This short paper explores the usability of Long Range Wide Area Network (LoRaWAN) technology for localization within the context of modern Industry 5.0 wireless networks. Traditional localization methods have often fallen short in providing meaningful accuracy in this domain. Our research addresses this gap by investigating the potential of LoRaWAN for localization, synthesizing key findings and advancements. Two primary contributions are presented: the analysis of two underground LoRaWAN datasets, valuable resources for researchers and practitioners, and the proposal of two innovative -nearest neighbors (-NN) algorithms designed to enhance position estimation accuracy through optimized nearest neighbor selection. By integrating preprocessing strategies with these algorithms, an improvement in accuracy of up to 17% is achieved.
Improving safety protocols within industrial landscapes, especially in hazardous settings where humans
interface with autonomous machinery, is a continual pursuit reshaping operational paradigms. This
development is significantly augmented by the integration of cutting-edge technologies, notably within
the realm of Industry 5.0 [
Wearable devices have gained traction owing to their ability to enhance safety, particularly in
rugged environments [
LoRaWAN, a robust communication solution for IoT applications, including industrial wearables, is
very heavily used in IoT deployments [
This paper delves into the feasibility of harnessing LoRaWAN for localization tasks based on real
life datasets, leveraging both preprocessing and postprocessing methodologies to augment accuracy,
particularly for wearable devices employed in safety-critical scenarios. Building upon antecedent
research that furnished open fingerprinting datasets acquired by the authors in "wild" conditions, this
study progresses by profering two novel -NN-based algorithm modifications customized to optimize
nearest neighbor selection for refined position estimation. Additionally, it reflects on two underground
LoRaWAN-based datasets, augmenting algorithmic refinement and serving as invaluable assets for the
broader research community following work [
The paper assesses LoRaWAN-based localization methodologies, finding conventional techniques wanting, while ML, notably -NN, substantially curtail errors. Diverse preprocessing strategies further impacting the accuracy, with indoor datasets showcasing superior precision, followed by outdoor and subterranean environments. The proposed -NN algorithms and associated strategies yield a marked accuracy enhancement of up to 17%. While not rivaling leading-edge technologies in precision, LoRaWAN evinces promise for applications such as logistics and agriculture, demonstrating commendable accuracy under appropriate conditions.
The rest of the paper is organized as follows. Section 2 provide a brief background information. Further, Section 3 outlines the main k- related modifications. Next, Section 4 depicts selected numerical results based on real-life datsets for both underground and traditional localization cases. The last section concludes the work.
This section draws upon previously gathered open-access LoRaWAN datasets documented in [
Two distinct measurement campaigns were conducted at Brno University of Technology (BUT)
and Politehnica Bucures, ti National University for Science and Technology (UNSTPB). Each campaign
adhered to a consistent two-stage protocol: an "ofline" phase encompassing map preparation and
LoRaWAN LG308 gateway (GW) deployment, succeeded by an "online" phase involving direct data
collection utilizing LoRa Field Test Devices (FTDs) [
This endeavor yielded seven distinct datasets [
• ds1: BUT, Building, GW place – Building, Spacing – 1m, GWs – 7, Indoor; Mean error: 3.4m; • ds2: BUT, Parking (fl. -1), GW place – Building, Spacing – 3m, GWs – 7, Underground; Mean error: 11.6m; • ds3: BUT, Parking (fl. -2), GW place – Building, Spacing – 3m, GWs – 7, Underground; Mean error: 13.3m; • ds4: BUT, Parking (fl. -1) and Parking (fl. -2), Spacing – 2.5m, GWs – 6, Underground; Mean error: 5.6m; • ds5: BUT, Parking (fl. -2) and Parking (fl. -2), Spacing – 2.5m, GWs – 6, Underground; Mean error: 7.2m; • ds6: UNSTPB, Building, GW place – Building, Spacing – 1m, GWs – 9, Indoor; Mean error: 3.3m; • ds7: UNSTPB, Alley, GW place – Building, Spacing – 1m, GWs – 8, Outdoor Mean error: 5.3m.
Initially, localization accuracy was approximated using conventional methodologies such as trilateration,
Weighted Centroid Algorithm (WCA), and -Nearest Neighbors (-NN). Trilateration and WCA, being reliant on
receiver positions, were directly applicable without significant data manipulation. Conversely, -NN necessitated
redundancy exclusion stemming from multiple fingerprints per location. In our baseline -NN implementation,
the "random choice" selection strategy was adopted, wherein one fingerprint per location was randomly chosen
for subsequent calculations. Following redundancy removal, the remaining data was partitioned into training
(20%) and testing (80%) subsets. Additional insights into the application of these localization algorithms can be
found in the seminal work [
Localization error estimation employed the Root Mean Square Error (RMSE) metric, widely utilized for accuracy assessment: = √︁ ∑︀((− ^)2+(− ^)2) , where and denote real coordinates, ^ and ^ represent predicted object coordinates (utilizing Trilateration, WCA, -NN), and denotes the number of MPs to be estimated.
Overall, our previous observation proved that trilateration and WCA exhibit notably low accuracy within a few tens of meters, rendering them relatively inefectual for localization purposes. Conversely, -NN yields markedly superior and generally promising results across most scenarios, particularly for indoor scenarios (ds1 and ds6), boasting an accuracy up to 10 times greater than methods reliant on GW location.
Consequently, -NN with random choice was designated as the baseline, with trilateration and WCA omitted from further comparisons to streamline visualization. The baseline findings revealed an average error of 3.4m for the indoor environment and 5.3m for outdoor scenarios.
In pursuit of heightened localization precision, we conducted experiments exploring various redundancy
reduction techniques essential for -Nearest Neighbors (-NN). In addition to the "random choice" strategy, we
investigated "averaging" (based on averaged fingerprints per location) and "maximum" (relying on fingerprints
with the highest average Received Signal Strength Indicator (RSSI) level) methodologies. As noted in [
Despite surpassing initial expectations, the current localization accuracy estimates by LoRaWAN still fall short
of leading technologies such as Ultra-Wideband (UWB) or Global Positioning System (GPS). Consequently, we
continued our quest for accuracy refinement. The noted in [
Upon scrutinizing the dataset using -Nearest Neighbors (-NN), a discernible observation emerges: the shortest
distances in the distance matrix may correspond to measurement points (MPs) quite distant from the point under
estimation. This phenomenon, extensively elaborated in our prior conference paper dedicated to LoRaWAN-based
outdoor localization [
The Modification 1 strategy proposes to reassess the relevance of selected nearest neighbors based on the Euclidean distance between them and the origin points. Here, one of the corners of the measurement area perimeter serves as the origin point. The rationale behind this approach posits that the nearest neighbor group should ideally reside within the same vicinity. Consequently, the proposition involves computing Euclidean distances from the origin to each neighbor, delineating an approximate localization zone based on mean distance, and discarding outliers, i.e., neighboring points outside this zone.
The Modification 2 follows a similar trajectory, albeit with a slight variation. Instead of computing Euclidean distance to the origin point, it’s calculated from each neighbor to the preliminary estimated value provided by -NN. Consequently, the localization zone is established by one border, extending from 0 to the mean value of Euclidean distances plus a designated allowance.
However, Modification 1 acknowledges certain limitations, particularly its eficacy at low values, predicated on the assumption that the outlier count in the group is fewer than relevant neighbors. Nevertheless, a potential issue arises: the likelihood of a larger outlier group compared to ’valid’ neighbors, i.e., those within the localization zone. To address this, a condition is proposed wherein irrelevant neighbors are eliminated only if their count is three times less than the total neighbor count : ≤ /3.
For the Modification 2 , this work investigates two options: Modification 2.1 employs unstable when the number of the nearest neighbors, based on which the estimated position is calculated, is not replenished after the elimination of outliers. Modification 2.2 uses stable when the number of removed outliers is replenished by the next neighbors from the distance matrix .
To estimate the general gain from the modifications proposed in this work, it is necessary to compare the strategies against the baseline using preprocessing, i.e., averaging redundancy reduction, which has proven to be the most efective. The results for this setup are presented in Figure 1.
The results are summarized in Table 1 show that Modification 2 remains the most successful, with which it was possible to achieve an improvement in localization quality by 17.2% on average across all datasets, while for ds1 and ds5 this number exceeds 25%. Among the datasets, the accuracy improvement exceeds 13% for all of them except ds2 and ds3 – two underground datasets with non-local placement of the GWs.
The charts indicate significant improvements in LoRaWAN localization accuracy, up to 17.2% on average, with the implementation of a simple redundancy reduction technique and proposed algorithms. While underground datasets posed challenges, indoor and outdoor scenarios showed results comparable to Wi-Fi and GPS, respectively, albeit within a few meters of accuracy. This level of precision may sufice for many industrial applications where extreme precision is not critical, including logistics, agriculture, and smart factories.
Basic -NN emerged as the most precise among the tested ML algorithms. Among the proposed accuracy-enhancing strategies, reassessment from the estimated value with stable (Modification 2.2 ) yielded the most significant increase in accuracy, averaging 17.2% with preprocessing. Other strategies also performed well, with an average performance exceeding 15%.
Notable error reductions were observed in indoor (ds1), outdoor (ds7), and certain underground datasets (ds4 and ds5), with over 6% reductions seen in the remaining datasets.
While certain conditions, like a concentration of 7 GWs per 50,000 square meters with uniform distribution, seem necessary for indoor and outdoor scenarios, the implications for underground environments remain uncertain. Although the presented underground datasets exhibited low localization accuracy, it’s unclear if this is representative of the broader environment or specific to the measurement campaign.
Despite proposed accuracy improvements, LoRaWAN-based localization still lags behind leading technologies. Nonetheless, it may find utility in sectors where extreme precision is not essential, leveraging its existing infrastructure for communication in logistics, agriculture, or smart factories.
To draw the intermediate conclusion, we start with the trilateration and WCA methodologies that exhibit ineficiency, yielding an approximate error of 30 meters. Conversely, Machine Learning (ML) algorithms showcase significant error reduction, often by several folds. Notably, -Nearest Neighbors (-NN) emerges as the most precise ML technique, reducing mean localization error by up to a factor of 10 compared to trilateration. Among the -NN extensions proposed in this paper, leveraging reassessment from the estimated position with unstable shows promise.
Baseline -NN, employing random fingerprint selection, is eclipsed by "averaging" and "maximum" preprocessing strategies in spite of redundancy reduction strategy. Of the two, the "averaging" strategy slightly outperforms the "maximum" approach. Optimal strategy selection hinges on dataset signal levels: "averaging" proves more efective for low signal levels, while "maximum" is preferable otherwise.
Notably, superior accuracy is observed in indoor datasets (∼ 2.6 m), followed by outdoor (∼ 4 m) and underground datasets (∼ 5 − 12 m).
In conclusion, this paper represents a culmination of eforts in exploring LoRaWAN’s localization potential. It introduces new underground datasets, contributing to algorithm refinement, and proposes novel -NN-based algorithms aimed at enhancing position estimation accuracy. The integration of these algorithms with redundancy reduction strategies yields a notable accuracy improvement of up to 17%. While not competing with leading precision technologies, LoRaWAN-based localization remains viable for specific applications such as logistics, agriculture, or smart factories, where extreme precision is not paramount.
X 3 Y 5.348
X 5
Y 3.445 X 4 Y 2.548 (c) ds3 (e) ds5
X 5 Y 13.3 X 5 Y 7.187 k-NN (preprocessing) Modification 1.1 (30%) Modification 1.2 (50%) Modification 2.1 (unstable k) Modification 2.2 (stable k) k-NN (baseline) k-NN (preprocessing) Modification 1.1 (30%) Modification 1.2 (50%) Modification 2.1 (unstable k) Modification 2.2 (stable k) k-NN (baseline) The authors gratefully acknowledge funding from European Union’s Horizon 2020 Research and Innovation Programme under the Marie Skłodowska Curie grant agreements No. 813278 (A-WEAR), 956090 (APROPOS), and support grant PN-III-P3-3.6-H2020-2020-0124 from the Romanian Ministry of Research.