HAL Id: hal-03083144https://hal.archives-ouvertes.fr/hal-03083144

Submitted on 18 Dec 2020

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

A Robustness Comparison of Measured NarrowbandCSI vs RSSI for IoT Localization

Ahmed Ghany, Bernard Uguen, Dominique Lemur

To cite this version:Ahmed Ghany, Bernard Uguen, Dominique Lemur. A Robustness Comparison of Measured Nar-rowband CSI vs RSSI for IoT Localization. 2020 IEEE 92nd Vehicular Technology Conference:VTC2020-Fall, Nov 2020, Victoria (virtual), Canada. �10.1109/VTC2020-Fall49728.2020.9348854�.�hal-03083144�

A Robustness Comparison of MeasuredNarrowband CSI vs RSSI for IoT Localization

Ahmed Abdel Ghany, Bernard Uguen, Dominique LemurInstitute of Electronics and Telecommunications of Rennes (IETR)

University of Rennes 1Rennes, France

{ahmed.abdelghany, bernard.uguen, dominique.lemur}@univ-rennes1.fr

Abstract—Received Signal Strength Indicator (RSSI)-basedfingerprinting is currently viewed as an important techniquefor the positioning capabilities in the Internet of Things (IoT).However, in the case of practical measurement, the localizationmethods based on RSSI are easily affected by the temporal andspatial variation, which contributes to most of the estimationerrors in current systems. In this paper, the feasibility ofutilizing the Channel State Information (CSI) for localization isstudied, after knowing that the CSI contains information aboutthe channel between the sender and receiver at the level ofindividual data subcarriers. Unlike most of the previous work,the intended approach is to use the entire subcarrier magnitudeswithout averaging or any reduction of the obtained narrowbandCSI. Moreover, the frequency hopping in the LoRa systemsshould be a profit for localization by getting access to a widerband. In order to obtain a reliable basis for this approach, anoutdoor measurement campaign is performed in the area of theCampus Beaulieu in Rennes to estimate the CSI of transmittedLoRa signals from different locations. For this, it is necessaryfor the individual channels from each different position to beappropriately different from one another to achieve significantlocalization gain. Hence, a comparison is done investigating theattainable evolution in the CSI at each location based on theCSI slope versus its average amplitude. In the given results, thefeasibility of using the proposed technique is asserted by thedrastic stability of the CSI slope over time and space, in contraryto the CSI average amplitude. This manifests the robustness ofthe CSI to the signal fluctuations and its more valuable renderingthan the RSSI.

Index Terms—IoT, LoRa, Localization, CSI, RSSI

I. INTRODUCTION

Beyond the traditional voice, video and data serviceswhere data throughput is the main purpose in the context ofthe Internet of Things (IoT), the focus here is on low-costdeployments with large coverage areas [1]. For providing thisconnectivity, Low Power Wide Area Networks (LPWAN) areconsidered the major technology, especially the LoRaWANwhose design compromises between low energy consumptionand a large communication range [2].

Localization is one of the essential features of many IoTmodules due to the very nature of the data collected from thedevices [3]. These usages involve location detection of assetsin a warehouse, patient tracking inside the building of thehospital, and emergency personnel positioning in a disaster

area. Therefore, precise localization is still a critical missingcomponent and it has been gaining growing interest from awide range of applications. Although equipping each sensorwith a global positioning system (GPS) chip is considereda tempting option, it is not a simple solution because itrequires adding a GPS tracker to a device which will increaseboth cost and power consumption [4]. Subsequently, agreat number of researches have been done to address thedomain of GPS-Free localization in IoT. These methods areinvestigated based on Received Signal Strength Indicator(RSSI), Angle of Arrival (AoA), Time of Arrival (ToA), TimeDifference of Arrival (TDoA) and their multiple integrations[5]. For localization, these measurements are utilized by thegateways to determine their relative position relations with thesource. However, using the AoA and ToA techniques alwaysrequire a precise calibration or an additional hardware [6].Therefore, these two kinds of techniques are not so practical.While TDoA is considered the most popular technique forlocalization as it does not require the transmitter to besynchronized with the receivers but only the gateways arerequired to have synchronized clocks [7]. Nevertheless, thisaccurate time synchronization can only be achieved by addinga GPS receiver at each gateway which needs an additional cost.

Among all the localization technologies, wireless RSSIfingerprinting has proven as an effective positioning techniquedue to its simplicity and deployment practicability [8].Fingerprinting based localization avoids hardware deploymentcost and effort by relying on existing network infrastructure.It just relies on the received signal strength at each gatewayto localize the node. However, RSSI-based fingerprintinglocalization methods may have poor positioning performanceas RSSI always vary due to the large signal power fluctuationsboth in time and space, moreover, each measured RSSI valuedepends on the hardware accuracy as well as the systemcalibration for every measurement [9]. Besides, RSSI containscoarse information so as to not fully utilize the abundantchannel information in each subcarrier. Therefore, some littlework has been published whose aim is to take the averagevalue of the whole Channel State Information (CSI) subcarrieramplitudes which is proven to be more temporally stable indifferent environments and helps maintain the performance

over time in comparison with RSSI [10].

In this paper, it is favorable to show the possibility toleverage CSI for improving the performance of positioningby investigating the profit of using the entire subcarriermagnitudes without averaging or any reduction. Thus,an initial measurement campaign is done to compute thenarrowband CSI of transmitted LoRa signals from differentlocations. To achieve significant localization gain, it isnecessary for the individual channels from each differentposition to be uncorrelated with one another. Thus, thepresented short time data indicate that channel slopes witheven short separated distances will be quite stable and showweak intercorrelations between them. Furthermore, the CSI ineach position after a while is more correlated with itself whichcan achieve a significant diversity gain in comparison with themean amplitude of the CSI. This trial allows us also to deriverecommendations for the use of diversity at the receivingsite in short-range outdoor-to-outdoor transmission systems,asking questions like: ”At what distance must verticallypolarized antennas be placed such that intercorrelation islow and hence gateway diversity is potentially beneficial forlocalization?” or ”Can also the LoRa frequency hopping adda diversity gain for a static transmitting node scenario?”

The remainder of this document is organized as follows.Section II presents the measurement overview and Section IIIprovides sufficient detail of the proposed post-processing algo-rithm. The results of the channel correlation analysis are thenpresented and commented in Section IV. Finally, Section V isdedicated for conclusions.

II. SYSTEM AND MEASUREMENT SETUP

The main concept of the proposed experiment istransmitting repeated up-chirps signal to sense thechannels consecutively at the typical uplink frequencybands, i.e 9 channels with center frequency fk ∈{867.1, 867.3, 867.5, 867.7, 867.9, 868.1, 868.3, 868.5, 868.8}MHz, and 125 kHz bandwidth. This is considered as atraditional channel sounder with a typical structure whichhas the Transmitter (Tx) and Receiver (Rx) placed at twodifferent locations, as shown in Figure 1. Thus, the Rxantenna is fixed on the roof of the university building asshown in Figure 2a. While the Tx has a mobile structure witha laptop and a Universal Software Radio Peripheral (USRP)as it is described in Figure 2b. First, the Tx, i.e. located inspecific positions within the area of the Campus Beaulieu inRennes, should generate a signal by a laptop and transmit itusing the USRP to sample the channel for a specific timeinterval. While the stationary Rx, whose antenna is locatedabove the building, should receive the signal with its USRPat the same time interval of transmitting. After the Rx picksup the signal, the desktop computer stores it to perform anessential post-processing algorithm on the received signal tomitigate the imperfections and obtain the channel transfer

function as detailed in the following subsection.

Fig. 1: The locations of the three measured points in the area ofthe Campus Beaulieu. Positions of Tx and the Rx are labeledby black, and red markers respectively. (©by OpenStreetMapContributers)

Whereas the aforementioned emulated preamble LoRa sig-nal is generated using python from the mth cyclic time shiftof the basic chirp such that x[n] is expressed as:

xm[n] =

√1

2SFexp(j2π

(((m+ n) mod 2SF )− 2SF−1)2

21−SF ),

(1)with m = 0 for the basic chirp symbol, while the Spreading

Factor (SF) is chosen to be equal to 7. Where n depicts thesample index n = 0, 1, 2, ..., 2SF − 1.

III. DATA PROCESSING

On the Rx side, the saved file is imported to be analysedfor each center frequency fk. Thus, the following main signalprocessing techniques are carried out with the same order.

A. Frequency synchronization

The frequency and time synchronization is applied at first,respectively. The objective of frequency synchronization is toestablish the subcarrier orthogonality by correcting the phaseas an initial step before applying any further processing [11].Let define the Carrier Frequency Offset (CFO) by ∆fk asbeing the difference between the up and down conversionfrequencies. This CFO results in a phase offset ∆φk = 2π∆fk

fsbetween two samples with the same index in consecutiveupchirps. The residual part of this offset can be estimated bytaking the average across the received entire symbol yk[n] as:

(a) Position of the Rx monopole antenna fixed on the roof.

(b) A trolley shelf with the different parts of the transmissionequipment at Tx location 2.

Fig. 2: Views from the Tx and the Rx sites.

∆̂φk[l] = arg (

2SF−1∑n=0

yk[l + n]y∗k[l + n+ 2SF ]). (2)

This method of detecting the CFO is described as thefrequency acquisition algorithm and is utilized over thewhole handled signal portion. As shown in Figure 3, theangles of the differential correlation function indicate thatthere are some phase deviations. Hence, the phase erroris compensated and the angles of the corrected differentialcorrelation function become concentrated around zero.

B. Channel estimation

The up-chirps in the LoRa preamble, i.e. typically consist ofeight symbols, are considered as a channel sounder. Therefore,the least squares estimate of the raw channel transfer functionHraw can be estimated for a simple division as [12]:

l

Fig. 3: The angles of differential correlation function ∆̂φk[l]before and after correction of the whole portion of about 30symbols.

Hraw =Y

X, (3)

where X and Y are the training and the received symbolin the frequency domain respectively. Hence, the proposeddenoising technique is applied on the raw channel transferfunction Hraw. This imposes doing an incipient step, i.e.removing the thermal noise from the subcarriers in the channeltransfer function. Thus, the raw channel transfer functionHraw is low-passed in the frequency domain using FIR filterto maintain the property of linear phase as:

Hfiltered = Rh ·Hraw, (4)

where Rh is the autocorrelation matrix of the channel andthe filtered channel Hfiltered is also defined as Hl

fk. Where

fk and l are the center frequency value in MHz and thelocation number, respectively. As shown in Figure 4, one canobserve that the obtained CSI is smooth without any noise atthe three different locations. Moreover, it is obvious that thechannels have a different magnitude of attenuation at the samesubcarrier through the different locations.

0 20 40 60 80 100 120Subcarrier index

1.2

1.0

0.8

0.6

0.4

0.2

0.0

Mag

nitu

de [d

B]

H1868.5

H2868.5

H3868.5

Fig. 4: The normalized CSI with 125 kHz bandwidth.

IV. CORRELATION ANALYSIS

To check the plausibility of utilizing the CSI for localization,the CSI spatial and temporal evolution at each location isanalysed using a proposed method based on the CSI slope.Furthermore, this technique is also compared to the traditionalmethod which determines each position from its CSI averageamplitude as detailed in the following subsections.

A. CSI slope

Based on the LoRa narrow-band characteristic, i.e. have abandwidth W = 125 kHz in Europe, a linear variation of thechannel transfer function is assumed over frequency bandwidthcentered on fk. This condition is generally valid under theconsidered reality of such a very narrow bandwidth, thus,all the obtained channel transfer functions are almost flat asshown in Figure 4. Nevertheless, it is clear that each CSI hasa unique slope. In this paper, the CSI slopes are estimated toanalyse the CSI evolution from each location i as:

sifk =| Hi

fk[0] | − | Hi

fk[2SF − 1] |

W, (5)

Then the Normalized Slope Distance (NSD) between anytwo positions for the whole K uplink frequency bands isdefined as:

Si,j =1

2K

K−1∑k=0

| s̃ifk − s̃jfk| (6)

with

s̃ifk =sifkSmax

, (7)

where Smax is equivalent to the absolute value of themaximum observed slope in the measured dataset. This ischosen to be the scaling factor for the slope normalization.Consequently,

0 ≤ Si,j ≤ 1. (8)

The magnitudes of the aforementioned normalized slopes̃ifk are given in Figure 5a for each different location i andcenter frequency fk. Thus, the full scale is normalized andlies in [−1, 1], whereas every amplitude can vary on a severalorder of magnitude levels. Each frequency of the K uplinkfrequency bands is labeled with a specific color. In order toanalyse the time variation of the CSI slope at each position,over a time duration τ of about 10 minutes, an arrow isdrawn for every center frequency from the first time instantto the second one. This measurement dataset demonstratesthe slow alteration of the CSI slope with respect to time.

Moving to the NSD Si,j values in Figure 6a, one canobserve the high level of NSD values in the three locationcombinations, which are S1,2, S1,3 and S2,3. This indicatesthat CSI models from different positions are distinctly differ-ent. While the CSI for each location significantly underlines

a low NSD value with itself after a τ time interval. Thisis obvious in S1,1τ , S2,2τ and S3,3τ whose values are nearzero as well as they are far away from the other NSD values,i.e. labeled by circles. So the environment of the propagationpaths for each specific position is stable with only marginalmodification during this duration.

B. CSI average magnitude

On the other hand, the channel evolution at each location isinvestigated by the mechanism, which depends on calculatingthe CSI mean magnitude. This method is considered as analternative way rather than estimating the RSSI value. Thus,the CSI average amplitude is estimated for each location i as:

rifk =1

2SF

2SF−1∑n=0

| Hifk

[n] |, (9)

Hence, the Amplitude Distance (AD) between any twopositions for the entire K uplink frequency bands is computedas:

Ri,j =1

K

K−1∑k=0

| rifk − rjfk| . (10)

For the whole evaluated locations and frequency bands,the measured values of the CSI average amplitude rifk aregiven in Figure 5b. It is foremost supposed that the CSI slopevariability from one sub-band to another is more informativethan in the CSI average magnitude whose amplitudes atvarious bands are evaluating roughly at near levels. Moreover,it seems that the CSI average amplitude has more alternationin time than the CSI slope, which is particularly clear atlocation 2 when considering the shift between the values of2 and 2τ .

As depicted in Figure 6b, the AD Ri,j values obtained forall the location combinations preserve proper values. However,the AD values, i.e. labeled by triangles, for each location withitself after the time interval τ are near to the AD values ofdifferent location combinations, i.e. labeled by circles. Thiscould be observed explicitly in the high AD value of R2,2τ .This confirms the proposed hypothesis which asserts that theCSI is more robust to the signal fluctuations than the RSSI,because, the CSI average magnitude values don’t have thesame temporal stability as the CSI slopes.

V. CONCLUSIONS

Rather than using RSSI fingerprinting, in this article, thefeasibility of utilizing the CSI for localization is presented.This hypothesis intends to improve the accuracy of positioningby utilizing the rich channel information in each subcarrier aswell as the frequency hopping in the LoRa systems. To allowa dynamic study of this approach, an outdoor measurementcampaign is carried out in the area of the Campus Beaulieuin Rennes to estimate the CSI of transmitted LoRa signals

1 1 2 2 3 3Location i

0.750.500.250.000.250.500.751.00

Mag

nitu

de

f0f1f2

f3f4f5

f6f7f8

(a) The normalized slope s̃ifk values.

1 1 2 2 3 3Location i

0.02

0.04

0.06

0.08

0.10

Mag

nitu

de

f0f1f2

f3f4f5

f6f7f8

(b) The CSI average amplitude rifk values.

Fig. 5: A comparison between the two families of the distinctobservables.

S1, 2 S1, 3 S2, 3 S1, 1 S2, 2 S3, 30.000.050.100.150.200.250.300.35

Mag

nitu

de

(a) The NSD Si,j values.

R1, 2 R1, 3 R2, 3 R1, 1 R2, 2 R3, 30.0000.0050.0100.0150.0200.0250.0300.0350.040

Mag

nitu

de

(b) The AD Ri,j values.

Fig. 6: The distance values.

from different locations. Hence, the individual channels fromeach different location have to be appropriately different withone another to achieve significant localization gain. Thisdifference in the CSI for every two locations is done based ontwo aspects, i.e. CSI slope and its average amplitude. Thus,the presented data indicate that CSI slopes are more stableand robust to the signal imperfections than the CSI averageamplitudes. This result demonstrates the high performanceof the CSI-based fingerprinting for positioning than RSSI, aswell as its temporal stability.

In future work, CSI-based fingerprinting could be moreefficient by using more than one gateway i.e., a realistic valuein the near future, to obtain more than one CSI for the instantsingular center frequency. Moreover, it could be improved withclassical machine learning techniques for merging other radioobservables such as RSSI, angle of arrival estimates or thepropagation model.

REFERENCES

[1] M. Centenaro, L. Vangelista, A. Zanella and M. Zorzi, ”Long-rangecommunications in unlicensed bands: the rising stars in the IoT andsmart city scenarios,” in IEEE Wireless Communications, vol. 23, no.5, pp. 60-67, October 2016.

[2] F. Adelantado, X. Vilajosana, P. Tuset-Peiro, B. Martinez, J. Melia-Seguiand T. Watteyne, ”Understanding the Limits of LoRaWAN,” in IEEECommunications Magazine, vol. 55, no. 9, pp. 34-40, Sept. 2017.

[3] L. Mainetti, L. Patrono, A. Secco and I. Sergi, ”An IoT-aware AAL sys-tem for elderly people,” 2016 International Multidisciplinary Conferenceon Computer and Energy Science (SpliTech), Split, 2016, pp. 1-6.

[4] A. H. Sayed, A. Tarighat and N. Khajehnouri, ”Network-based wirelesslocation: challenges faced in developing techniques for accurate wirelesslocation information,” in IEEE Signal Processing Magazine, vol. 22, no.4, pp. 24-40, July 2005.

[5] Z. Li, T. Braun, X. Zhao, Z. Zhao, F. Hu and H. Liang, ”A Narrow-BandIndoor Positioning System by Fusing Time and Received Signal Strengthvia Ensemble Learning,” in IEEE Access, vol. 6, pp. 9936-9950, 2018.

[6] Zafari, F., Gkelias, A., and Leung, K. K. (2019). A Survey of IndoorLocalization Systems and Technologies. IEEE Communications Surveysand Tutorials, 1-1.

[7] B. C. Fargas and M. N. Petersen, ”GPS-free geolocation using LoRa inlow-power WANs,” 2017 Global Internet of Things Summit (GIoTS),Geneva, 2017, pp. 1-6.

[8] Choi, Wongeun and Chang, Yoon-Seop and Jung, Yeonuk and Song,Junkeun. (2018). Low-Power LoRa Signal-Based Outdoor PositioningUsing Fingerprint Algorithm. ISPRS International Journal of Geo-Information.

[9] K. Wu, J. Xiao, Y. Yi, D. Chen, X. Luo and L. M. Ni, ”CSI-BasedIndoor Localization,” in IEEE Transactions on Parallel and DistributedSystems, vol. 24, no. 7, pp. 1300-1309, July 2013.

[10] Q. Song, S. Guo, X. Liu and Y. Yang, ”CSI Amplitude Fingerprinting-Based NB-IoT Indoor Localization,” in IEEE Internet of Things Journal,vol. 5, no. 3, pp. 1494-1504, June 2018.

[11] R. Ghanaatian, O. Afisiadis, M. Cotting and A. Burg, ”Lora DigitalReceiver Analysis and Implementation,” ICASSP 2019 - 2019 IEEEInternational Conference on Acoustics, Speech and Signal Processing(ICASSP), Brighton, United Kingdom, 2019, pp. 1498-1502.

[12] J. -. van de Beek, O. Edfors, M. Sandell, S. K. Wilson and P. O.Borjesson, ”On channel estimation in OFDM systems,” 1995 IEEE 45thVehicular Technology Conference. Countdown to the Wireless Twenty-First Century, Chicago, IL, USA, 1995, pp. 815-819 vol.2.