HAL Id: hal-02968155
https://hal.archives-ouvertes.fr/hal-02968155
Submitted on 16 Oct 2020
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Energy-ecient GPS synchronization for wireless nodes
David Pallier, Vincent Le Cam, Sébastien Pillement
To cite this version:
David Pallier, Vincent Le Cam, Sébastien Pillement. Energy-ecient GPS synchronization for wireless
nodes. IEEE Sensors Journal, Institute of Electrical and Electronics Engineers, 2021, 21 (4), pp.5221
- 5229. �10.1109/JSEN.2020.3031350�. �hal-02968155�
IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2017 1
Energy-efficient GPS synchronization for
wireless nodes
David Pallier, Vincent Le Cam, Sébastien Pillement
Abstract— Synchronization is a challenging problem for wireless nodes, especially for applications demanding good
synchronization accuracy over wide areas. In that case, the GPS is a valuable solution as the nodes can independently
synchronize to UTC. However, the energy consumption of a GPS receiver (over 100 mW when switched on) is not
sustainable on a wireless node. Therefore, in this work, we developed a synchronization scheme based on periodic
extinctions of the GPS receiver. The goal is to study the GPS power switching effect on the synchronization accuracy.
To do so, a node with dedicated timestamping hardware was designed. Two clock models were compared to predict
the node time when the GPS is off and the impact of a Kalman filter, to remove the GPS noise, was evaluated. From
experimental data, we show that the choice of the clock model depends on the accuracy needed and that the Kalman
filter improves the estimation of the clock frequency for both models. In our design, the GPS can be off from 60% up to
95% of the time for mean synchronization errors of 20 ns to 420 ns, respectively. This work demonstrates that GPS power
switching is an efficient solution to reduce energy costs while maintaining a high synchronization accuracy.
Index Terms— Clock synchronization, WSN, Energy efficiency, FPGA, GPS, Kalman filter, Structural Health Monitoring
I. INTRODUCTION
M
ONITORING the health of structures often implies the
deployment of many sensor nodes to sample physical
parameters such as ambient temperature, constraints, vibra-
tions, acoustic waves, electrical fields, etc. Posted to a central
supervisor, in delayed or in real-time, those samples are fed
to Structural Health Monitoring (SHM) algorithms to assess
the infrastructure health. As most of these parameters are time-
dependent and sampled from different nodes, the system needs
to have a common time reference. The following examples un-
derline the need for synchronization and the accuracy needed
for SHM applications:
1) The natural frequency range of vibrations for structural
diagnostics (modal analysis) is typically comprised be-
tween 0 and 100 Hz. Thus, accelerations are usually
sampled at 1000 Hz. To ensure good modal reconstruc-
tion, the synchronization error between nodes must be
lower than 1 ms.
2) Acoustic wave propagation is often used to localize the
crack origin in a steel bar or a wire break in bridge
cables. The method based on Time Difference Of Arrival
(TDOA) is very dependent on the sensor’s ability to
timestamp acoustic wavefronts. With a typical velocity
of 5000 m/s and a localization accuracy expected to +/-
15 cm, synchronization error between two nodes must
be lower than 30 microseconds.
3) The TDOA method is also used to monitor the localiza-
tion of lightning on high electrical voltage lines. With
This work was supported by the WISE program of Pays-de-la-Loire
region
D. Pallier and V. Le Cam, Univ. Gustave Eiffel, Inria, COSYS-SII, I4S,
F-44344 Bouguenais, France
S. Pillement, Univ Nantes, CNRS, IETR UMR 6164, F-44000 Nantes,
France
a speed of 200 000 km/s and an expected localization
of +/- 50 m, synchronization error between two nodes
must be under 250 nanoseconds.
Time-counting on an electronic device, such as a wireless
node, is mainly based on the use of an oscillator. Unfortu-
nately, oscillators are highly sensitive to environmental condi-
tions such as temperature, acceleration, or voltage supply sta-
bility [1]. Consequently, the frequency of the oscillator drifts
from its theoretical value over time. Without a synchronization
protocol, each node will have its own local time and the offset
between the clocks of the nodes increases over time. In most of
the cases, this synchronization protocol relies on timestamps
exchange between the nodes. The synchronization accuracy
that can be obtained in the network depends on the accuracy
of the timestamps exchange, the synchronization update rate,
and the stability of local oscillators.
A wireless sensor network (WSN) is a cost-effective way
to deploy sensing nodes onto already existing large structures
(bridges, wind power plants, railway lines, etc). This is why
WSNs are more and more used for SHM applications. How-
ever, synchronization between nodes is harder in wireless net-
works due to non-deterministic propagation delays. Classical
wireless synchronization protocols such as [2], [3] and [4]
cannot achieve sub-microsecond accuracy, especially in multi-
hop scenarios. In this context the Precision Time Protocol
(PTP) has been extended over 802.11 communications with
hardware timestamping [5] or synchronization over UWB
communications [6] and can achieve sub nanosecond accuracy.
While these protocols can offer good synchronization accuracy
for the monitoring of small structures, they cannot be used
for large structures like bridges, railways or power-lines, as
they need close and iterative synchronization process. For large
outdoor systems (which are most of SHM applications) a satel-
Page 2 of 24
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
2 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2017
lite navigation system, like GPS, can be used to synchronize
each node to UTC as in [7]. This method is suitable for large
infrastructures and allows timestamping accuracy up to tens
of nanoseconds. The main limitation of this technique is the
energy cost of a GPS receiver. Since energy consumption
needs to be minimized in wireless systems, GPS solution
needs to be optimized according to the required timestamping
accuracy.
In this context, we developed a new synchronization scheme
based on GPS holdover. Namely, GPS is used for precision
purpose but it will be turned off to save power. During the
off-state, the local time of the node is estimated using a
clock model. The synchronization accuracy obtained with this
scheme depends on the chosen holdover parameters: from
tens of ns to 1 ms. To further improve the estimation of
the clock model parameters, the GPS signal can be filtered
when in on-state. The contributions of this work are: (1) the
design of a node with dedicated hardware for timestamping,
(2) the accuracy and energy efficiency evaluations of our
synchronization scheme for different clock models, and (3)
the study of the impact of frequency filtering using a Kalman
filter.
The paper is organized as follows. The related works
are summarized in Section II. Node architecture, timestamps
computation, the use of GPS for periodic re-synchronization
and the experimental setup are described in Section III. Section
IV introduces the clock models, and the Kalman filter. The
results are discussed in Section V, while Section VI resumes
our contributions and outlines future perspectives.
II. RELATED WORKS
Since synchronization in distributed systems was first for-
malized in [8] and later extended in [9], numerous solutions
have been developed to tackle the problem of synchroniza-
tion in networks. The Network Time Protocol [10] and the
Precision Time Protocol [11] are the most used network syn-
chronization protocols over the wired Ethernet. It is possible
to obtain microsecond synchronization accuracy with these
protocols, and an extension of the PTP [12] aims to achieve
sub-nanosecond accuracy.
Classical synchronization protocols for WSNs are based on
the exchange of dedicated timestamped RF beacons. These
protocols can be divided into two categories: the sender to
receiver and the receiver to receiver protocols. In the sender
to receiver scheme, a node sends its time to a receiver that
timestamp the arrival of the packet using its local time.
Then, the receiver synchronizes itself by estimating the offset
between its local time and the time of the sender. To estimate
the propagation delay a two-way exchange has been developed
in [3]. The same principle is used in [4] but on a mesh
topology instead of a tree and with the estimation of the packet
processing overhead. Other works [13] [14] [15] also rely on
the same principle but are optimized for specific nodes with
constrained resources like PicoRadios [16]. In the receiver
to receiver scheme [2] [17] a third party node broadcast a
beacon that doesn’t contain a timestamp. All the receivers
timestamp the arrival of the beacon with their local clock
and then exchange their observations pairwise to calculate
their offsets. The performances of all these protocols are
limited by the propagation delays and the packet processing
overhead. The interested reader can refer to [18] that outlines
the existing implementations of these techniques and their
limitations. While some implementations of these protocols
can achieve microsecond accuracy most of them are in the
tens of microsecond range. Moreover, the accuracy decreases
with the number of hops which can be problematic on large
structures such as bridges, railways, or power-lines. Signal
processing techniques have been described in [19] to deal with
non-deterministic propagation delays but, to the best of our
knowledge, these techniques have not been implemented on
real wireless nodes.
Other works focus on the implementation of clock syn-
chronization in wireless communication norms. ZigBee is the
most used protocol in the Low Rate Wireless Area Personal
Network (LR WPAN), which adds network and application
layers on top of the 802.15.4 protocol. In [20] the Flooding
Time Synchronization Protocol (FTSP) is used to synchronize
nodes over ZigBee to obtain a timestamping accuracy of 61
µs. The use of UWB for the physical layer was added to
IEEE 802.15.4 as an amendment in 2007 and merged in
IEEE 802.15.4-2011. In [6] UWB is used to synchronize
the clocks of a pair of chips on the same PCB with an
accuracy of 374 ps (standard error of 677 ps). Several im-
plementations of synchronization protocols over IEEE 802.11
are compared in [21]. IEEE 802.11 compliant synchronization
solutions can achieve µs clock synchronization accuracy [22]
and non compliant solutions using PTP (IEEE 1588) can
achieve nanosecond accuracy [5] or sub-nanosecond accu-
racy with a hardware timestamping mechanism. While these
synchronization schemes can offer a good synchronization
accuracy on small structures, they are not suited for large
structures, where the nodes can be distant of a few hundred
meters to several kilometers. More recently, Low-Power Wide-
Area Network (LPWAN) communications have gained interest
for industrial applications. Based on LoRa, Sigfox, or NB-
IoT, these protocols allow for communications over a few
kilometers in urban areas. In [23] LoRa and NB-Iot are used
for node communications in the context of machine vibrations
monitoring. Nodes are synchronized every 10 s using LoRa
with an accuracy of 5 µs.
Another approach to wireless node synchronization is to
use a satellite navigation system, like the GPS, to make
each node synchronized to UTC. This solution has been
implemented in [7] and [24]. The clock of the node is steered
by the PPS signal coming out of its GPS receiver. To achieve
an accuracy of tens of nanosecond, dedicated hardware is
used with high-speed counters. The use of hardware induces
deterministic timestamping which removes software overhead
and thus improves the timing accuracy. While this approach is
suitable for accurate synchronization, a GPS receiver requires
a lot of energy (up to 100 mW). This cost can be affordable
for applications requiring a high synchronization accuracy,
but for less critical applications that require less accuracy,
the energy consumption needs to be optimized. In this work,
we propose an adaptive synchronization protocol based on
Page 3 of 24
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
AUTHOR et al.: PREPARATION OF PAPERS FOR IEEE TRANSACTIONS AND JOURNALS (FEBRUARY 2017) 3
automatic on/off switching of the GPS receiver to minimize
the energy consumption under a defined accuracy constraint.
Adaptable synchronization accuracy has already been de-
scribed in [17] as an extension of RBS [2]. In [25] a
Kalman filter is used to track the offset of a local clock
from timestamped message exchange and an algorithm is
developed to find the optimal message exchange rate for a
parametric accuracy. Our work differs from these solutions as
it is designed on top of GPS synchronization (although any
GNSS receiver can be used) instead of beacons exchange and
relies on dedicated timestamping hardware to achieve up to
tens of nanoseconds synchronization accuracy.
III. WIRELESS NODE
A. Node Architecture
The node design presented in this paper is similar to
the one presented in [7]. It includes a dedicated hardware
timestamping unit. The timestamping unit is connected to
a GPS receiver and has inputs connected to digital signals
that have to be timestamped (see Figure 1). The node is
synchronized with the PPS signal and the epoch transmitted on
a serial link from the GPS receiver. A rising edge on the PPS
signal corresponds to the beginning of a UTC second. The
epoch corresponds to the number of seconds elapsed since
January 6th, 1980. It is used to compute the UTC date at the
next rising edge of the PPS signal. The detection of a rising
or a falling edge on a digital input of the node is called an
event. The events are timestamped with high-speed counters
in the timestamping unit and are stored until they are read by
the main control unit of the node. This node architecture has
several advantages:
• No processing overhead: The use of dedicated hardware
on FPGA suppress processing overhead at event detec-
tion. Since the delay between the rising edge on the input
and the timestamping is deterministic and not affected by
software overhead due to process scheduling and interrupt
latency, the timestamping accuracy is improved.
• Parallel high granularity timestamping: On most micro-
controller units, the number of parallel counters able to
timestamp independent events is limited. Another limiting
factor is the time granularity fixed by the quartz oscillator
frequency. With a custom design, the number of parallel
counters is only limited by the FPGA size. The frequency
is fixed by the PLL included in the FPGA and can be
higher than the quartz oscillator frequency, allowing a
higher granularity.
• Asynchronous data fetching and main control unit inde-
pendence: Since timestamping and data processing are
separated, the main control unit can fetch data (on digital
inputs) asynchronously on the timestamping unit. Thus
any operating system and any microcontroller can be used
as long as the microcontroller is equipped with a serial
link.
In this work, we focus on the timestamp unit design. For
the main control of the node, without loss of generality and
for quick prototyping purposes, we used a Raspberry Pi.
Fig. 1: Timestamp unit
B. Timestamp computation
We define the local estimation of the PPS signal (according
to the local quartz oscillator) as the fake PPS signal. The offset
θ is then the time between a rising edge of the true PPS signal
and the next rising edge of the fake PPS signal (issued by the
GPS). It corresponds to the time difference between the local
clock and the GPS time. ∆ is the time between a rising edge
of the fake PPS signal and a rising edge on a digital input (i.e.
an event). θ and ∆ are represented in the timing diagram in
Figure 2.
Fig. 2: Timing diagram
The date t[n] of a recorded event occurring during the n
th
local second is computed based on the date of the true PPS
rising edge t
pps
[n], that occurred during the fake PPS period
and the elapsed time between the event and this rising edge.
The elapsed time corresponds to the measured offset θ[n]
minus the measured ∆[n] for this particular event. This time
is positive when the event is detected after the true PPS edge
and negative when detected before. Hence:
t[n] = t
pps
[n] + θ[n] − ∆[n] (1)
Page 4 of 24
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
4 IEEE SENSORS JOURNAL, VOL. XX, NO. XX, XXXX 2017
C. Timestamping unit Architecture
The above timlestamp computation is implemented on a
Spartan 6 FPGA and is shown in Fig. 1. Since this design
aims at turning off the GPS, the clock synchronization is
different from the one used in [7]. In this design, the clock
counter is not reset by the PPS but instead is bounded by
N
0
. This integer corresponds to the number of clock beats
counted over one second at the nominal frequency f
0
of the
quartz oscillator. In the meantime, the local time offset θ is
recorded to convert the local timestamp to UTC. The fake PPS
signal is generated inside the FPGA, by dividing the local
clock frequency to obtain a 1Hz square signal. The offset
corresponds to the differential between this signal and the
true PPS signal from the GPS receiver. Tracking the local
time offset has been chosen over the control of the clock
as it allows for both offline and online synchronization. This
enables us to perform the timestamp conversion, from local
time to UTC, on the fly or post facto if the offset is recorded
alongside the events. This post facto synchronization allows
us to test different synchronization schemes from the same
events dataset for evaluation purposes.
The values θ and ∆ are not directly measured inside the
timestamp unit as they are integers corresponding to counter
values and not timing values. Let N
θ
and N
∆
be the counter
values corresponding to the offset counter and the event
counter, respectively. These integers are converted into time
values using the true frequency of the oscillator. Equation 1
becomes:
t[n] = t
pps
[n] +
1
f[n]
(N
θ
[n] − N
∆
[n]) (2)
With N
f
, N
θ
and N
∆
the counter values from the timestamp
unit, while t
pps
is obtained from the GPS receiver.
The number of local clock beats counted during a full period
of the true PPS signal is N
f
. Let T
GP S
be the time elapsed
between two rising edges of the true PPS signal. If the true PPS
signal is perfect then T
GP S
is one second. The true frequency
of the local oscillator can be computed as follows:
f[n] =
N
f
[n]
T
GP S
[n]
(3)
giving:
t[n] = t
pps
[n] +
T
GP S
N
f
[n]
(N
θ
[n] − N
∆
[n]) (4)
If the true PPS signal is perfect, the only residual error is
due to the granularity resolution as N
f
, N
θ
and N
∆
are integer
values. The frequency output of the PLL inside the FPGA was
set to 240MHz giving a counter resolution of 4.17 ns.
D. Periodic re-synchronization from GPS
As described in [26], GPS synchronization allows for a
timing accuracy under 60ns. However, it entails a substantial
energy cost as the GPS receiver is constantly on. When a
lower accuracy is needed, GPS can be used periodically to
re-synchronize the node. Between two synchronizations, the
GPS receiver is off to save energy. The accuracy that can be
obtained depends on the re-synchronization period. To be able
to deliver the PPS and the epoch, the GPS receiver needs to
decode parts of the navigation message broadcasted by the
GPS satellites. In normal conditions the GPS receiver has to:
• Generate PPS: In order to generate the PPS signal, the
GPS receiver needs to track the "in-view" satellites,
compute its time and position (called fix) and then lock
its clock. According to [26] the time to first fix is
under 1 second when the ephemeris are still valid. From
experimentations, we observed less than 1 second to get
a fix and less than 4 seconds to lock the clock.
• Refresh ephemeris: In order to track the in-view satellites
the receiver needs to compute its position, therefore the
ephemeris needs to be updated. In this work, we update
them at their typical update rate of two hours.
• Acquire a complete navigation message: Since the
ephemeris are already acquired every two hours the
complete navigation message is acquired once a day.
In between the above phases, we propose to turn on and
off the GPS to save energy. The off phase is called holdover
and correspond to the time between two consecutive GPS time
synchronizations. Figure 3 shows the on and off phases of the
GPS receiver for a 24 hours period. The T_NAV is 25 minutes
long every 24 hours. The T_EPH is 1 minute long every 2
hours. The GPS is on during k
on
seconds and in holdover
during k − k
on
seconds. In this mode, the GPS receiver
is switched off (except its memory to keep the ephemeris).
Equation 5 gives r, which represents the on/off ratio of the
GPS receiver over 24 hours as a function of the parameters k,
k
on
, T
NAV
and T
EHP
.
r = T
NAV
+
1
86400
.(k
on
.
7200 − T
NAV
k
+ 11.k
on
.
7200 − T
EP H
k
)
(5)
The minimum k
on
is 5 seconds to get at least one locked pulse
on the PPS signal, after this first delay the receiver delivers
one pulse per second. Two pulses are required to get N
θ
, N
∆
and N
f
. The main difficulty is then to track the offset of the
fake PPS, while the GPS is in holdover mode.
Fig. 3: Synchronization process from GPS
IV. OFFSET TRACKING
A. Clock model
The offset needs to be predicted during the holdover to have
accurate timestamping. To do so a clock model is used. Let
Page 5 of 24
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
