
For this, 10 000 symbols were generated for each SNR and small. When the SNR is small, an increase in symbol energy
decoded with the decoder proposed in III-2. The resulting will not lead to a lower SER. For low SNR, more bits just
points of this simulation were put in the curve fitting toolbox means more symbols and hence, more different values to pick
of MATLAB to estimate the BER curve for CSS. The results from and to fail.
 Probability of bit error
Es = kEb (22) 10−1

1 BPSK

p (2
10−4 10-bit CSS
10-bit FSK
)E
s 2Es BFSK
0 0 1 10−7
−5 0 5 10
Eb /N0

Fig. 3: Bit error rate of a BPSK, CSS symbol with bandwidth

(a) BFSK or 2 or- (b) BPSK or 2 an- 125kHz with 10 bits and FSK symbol with 10 bits. For lower
thogonal symbol tipodal symbol values of Eb /N0 , BPSK gives better results, but for higher
values, CSS is better.
Fig. 2: Difference in distance between 2 different symbols
(0 and 1), when mapped to the signal space. It is clear that

Throughput (in bits/s)

orthogonal symbols are closer together than antipodal symbols.
If more bits are added, the distance between the orthogonal
symbols stays the same. In PSK however, the symbols will be
closer together. 102

We can observe this behavior also in the equation for FSK BPSK
(Eq. 18). The latter part in the equation is the correlation CSS (12 bits)
of the correct wanted signal with noise, and this correlation
101
needs to be compared to the correlations with the other 2k − 1 0 2 4 6 8 10
symbols shown in the first part of the equation. The equation Distance (km)
shows that when Eb /N0 is high, the p first part goes to 0 in the
significant region where y is around 2kEb /N0 and the SER Fig. 4: Throughput of BPSK (sigfox-like) and CSS network
for FSK is almost equal for all k, resulting in an advantage for based on the range. The range of an UNB network is larger
symbols with more bits. When the SNR is low, more errors are than the CSS network.
observed because the probability is small that all the 2k − 1
different correlators give a lower value than the one of the
wanted symbol. In the case of BPSK, the distance between account the scalable data rate provided in LoraWAN, but only
the 2 symbols is larger, and hence, BPSK performs better for used the most reliable spreading factor. It should be noticed
small SNR. So, the bit error curve of BPSK and the bit error here, that for the sensor side, a CSS network is preferred, as
curve of FSK for k > 2 cross. the range is almost the same, but close to the base station, a
For CSS, a similar reasoning could be followed. But here, higher throughput can be achieved. The latter results in shorter
the non-orthogonal symbols need to be taken into account. A messages and less power consumption.
constant offset should be introduced in each of the correlators
depending on the symbol that was sent. To simplify the T = (1 − PP )RB (23)
equation, Eq. 21 was estimated in Matlab with the simulated
Pp = 1 − (1 − Pe )l (24)
data. The simulated bit curves show a square root over energy
per bit and contain a constant part that is only function of
the number of bits inside each symbol. The latter term is the C. Coexistence
result of the non-orthogonality proven in Sec. 1. Figure 5 shows the robustness of CSS (Eq. 6) against a
continuous wave (Eq. 3) with a given SIR (defined as Eq. 25).
B. Range A wideband symbol as CSS needs a certain SIR to be able to
When taken into account the wider bandwidth and the spread the interferer over it’s bandwidth. For spreading factor
higher bit rate, CSS results in a smaller coverage area. This 10, that SINR equals −16dB. Higher spreading factors require
can be seen in figure 4. The figure shows the throughput of less and lower more SINR.
both physical layers based on the packet error. The equation In a second simulation, also shown in figure 5, a CSS-
of the throughput is given in Eq. 23. In this formula, Pp symbol is demodulated while another CSS-symbol with differ-
is the probability of a packet error and can be calculated ent spreading factor is present. For spread spectrum systems,
with Eq. 24 with Pe the probability of a bit error described the demodulator should be able to ignore symbols with dif-
in previous section. In this simulation, we did not take into ferent spreading factors. This can be verified by the figure
 in Eq. 26 and Eq. 27. These equations show that overall a
Pe = probability of bit error
BPSK network can send four times as many messages then a
CSS network. Adding more spreading factors to increase the
10−1 throughput is possible, but these small spreading factors are
very sensitive to noise and require a higher SINR.

10−2 CSS with CW 12

1 X Rb (S)
CSS with 12-bit CSS TCSS = [ ] ⇒ TCSS = 18packets/s (26)
2e i=7 l
CSS with AWGN
10−3 1 RS B
−40 −35 −30 −25 −20 −15 TU N B = ⇒ TU N B = 72packets/s (27)
4e l b
SIR VI. C ONCLUSION
Fig. 5: BER for CSS symbol and LoRa symbol with 10 bits This paper showed a detailed analysis of chirp spread
per symbol and bandwidth B = 500kHz. spectrum. The symbols in CSS are not perfectly orthogonal.
The simulation results show that it is suited for LPWAN, and
the communication range is close to the ranges narrowband
because both the curve of a continuous wave as the one from solutions can reach. But, there is an advantage of 22dB
another spreading factor are close. the CSS-demodulator sees in coexistence and robustness against interference compared
the other CSS symbol with a different spreading as modulated to BPSK for spreading factor 10. Although CSS is robust
noise, like the continuous wave. This feature allows to send against interference and has a scalable data rate, the time-
simultaneously to the same base station with good power bandwidth footprint is significantly higher than the footprint
control and different spreading factors. of the narrowband solution. This results that for the entire
network, it is preferred to have a narrowband solution.
Psignal
SIR = 10log10 ( ) = −20log(ai ) (25) R EFERENCES
Pinterf erer
[1] B. Reynders et al., “Range and coexistence analysis of long range
It should be obvious that coexistence in CSS is more powerful unlicensed communication,” in Telecommunications (ICT), 2016 23rd
than coexistence in BPSK-based networks as described in [14]. International Conference on. IEEE, 2016, pp. 1–6.
For maximal BER of 10−3 , 22 dB less power is required for [2] S. Kartakis et al., “Demystifying low-power wide-area communications
for city iot applications,” in Proceedings of the Tenth ACM International
CSS based networks. Workshop on Wireless Network Testbeds, Experimental Evaluation, and
Characterization, ser. WiNTECH ’16. New York, NY, USA: ACM,
V. D ISCUSSION 2016, pp. 2–8. [Online]. Available: http://doi.acm.org/10.1145/2980159.
There are three important insights that need to be empha- 2980162
[3] M. Centenaro et al., “Long-Range Communications in Unlicensed
sized. First, CSS is indeed robust against interference due to Bands: the Rising Stars in the IoT and Smart City Scenarios,” IEEE
the spreading, but the time-bandwidth footprint is significantly Wireless Communications, vol. 23, October 2016.
bigger than BPSK. This makes it more prone to multiple [4] K. Mikhaylov et al., “Analysis of capacity and scalability of the lora
low power wide area network technology,” in European Wireless 2016;
interfering sources and get the SINR below the threshold. 22th European Wireless Conference; Proceedings of. VDE VERLAG
Secondly, it should be noted that 16 dB is small if you take into GmbH, 2016, pp. 1–6.
account the possible link budget of 150 dB. If a competing [5] A. Springer et al., “Spread spectrum communications using chirp sig-
nals,” in EUROCOMM 2000. Information Systems for Enhanced Public
device is significantly closer to the base station, the packet Safety and Security. IEEE/AFCEA. IEEE, 2000, pp. 166–170.
from the far away node will not arrive at all even when [6] “Sigfox,” http://www.sigfox.com/en, 2014, accessed: 2015-02-01.
colliding with a packet with a different spreading factor. [7] M. Hata, “Empirical formula for propagation loss in land mobile radio
services,” Vehicular Technology, IEEE Transactions on, vol. 29, no. 3,
The final remark related to the overall network throughput. pp. 317–325, 1980.
From [15] and [16], we can estimate the overall throughput in [8] LoRaWAN Specification, LoRa Alliance, Inc, Jan. 2015, rev. 1.
the network. As LoRaWAN uses unslotted aloha, the equations [9] 860 MHz to 1020 MHz Low Power Long Range Transceiver, Semtech
Corporation, 2014, rev. 2.
of [15] apply. The maximal network throughput occurs when [10] C. Goursaud and J. M. Gorce, “Dedicated networks for iot: Phy / mac
on average 1/2 packet with spreading factor S is sent during state of the art and challenges,” EAI Endorsed Transactions on Internet
the transmission of one packet (l/Rb (S) with both parameters of Things, vol. 15, no. 1, 10 2015.
[11] L. W. Couch, II, Digital and Analog Communication Systems, 6th ed.
defined above) with the same spreading factor with a packet Upper Saddle River, NJ, USA: Prentice Hall PTR, 2000.
delivery ratio (PDR) of 1/e. This can be summed over different [12] J. G. Proakis, “Digital communications fourth edition, 2001,” 1998.
spreading factor assuming they do not interfere each other. The [13] A. RF, “Link budget analysis: Digital modulation, part,” http://www.
atlantarf.com/FSK Modulation.php, 2013, accessed: 2015-08-23.
BPSK network is narrowband and can hence use B/b channels [14] R. Axford, “Effects of cw-and bpsk-signal interference on a standard
to have the same bandwidth, where b is the BPSK bandwidth bpsk digital communications system,” DTIC Document, Tech. Rep.,
(100 Hz). The authors in [16] show that the maximal network 1992.
[15] A. S. Tanenbaum, “Computer networks. 5th,” 2011.
throughput occurs when on average 1/4 packet is sent per [16] C. Goursaud and Y. Mo, “Random unslotted time-frequency aloha:
channel during the transmission of one packet (l/RS with Theory and application to iot unb networks,” in Telecommunications
RS = 100 bps) with a PDR of 1/e. This is summarized (ICT), 2016 23rd International Conference on. IEEE, 2016, pp. 1–5.

View publication stats