You are on page 1of 7

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

net/publication/325493055

Building a connected BLE mesh: A network inference study

Conference Paper · April 2018


DOI: 10.1109/WCNCW.2018.8369004

CITATIONS READS
5 345

4 authors, including:

Alessandro Chiumento Brecht Reynders


Trinity College Dublin KU Leuven
38 PUBLICATIONS   483 CITATIONS    11 PUBLICATIONS   296 CITATIONS   

SEE PROFILE SEE PROFILE

S. Pollin
KU Leuven
281 PUBLICATIONS   3,717 CITATIONS   

SEE PROFILE

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

Novel & Unconventional SWIPT Modulation Techniques View project

Wireless Communication and Networking with Unmanned Aerial Vehicles View project

All content following this page was uploaded by Alessandro Chiumento on 20 June 2018.

The user has requested enhancement of the downloaded file.


Building a connected BLE mesh: a network
inference study
Alessandro Chiumento, Brecht Reynders, Yuri Murillo, and Sofie Pollin
KU Leuven, Department of Electrical Engineering, Kasteelpark Arenberg 10, 3001 Heverlee, Belgium
Email: alessandro.chiumento@esat.kuleuven.be

Abstract—Bluetooth low energy (BLE) is on the way of between network variables have been presented in [13]
becoming the next standard for low-power, low-datarate [14] and [15]. In [13] a Bayesian Network has been
applications. While not being designed directly for mesh used to predict wireless network service reliability in 5G
operation, recent works have shown that both connected
and broadcasts mesh are possible, this latter one being networks by determining a causal relationship between
ultimately included in the standard. For any robust opera- network measurements such as latency, signal to noise
tion in a connected BLE mesh network, especially for high ratio, throughput and reliability.
reliability and low-latency operations like healthcare, the
control parameters need to be carefully chosen in order In [14] and [15] the usage of structure learning tech-
to avoid congestion and packet loss but the relationships niques to infer interdependencies in network variables
between controllable parameters and final network per- in cognitive networks has been discussed. Both works
formance have not yet been investigated in BLE mesh
networks. In this work, we show that it is possible to infer focus on the importance of using probabilistic graphical
the relationships between the controllable and observable models to infer the network structure but, in both cases,
network parameters by using a mutual information based the authors make strong assumptions about the nature
structure learning approach; we show, in fact, how each of the actual wireless network studied thus gravely
setting such as transmit power, connection interval, source limiting the effectiveness of the proposed techniques
rate, impact overall network performance figures of merit
such as end-to-end delay, packet delivery ratio and network when applied to different communication technologies
build time. or network topologies.

I. INTRODUCTION In this work a mutual information based structure


Bluetooth Low Energy (BLE) was originally intro- learning approach is used to infer the interdependencies
duced in the Bluetooth Core Specification 4.0 as a pos- between network variables in a BLE mesh network from
sible way to bring Bluetooh connectivity to low-power actual measurements. The approach, based on [16], uses
devices which necessitate to operate for long periods of information theoretic concepts to first determine the
time without battery replacement. Increasing interest has statistical closeness of the network variables and, second,
been shown by the industrial community to use BLE assigns directionality between the variables. The struc-
for critical, indoor applications such as localisation [1], ture of the network is recovered from available traffic
continuous biometric readings in health structures [2], data time-series without making any assumption about
[3] and industry floor monitoring [4]. Particularly for the underlying structure of the data or prior knowledge
these indoor applications, it has been shown that, because about the system.
of the BLE devices’ limited range, a great number of The main contribution of this work is to provide an
nodes is needed to cover office environments [5]; in analysis of network inference as a tool to choose which
order to alleviate this problem, multi-hop communication design and run-time parameters in a BLE mesh network
protocols have been presented in recent works [6], [7]. need to be addressed in order to provide concrete insight
The first version of BLE mesh has recently been released on the network’s performance without relying on ad-hoc
[8], although no commercially available node supports assumptions or expert knowledge but allowing the net-
it yet nor any use case has been reported at the time of work dynamics to teach the system’s designer directly.
writing this paper.
While network inference, as structure learning, has The remainder of this work is structured as follows.
seen mounting interested in the research community Section II introduces the connected BLE mesh protocol
to learn interdependencies between variables in fields used for experimental testing. Section III provides an
such as genetics [9] and data mining [10], it has been account of mutual information-based inference and Sec-
used in wireless networks mostly to infer the physical tion IV depicts the measurement set-up and discusses the
topology rather than how the network variables influence network metrics available. Section V presents the results
one another [11], [12]. Approaches to learn interactions and, finally, Section VI concludes this work.
II. BLE CONNECTED M ESH N ETWORKING the time between advertisements to reduce the overhead.
A. Mesh network building and management If a connection is lost then the connection partner or the
neighbours of a failed node switch to higher discovery
The connection based mesh scheme chosen for this mode. As a self-healing mechanism, alternative routes
work was originally presented in [17]. The phases com- are generated every time a node disconnects. In this con-
posing the protocol are outlined below. text, regular exchange of advertising packets constitute
Node Initialisation: When a node is activated, it the key element for ensuring link maintenance and quick
generates automatically a Node ID and a Cluster ID. reaction to link changes.
The Node ID identifies the node in the network and the Routing: Routing is implemented in a neighbour-only
Cluster ID identifies to which cluster the node is part fashion. No routing tables are present and each node
of. At starting time the Cluster size is set to 1. As new is only aware of the nodes around it. When a sink is
nodes appear, the algorithm will force them to connect present in the network, this information is broadcasted
to bigger clusters until, in fact, all nodes in the network over the network. Each node then, when it has data to
are part of the same cluster and thus reachable by every send, selects the neighbour with the smallest distance to
other device. the sink (in number of hops) and sends all the packets
Discovery Phase: After initialisation, a node broad- to that device.
casts its Node Id, Cluster Id and its Cluster size as adver-
tisements in the so called JoinMe packets. In the mean B. Mesh Observable Parameters and Figures of Merit
while, it scans on all 3 broadcasting channels for similar The controllable and observable parameters used in
messages from its neighbours. This is continued until a this work are dependent on the technology used for
route is set-up. The discovery phase can have high to low communication, the application and the hardware im-
broadcasting interval duration depending on the stability plementation of the protocols; they are summarised in
of the network: the longer the network remains stable, Table I and discussed below.
the longer the time between advertisements. Figure 1
TABLE I: Controllable and Observable variables
shows an example of the data exchanged by nodes during
discovery. Controllable Observable
Name Symbol Name Symbol
Transmit Power PT X Packed Delivery Ratio P DR
Connection Interval CI End-to-End Delay De2e
Source Rate Rs Hop Count Hcount
Slave Connections S Build Time BT

The controllable parameters are addressable directly


in each BLE node and impact local behaviour between
neighbouring nodes:

Transmit power PT X : in order to avoid hidden


Fig. 1: Node state with connections. nodes problems due to asymmetric power profiles in the
mesh network [18], the power of all nodes is set to be
Route Set-up: After a predefined interval, each node identical and varied at the same time.
decides to which Cluster to connect based on the re- Connection Interval CI: the time between two data
ceived advertisement packets. The neighbours are ranked transfer events is defined as Connection Interval [8]. A
(based on signal strength and number of free connec- lower CI means that more data can be communicated
tions) and the node with the highest ranking becomes within the same time as the transmissions are closer
cluster head. The connection of two nodes belonging to together. The CI can range between 7.5ms and 4s [8].
the same cluster is avoided as information on Cluster Source Rate Rs : the amount of data generated by
ID is sent during the handshake. As a result the whole a source is application dependent and may be tweaked
network is loop free. If the Cluster IDs are indeed on demand. In this work the Source Rate is expressed
different, the node with the smallest cluster size connects in packets per second.
to the bigger cluster, possibly disconnecting its neighbor Slave Connections S: each BLE node may have
as only one connection as slave is allowed. Information multiple slaves, the number of possible slaves can influ-
on the new cluster sizes is then propagated via update ence the topology of the network at build-up time [19],
messages over the connected mesh to the nodes on both [17]. The maximum number of slave connections is then
clusters. denoted as S and is set equal for all the nodes in the
Self-healing and Route Maintenance: As the network network.
reaches stability and all the nodes are part of the same The observable variables are the figures of merit which
cluster, the discovery mode is switched to low increasing indicate the performance of the network. They denote
how the network behave as a whole and are not linked uncertain, the more information that process carries [20].
with a single node settings. When two variables are considered, the concept of joint
Packet Delivery Ratio P DR: The PDR represents entropy H(X, Y ), or how much uncertainty is associated
the ratio between the number of correctly received pack- with multiple variables is used:
ets at the sink NRx,Sink and the number of total packets XX
generated at the source NTx,Source . H(X, Y ) = − p(x, y) log p(x, y) (4)
x∈X y∈Y
NRx,Sink
PDR = . (1) where p(x, y) is the joint probability of X = x and Y =
NTx,Source
y happening at the same time. The Mutual Information,
End-to-end Delay De2e : The De2e represents the which quantifies the amount of information one variable
time difference between when a packet is generated at contains about the other, can then be derived directly
the source tGen,Source and when it is correctly received at from the joint entropy as [16]:
the sink tRx,Sink ,
XX p(x, y)
De2e = tRx,Sink − tGen,Source . (2) M I(X, Y ) = − p(x, y) log (5)
p(y)p(y)
x∈X y∈Y
Hop Count Hcount : given the multi-hop nature
of the mesh protocol, there might be multiple nodes Intuitively, if two variables interact closely, their mutual
between a source and the sink. As the connected mesh information will be large; if they are not related, their
protocol investigated builds a network between neigh- mutual information will be theoretically zero [16]. The
bouring nodes at start-up, the network topology may distance between the two time-varying processes can
be different every time the network is re-initialized. then be obtained by minimizing the distance between
The path between any source and sink may then be the variables X and Y, defined as e−M I(X,Y ) , for all
completely different every time the network is rebuilt. possible values of the time-delay τ between the two
Build Time BT : as the network grows organically variables:
(each new node tries to become part of the biggest d(X, Y ) = min e−M I(X(t±τ ),Y (t)) (6)
cluster), the network build time may vary considerably τ

depending on how the nodes connect to one another. The B. Reducing entropy to detect indirect interactions
Build Time BT is then defined as the time difference
Equation 5 implies that if the two random variables
between all nodes are turned on and the instant when all
X and Y are correlated, the joint entropy is actually
nodes are part of a single fully-connected cluster.
smaller than for uncorrelated variables [21]. By adding
III. N ETWORK I NFERENCE AND M UTUAL a new variable X ∗ connected to Y , there must then be
I NFORMATION an entropy reduction of:
The network inference tool used in this work infers the H(Y |X ∗ ) − H(Y |X ∗ , X ∗ )
relationships between the variables in three steps. First, ER (Y, X ∗ ) = (7)
H(Y )
it represents the distances between the variables in terms
of their statistical closeness, then it determines how In which H(Y |X ∗ ) is the conditional entropy,
strongly correlated the links variables are by providing ER (Y, X ∗ ) is the entropy reduction of Y due to the
links between them. Finally, it refines the prediction of introduction of the new variable X ∗ . As the MI defines
the inferred links by assigning directionality. the distance between the random variable the entropy
reduction ER defines the strength of the correlation, as
A. Mutual Information as distance between variables two random variables may be close but only indirectly
The interdependencies between the observed and con- correlated [16].
trollable variables are determined automatically from the
C. Determining directionality
time-series measurement data about the network perfor-
mance. The method used, employs a mutual information Mutual information is undirected, and thus it does
entropy reduction technique to deduce, from statistical not provide information on which random variable is
features of the data, how variables are correlated. In influencing the other ones [22]. Time series, on the
order to define mutual information (M I), the Shannon other hand, can be used to extract causality by using
entropy H(X) of a random variable X needs to be the transfer entropy, which is an asymmetrical measure
introduced: used to quantify the reduction in uncertainty in future
X values of a random variable Y by knowing the past and
H(X) = − p(x) log p(x) (3) present values of X and Y . By determining the causal
x∈X
relationship, in time, between the two random variables
where p(x) is the probability of X = x; Shannnons it is then possible to assign directionality to the links
entropy measures the uncertainty in a process: the more between variables [16], [23].
TABLE II: Experimental settings
IV. M EASUREMENT S ETUP
Experiment Rs [pps] PT x [dBm] CI [ms] S
The testbed used for this work consists of an indoor
1 5 -20 10 7
lab with a total of 12 nodes with a grid-like topology as 2 5 -16 10 7
depicted in Figure 2. One node always acts as a sink, 3 5 -12 10 7
collecting the data while any other node in the network 4 5 -8 10 7
5 5 -4 10 7
may be used as a source. At any given moment, there 6 5 0 10 7
may be only one source active. The nRF52 development 7 1 4 10 3
boards manufactured by Nordic Semiconductor have 8 1 4 10 7
been used in this work [24] and each runs the connected 9 5 4 10 3
10 5 4 10 7
mesh protocol discussed above [17]. 11 10 4 10 3
12 10 4 10 7
13 1 4 100 3
14 1 4 100 7
15 5 4 100 3
16 5 4 100 7
17 10 4 100 3
Sink 18 10 4 100 7
19 1 4 500 7
1.5 m

20 5 4 500 7
21 10 4 500 7
22 5 4 10 7

Fig. 2: Measurement setup. All vertical and horizontal dis-


tances are set to 1.5 m.

The experimental scenarios were performed as fol-


lows, each node in the network has been initialized
at the same time with the same parameters, then a
node randomly chosen as source started transmitting data
(256 packets) which is relayed to the sink; the same
experiment is repeated 20 times to account for channel
variations (i.e. a total of 5120 packets per experiment).
Finally, the network performance is collected for each
run. Before the beginning of the next experiment, new
node parameters have been chosen and the process is
repeated. Table II presents the experiments performed
and the values of controllable parameters for each ex-
periment.

V. I NFERENCE R ESULTS Fig. 3: Network structure learned from measurements


This Section presents the results of the network infer-
ence engine discussed in Section III on the data collected
with the measurements presented in Section IV. Source Rate: The Source Rate impacts strongly the
In Figure 3 the inferred structure of the network of PDR and weakly the End-to-End delay. From Figure 7 it
the observed and controlled variables is presented. The can be seen that the Source Rate is also a good predictor
thickness of the links connecting the nodes is indicative for the PDR. It is interesting to note that the Source Rate
of the strength of the dependency. and the Connection Interval have strong correlation; this
Connection Interval: The Connection Interval means that they influence one another greatly and thus
presents strong impact on both the PDR and the End- need to be taken both into account if optimizing for PDR.
to-End delay, this is visible also from the mutual infor-
mation plot in Figure 4. From the figure it is visible Transmit Power: The Transmit Power has effect
that the MI between the CI and the PDR is very high on the PDR and the De2e and quite lower impact on
(almost 1) for all ten time lags. This means that any the Hcount . This is due to the main effect that all
knowledge of the observed CI at time t provides well nodes seemed to be in communication range with one
correlated information about the behaviour of the PDR another, thus reducing the effect of the Transmit Power.
between times t−5 and t+ 5 making the CI a very good It is important to notice then that, different physical
predictor of PDR performance. network topology would have considerable effect on the
Fig. 4: Normalised MI between the CI and the observed Fig. 6: Normalised MI between the PT X and the observed
variables variables

Fig. 5: Normalised MI between the Rs and the observed Fig. 7: Normalised MI between the S and the observed
variables variables

measurement results and thus the correlation between


power and the observed variables would be different network behaviour. Interestingly, because of the network
Slave Connections: The number of available slaves build-up protocol, the Build Time is entirely unaffected
connections impacts network performance in a way by all the other variables, as long as communication
similar to the Transmit Power for the topology used in with neighboring nodes is possible; this entails that a
the experiments. new protocol needs to be designed to optimise network
A note on the Build Time: It is possible to see build-up rather than adjusting controllable parameters.
from Figure 8 that the Build Time is not impacted by Communication protocol design, network topology and
any of the controllable parameters. Furthermore, any run-time application requirements such as Source Rate
present knowledge of the Build Time does not give any have strong impact on the network and the inference
indication of future behaviour as visible from the only method is able to discern which parameter influences
peak in the figure at time instant t = 0. This is solely due which figure of merit on real network data. The insight
to the connected mesh protocol which builds the network given by such techniques on actual network operation is
randomly without taking into account node characteris- then valuable to design future networking solutions.
tics or behaviour (such as RSSI, buffer occupancy ...).
This behaviour showcases that network inference can ACKNOWLEDGMENTS
be used as a powerful tool to determine the impact of
protocol choices on actual network performance This work is supported by IWT CELTIC O&O project
ASUA/RoCCS. Yuri Murillo is funded by a PhD fellow-
VI. C ONCLUSION ship of the Research Foundation Flanders (FWO).
Network inference has been used in this work to
determine the impact of design choices and node settings
on one another and overall network performance in a
BLE connected mesh. The results show that the process
is powerful and able to infer relationships between
variables from data alone without expert knowledge. For
example, the inference system is able to pick up that in
order to keep PDR high and End-to-end Delay low the
Connection Interval needs to be carefully chosen. This
is not an unexpected result as similar observations have
been presented in [19] but the inference engine has de-
livered this observation without any prior knowledge of Fig. 8: Normalised MI between the BT and all the variables
R EFERENCES [13] B. Erman and S. Yiu, “Modeling 5g wireless network service
reliability prediction with bayesian network,” in 2016 IEEE In-
[1] R. Faragher and R. Harle, “Location fingerprinting with blue- ternational Workshop Technical Committee on Communications
tooth low energy beacons,” IEEE Journal on Selected Areas in Quality and Reliability (CQR 2016), May 2016, pp. 1–6.
Communications, vol. 33, no. 11, pp. 2418–2428, Nov 2015. [14] N. Du, Y. Bai, L. Luo, W. Wu, and J. Guo, “Building the
[2] H. H. Strey, P. Richman, R. Rozensky, S. Smith, and L. Endee, knowledge base through bayesian network for cognitive wireless
“Bluetooth low energy technologies for applications in health networks,” in 2011 IEEE 17th International Conference on
care: proximity and physiological signals monitors,” in 2013 10th Parallel and Distributed Systems, Dec 2011, pp. 412–419.
International Conference and Expo on Emerging Technologies for
[15] R. K. Yadav and B. S. Manoj, “On the use of undirected
a Smarter World (CEWIT), Oct 2013, pp. 1–4.
probabilistic graphical modeling for cognitive wireless networks,”
[3] C. N. C. H. B. M. E. J. G. R. W. B.-G. S. P. J. Ciemins
in 2013 Annual IEEE India Conference (INDICON), Dec 2013,
Elizabeth L., Arora Anupama and C. P. J., “Improving blood
pp. 1–6.
pressure control using smart technology,” in Telemedicine and
e-Health, ahead of print, Sept 2017. [16] A. F. Villaverde, J. Ross, F. Morn, and J. R. Banga, “Mider:
[4] R. Rondón, M. Gidlund, and K. Landernäs, “Evaluating Network inference with mutual information distance and entropy
bluetooth low energy suitability for time-critical industrial iot reduction,” PLOS ONE, vol. 9, no. 5, pp. 1–15, 05 2014. [Online].
applications,” International Journal of Wireless Information Available: https://doi.org/10.1371/journal.pone.0096732
Networks, vol. 24, no. 3, pp. 278–290, Sep 2017. [Online]. [17] M-Way Solutions, “Fruitymesh project page,” https://github.com/
Available: https://doi.org/10.1007/s10776-017-0357-0 mwaylabs/fruitymesh/wiki, accessed: 2017-07-28.
[5] P. D. Marco, R. Chirikov, P. Amin, and F. Militano, “Coverage [18] V. Shah, E. Gelal, and S. V. Krishnamurthy, “Handling
analysis of bluetooth low energy and ieee 802.11ah for office sce- asymmetry in power heterogeneous ad hoc networks,”
nario,” in Personal, Indoor, and Mobile Radio Communications Computer Networks, vol. 51, no. 10, pp. 2594 – 2615,
(PIMRC), 2015 IEEE 26th Annual International Symposium on, 2007. [Online]. Available: http://www.sciencedirect.com/science/
Aug 2015, pp. 2283–2287. article/pii/S138912860600346X
[6] H. S. Kim, J. Lee, and J. W. Jang, “Blemesh: A wireless mesh [19] Y. Murillo, B. Reynders, A. Chiumento, S. Malik, P. Crombez,
network protocol for bluetooth low energy devices,” in 2015 3rd and S. Pollin, “Bluetooth now or low energy: Should ble mesh
International Conference on Future Internet of Things and Cloud, become a flooding or connection oriented network?” in 2017
Aug 2015, pp. 558–563. IEEE 28th Annual International Symposium on Personal, Indoor,
[7] S. M. Darroudi and C. Gomez, “Bluetooth low energy mesh and Mobile Radio Communications (PIMRC) - Special Session
networks: A survey,” Sensors, vol. 17, no. 7, 2017. SP-04 on ”Resource-Efficient, Reliable and Secure Internet of
[8] Bluetooth SIG, “Bluetooth core specification 5,” December 2016. Things in the 5G Era, Oct 2017.
[9] J. Yan, S. L. Risacher, L. Shen, and A. J. Saykin, [20] C. E. Shannon, “A mathematical theory of communication,”
“Network approaches to systems biology analysis of complex SIGMOBILE Mob. Comput. Commun. Rev., vol. 5, no. 1, pp.
disease: integrative methods for multi-omics data,” Briefings 3–55, Jan. 2001. [Online]. Available: http://doi.acm.org/10.1145/
in Bioinformatics, p. bbx066, 2017. [Online]. Available: 584091.584093
+http://dx.doi.org/10.1093/bib/bbx066 [21] C. Axenie, C. Richter, and J. Conradt, “A self-synthesis approach
[10] M. Zanin, D. Papo, P. Sousa, E. Menasalvas, A. Nicchi, to perceptual learning for multisensory fusion in robotics,” Sen-
E. Kubik, and S. Boccaletti, “Combining complex networks and sors, vol. 16, no. 10, 2016.
data mining: Why and how,” Physics Reports, vol. 635, no. [22] P. Zoppoli, S. Morganella, and M. Ceccarelli, “Timedelay-aracne:
Supplement C, pp. 1 – 44, 2016, combining complex networks Reverse engineering of gene networks from time-course data
and data mining: Why and how. [Online]. Available: http:// by an information theoretic approach,” BMC Bioinformatics,
www.sciencedirect.com/science/article/pii/S037015731630062X vol. 11, no. 1, p. 154, Mar 2010. [Online]. Available:
[11] Y. E. Sagduyu, Y. Shi, A. Fanous, and J. H. Li, “Wireless network https://doi.org/10.1186/1471-2105-11-154
inference and optimization: Algorithm design and implementa- [23] T. Schreiber, “Measuring information transfer,” Phys. Rev.
tion,” IEEE Transactions on Mobile Computing, vol. 16, no. 1, Lett., vol. 85, pp. 461–464, Jul 2000. [Online]. Available:
pp. 257–267, Jan 2017. https://link.aps.org/doi/10.1103/PhysRevLett.85.461
[12] Y. Liang and R. Liu, “Routing topology inference for [24] Nordic Semiconductor, “Nordic nrf52 development
wireless sensor networks,” SIGCOMM Comput. Commun. Rev., board,” https://www.nordicsemi.com/eng/Products/
vol. 43, no. 2, pp. 21–28, Apr. 2013. [Online]. Available: Bluetooth-low-energy/nRF52-Preview-DK, accessed: 2017-
http://doi.acm.org/10.1145/2479957.2479961 07-28.

View publication stats

You might also like