Conference PaperPDF Available

Enhanced Reliability of Mobile Robots with Sensor Data Estimation at Edge

Authors:

Abstract

The proliferation of sensing equipment serving an expansive range of applications has led the Internet of Things (IoT) paradigm to cover technologies beyond Wireless Sensor Networks (WSN). Extensive advancement in electronics, communication methods and sensors has made it possible to leverage advanced technologies such as Machine Learning and Prob-abilistic Modeling in resource-constrained embedded systems. These techniques increase reliability and enhance interactions among physical elements of an IoT-based system in which data loss or corruption seems inevitable. However, traditional data estimation and reconstruction methods cannot be directly applied considering the computational limitations at the edge of the network. Therefore, mobile robots would greatly benefit from a resource efficient sensor data recovery procedure, capable of generating near-accurate estimates at the resource-constrained Edge layer. In this paper, we introduce a novel Bayesian filtering-based data reconstruction scheme, with real-time performance and precision for incoming semantic and geometric information from a varied set of sensors to increase reliability of autonomous navigation of mobile robots. Afterwards, we corrupt each stream of observations to validate model performance against a baseline. Furthermore, we also provide benchmark on execution latency, CPU usage and current draw while running the models in a practical setup.
Enhanced Reliability of Mobile Robots
with Sensor Data Estimation at Edge
Victor Kathan Sarker1, Prateeti Mukherjee2and Tomi Westerlund1
1Department of Future Technologies, University of Turku, Turku, Finland
2Department of Computer Science and Engineering, IEM, Kolkata, India
Email: 1{vikasar, tovewe}@utu.fi, 2mukherjeeprateeti01@gmail.com
Abstract—The proliferation of sensing equipment serving an
expansive range of applications has led the Internet of Things
(IoT) paradigm to cover technologies beyond Wireless Sensor
Networks (WSN). Extensive advancement in electronics, commu-
nication methods and sensors has made it possible to leverage
advanced technologies such as Machine Learning and Prob-
abilistic Modeling in resource-constrained embedded systems.
These techniques increase reliability and enhance interactions
among physical elements of an IoT-based system in which data
loss or corruption seems inevitable. However, traditional data
estimation and reconstruction methods cannot be directly applied
considering the computational limitations at the edge of the
network. Therefore, mobile robots would greatly benefit from
a resource efficient sensor data recovery procedure, capable of
generating near-accurate estimates at the resource-constrained
Edge layer. In this paper, we introduce a novel Bayesian filtering-
based data reconstruction scheme, with real-time performance
and precision for incoming semantic and geometric information
from a varied set of sensors to increase reliability of autonomous
navigation of mobile robots. Afterwards, we corrupt each stream
of observations to validate model performance against a baseline.
Furthermore, we also provide benchmark on execution latency,
CPU usage and current draw while running the models in a
practical setup.
Index Terms—Reliability, Edge, Mobile Robot, Sensors, State-
Space Models, Data Estimation, Bayesian Filtering.
I. INTRODUCTION
Autonomous and semi-autonomous mobile robots have
gained popularity over the past few years due to the ad-
vancement in efficient algorithms, and the development of
powerful embedded hardware and sensor technology. Aiding
the implementation of Industry 4.0 objectives, these robots find
a wide variety of use in our daily life, with little to no human
intervention [1]. The support of these units greatly increase the
accuracy and pace of work, reduce manual labor and risk of
injuries, and help perform tasks in a more systematic manner.
In addition, these units can operate in both independent as well
as coordinated manner in highly interconnected systems, with
a centralized or distributed architecture through sharing of data
with other system nodes. With the evolution of the Internet of
Things (IoT), these systems have taken a new dimension as
autonomous and semi-autonomous robots have been developed
to be more interactive and environment-aware entities for inte-
gration in numerous applications. In an industrial setting, these
robots play an important role, working either as individual
units or alongside humans. Examples of these setups include
the use of robots for menial tasks such as sorting, in-factory
carriage and cleaning [2]–[4]. Successful execution of these
duties, however, requires the robots to be aware of surrounding
environment. To achieve this, autonomous robots make use of
a variety of sensors such as proximity and distance sensors
to understand their surroundings and navigate accordingly.
A greater number of sensors result in more data, burdening
the resource-constrained processing units onboard the robot.
Furthermore, data loss is unavoidable in these systems con-
sidering unreliable wireless links and hardware failures at
system nodes [5]. Malfunctioning nodes and communication
breakdowns result in corrupted data transmission, a prevalent
issue in both industrial automation setups as well as research.
Controlled and cleverly designed systems also suffer from data
loss, thereby reducing the statistical rigor in processing and
leading to invalid conclusions [6].
Although robots can be powered by small fuel-powered
engines, most autonomous indoor robots run on batteries. On
account of the limited supply capacity of batteries, the com-
putational unit and other electrical and electronic components
within the robot must be as energy-efficient as possible. In
order to ensure longer run-time of the system, it is crucial
that the algorithms running onboard be of lower complexity
and run efficiently on embedded controllers with inferior
computational resources. Moreover, high functional accuracy
of sensing equipment and low rate of faulty data contribute
to a low processing cycle count since repeated measurements
and analysis are not required. However, factors such as noise,
interference, non-linearity due to temperature variations, me-
chanical errors and faults in circuitry often lead to missing
observations and faulty acquisition. In such circumstances,
approximating these observations and reads based on the
nature of the sensor and previous observations can greatly aid
proper operation of the robot units.
While traditional IoT-assisted robotic systems bear great
potential, they suffer from issues such as inadequate bandwidth
and high communication latency when the volume of data
to be forwarded to cloud servers is large. To overcome
these problems, the use of Edge and Fog gateways plays an
important role for enhancing energy-efficiency and operational
performance of resource-constrained indoor robots [7]. Al-
though the hybrid Edge-Fog-Cloud architecture is an impactful
approach to efficiently offload computationally expensive tasks
from nodes with limited processing capabilities, the data
transmission phase is still susceptible to noise, interference,
and data loss. If a major chunk of the data is lost, the offloading
process will inevitably cause faulty decisions, leading to major
system performance concerns. Therefore, while the benefits of
offloading in IoT ecosystems are abundant, it is important to
account for the hazards and ensure that algorithms are in place
to mitigate data corruption and loss.
In this work, we exploit Machine Learning (ML) techniques
to reconstruct missing sensor data in small autonomous mobile
robots that constitute the resource-constrained Edge layer. In
particular, we employ Bayesian Statistics and the concept of
Hidden Markov Models (HMM) to develop two estimation
models with linear time complexity in the number of time-
steps. Furthermore, we simulate an experimental setup with
2D Light Detection and Ranging (LIDAR), Gyroscope, and
Global Positioning System (GPS) observations for state vector
representation to evaluate performance of proposed models.
We then corrupt the observations and illustrate how different
parameters affect the estimation outcome. The specific contri-
butions of the paper are:
Transform existing mathematical concepts of Recursive
Bayesian State Inference to develop two domain-specific
models for sensor data estimation,
Experiment to validate estimation at the Edge with data
from LIDAR, gyroscope and GPS module, and,
Analyze results and discuss the effect of different param-
eters on results of estimation.
The rest of the paper is arranged as follows: Section II
enumerates related works and state of the art, Section III
describes the proposed method for data estimation at the Edge,
Section IV demonstrates the experimental evaluation of the
proposed method through simulations and presents the results.
Finally, Section V concludes the paper and leaves directions
for future work.
II. RE LATE D WOR K
Many research works exist in the field of missing data
recovery and estimation, however, seldom are those ad-
vanced algorithms suitable for resource-constrained compu-
tational units. Izonin et al. [8] used stochastic integrals to
increase dimensionality of their task in combination with the
AdaBoostRegressor to search for decomposition coefficients.
The authors simulated their model on real data acquired from
an IoT-based air quality assessment system and compared the
performance against existing schemes. While it resulted in a
6% increase in accuracy, the paper disregards the importance
of other parameters such as resource utilization and energy-
efficiency, if run on a resource-constrained Edge device. Ullah
et al. [9] presented a method for localization based on extended
Kalman Filter with Edge computing in WSN. The authors
ran simulations with GPS data at different velocity settings.
However, the execution impact on Edge computation units is
not considered and real-life benchmarks are not performed,
thus limiting the usability when multiple nodes simultaneously
require processing of localization data at the Edge.
Fekade et al. [10] presented a similar approach in which
Probabilistic Matrix Factorization (PMF) is performed within
a group of sensors to detect similarities among the obtained
data points to subsequently split these points into clusters using
the k-means clustering scheme. The metrics for comparison of
their proposed approach against the standard Support Vector
Machine (SVM) and Deep Neural Network (DNN)-based
algorithms include Root Mean Squared Error (RMSE) and
execution time for normalized data. However, other crucial
parameters are not considered and the computation-intensive
process can only serve feasible results on a handful of high-
end devices at the Edge.
Peng et al. [11] proposed an Incremental Space-Time-
based Model (ISTM) to recover missing data points in real-
time. The Incremental Multiple Linear Regression forms the
basis of the ISTM scheme in which the model is continually
updated in accordance with the intermediary data matrix. In
addition, estimations for missing values are drawn considering
nearly historical data along with the observations supplied
by the neighboring sensing equipment. This approach results
in increased computation time. However, given the repeated
amendments to the model, the process bears potential in
serving important tasks related to data imputation such as
outlier detection in sensor data streams.
Lujic et al. [12] suggested an interesting semi-automatic
recursive approach for missing data recovery at the Edge. In
their work, the indices pertaining to gaps in data are marked
prior to starting the recovery cycle which identifies the amount
of missing and faulty data points. This process is repeated
until all missing values have been traversed. Subsequent
functional steps include data analysis to find the right setup of
forecasting methodology, replacing gaps with predicted values,
and checking conditions for the succeeding recovery cycle.
This produces good results in terms of accuracy, however, the
performance and utility of the model could be enhanced with
the use of deeper statistical analytics and ML tools.
An interesting odometry estimation method in 3D LIDAR
scans with Convolutional Neural Networks (CNNs) is sug-
gested by Velas et al. [13]. Here, the sparse data points are first
passed through an encoder to generate 2D matrices, which are
then fed as inputs to the Neural Networks to obtain rotation
probabilities. However, given the complexity of CNNs, the
work lacks the computational efficacy to function in Edge de-
vices. In a different work [14], missing points in the trajectory
of a moving vehicle are recovered through microscopic traffic
flow models, wherein, calibration is done by assigning weights
to known data points based on their proximity to the gaps using
the tri-cube weight function. Extensive comparisons are drawn
with Newell’s, Pipes, Intelligent-Driver Model (IDM) and
Gipps’ car following models, and several curves are plotted
to illustrate the contrasting performance.
Vijayakumar et al. [15] suggests the use of Kalman Filters
to predict missing events in sensor data streams. In their
approach, the filter is applied to SQL-based event processing
systems using the standard implementation in the Bayes++
software library. While the work inspires the proffered models,
the use of already existing libraries hinders domain-specific
progress, and is hence avoided in our work. In addition, the
presented architecture bears no correlation to practical units
in an autonomous or semi-autonomous setting, thereby disre-
garding prevalent problems at the Edge. Furthermore, applying
the classical form of the filter to modern-day problems would
be an outdated approach. Transforming the mathematical con-
cepts to suit the needs of a particular field is crucial in order
to generate meaningful results and to ensure that the solutions
adhere to domain-specific challenges and requirements.
III. DATA ESTIMATION SCHEME
While a multitude of research exists on sensor data retrieval,
handful of these works consider the hardware limitations of
computational units. However, given the promising capabilities
and advantages of Edge computing [16], it is crucial to explore
the consolidation of statistical concepts and ML techniques
to serve real-life applications where system resources are
limited. To this end, the proposed algorithms, albeit based
on fundamental principles of Bayesian filtering, are developed
to suit the needs of the problem at hand, keeping in mind
the computational limitations and power supply constraints in
IoT ecosystems, to facilitate improved safety and reliability of
autonomous operation.
The state vector Ztfor our problem is:
Zt=xtytφtvtT(1)
where, xtand ytrepresent the mapped position of a vehicle
using the data transmitted by the 2-D LIDAR sensor in
real-time [17], φtdenotes the orientation as realized by the
gyroscope, and vtrepresents the velocity calculated from
coordinates provided by the GPS module.
The state transition equations Ztand Ytare defined as:
Zt=f(Zt1) + qt(2)
Yt=HZt+rt(3)
where, qt∼ N(qt|0, Q)is the process noise, rt∼ N(rt|0, R)
is the observation noise, and the non-linear state transition
function fis:
f
x
y
φ
v
=
vtcosφ
ytsinφ
ωt
at
where, ω=
dt is the angular velocity and ais the acceleration.
The process noise covariance Qand the observation noise
covariance Rare:
Q=
σ2
x0 0
0σ2
y
σ2
φ0
0 0 σ2
v
, and, R=σ2
rx 0
0σ2
ry
where, the emission matrix His described as
H=1000
0100
Markovian representations summarize the data obtained
from past and present observations for future predictions, in
the best estimate of the current state vector. As stated by
Kalman [18], the difference between the predicted value from
this estimate at the current state and the next available observa-
tion is orthogonal to earlier observations, resulting in a simple
recursive algorithm for calculation of likelihood in Gaussian
processes. Furthermore, in the absence of observations, the
prediction means in Kalman Filter models are mathematically
equipped to serve as a plug-in replacement for the lost data
points since they represent the Bayes’ optimal solution [19].
Considering the nature of our problem and the computational
efficiency of state space models coupled with the robustness of
Kalman Filter against statistical noise and other inaccuracies,
we formulate two models based on Extended Kalman Fil-
ter (EKF) and Cubature Kalman Filter (CKF) for the proposed
system state.
1) EKF-based Model: The non-linearity of EKF invalidates
the presumed Gaussian nature of the Process and Measurement
models in regular Kalman Filters. The EKF approximates
the non-linear model by a local linear model obtained
from a first order Taylor expansion around the current
estimate, subsequently applying the filtering process to this
approximation. The matrix of partial derivatives for the
proposed scheme is:
Fz
x
y
φ
v
=
10vtsinφ tcosφ
0 1 vtcosφ tsinφ
0 0 1 0
0 0 0 1
With initial mean (m0), initial covariance (P0) and obser-
vation (yk) as input, we transform the Prediction step and
introduce the Imputation step as described in Algorithm 1.
2) CKF-based Model: While EKF is a powerful mathe-
matical tool, it might not produce satisfactory results when
the system uncertainties are large. Short signal interruptions
may result in a quickly diverging EKF, thereby hindering the
performance of the proposed model in urban environments.
Therefore, we make use of the Cubature Rule to numerically
approximate the moment integrals in Bayesian filtering, while
completely preserving the second-order information about the
state that is contained in the sequence observations [20]. The
Spherical Cubature Approximation for Gaussian Integrals [21]
states that for any function h,
Zh(x)N(x|m, P )1
2n
2n
X
n=1
h(m+i)
where, Lis the Cholesky decomposition of Pk1, and,
ζi=(nei;i= 1...n
nein;i=n+ 1...2n
Algorithm 1 EKF-based Model
Input: m0, P0,{yk}T
k=1
1: for k= 1,2,3, . . . , T do
2: Prediction Step
3: Solve for m
kand P
k
4: m
k=f(mk1)
5: P
k=Fz(mk1)Pk1FT
z(mk1) + Q
6: If ykis not missing
7: Update Step
8: vk=ykHmk
9: Sk=H P
kHT+R
10: Kk=P
kHTS1
k
11: mk=m
k+KkVk
12: Pk=P
kKkSkKT
k
13: Else
14: Imputation Step
15: mk=m
k
16: Pk=P
k
17: end for
Since the problem scenario is not altered, the state vector
remains the same as described earlier in (1), as do the state
transitions (2) and (3).
m
k=Zf(xk1)N(xk1|mk1, Pk1)dxk1
1
2n
2n
X
n=1
f(mk1+Lk1ζi)(4)
P
k=Z(f(xk1)m
k)(f(xk1)m
k)T
N(xk1|mk1, Pk1)dxk1+Q
1
2nX[f(mk1+Lk1ζi)m
k]
[f(mk1+Lk1ζi)m
k]T+Q(5)
For the CKF-based model, the Prediction Mean m
kand
Prediction Covariance P
kare described as in (4) and (5),
respectively. These result in Algorithm 2 for the CKF-based
model towards missing data estimation.
IV. EXP ER IM EN T AN D RES ULTS
We simulated the proposed algorithms on the aforemen-
tioned state transition model with state trajectories sampled
for 1000 time steps at 10 ms intervals. For the experiments
in this work, the algorithms are implemented using Python
programming language and executed on the Raspberry Pi 3B+,
Raspberry Pi 4B, and the Aaeon Intel Up Gateway [22], [23]
which serve as the Edge computation unit. In addition, we
tested the performance against computers with Intel Core i5-
5200U and Intel Core i7-4770 processors, both running the
Microsoft©Windows 10 Pro. The comparison of their per-
formance in terms of execution latency, CPU utilization and
average current draw is listed in Table I. The execution latency
Algorithm 2 CKF-based Model
Input: m0, P0,{yk}T
k=1
1: for k= 1,2,3, . . . , T do
2: Prediction Step
3: Solve for χ(i)
k1, m
kand P
k
4: χ(i)
k=mk1+Lk1ζi;i= 1...2n
5: ˆχ(i)
k1=f(χ(i)
k1)
6: m
k=1
2nP2n
i=1 ˆχ(i)
k1
7: P
k=1
2nP2n
i=1( ˆχ(i)
k1m
k)(ˆχ(i)
k1m
k)T+Q
8: If ykis not missing
9: Update Step
10: vk=ykHmk
11: Sk=H P
kHT+R
12: Kk=P
kHTS1
k
13: mk=m
k+KkVk
14: Pk=P
kKkSkKT
k
15: Else
16: Imputation Step
17: mk=m
k
18: Pk=P
k
19: end for
denotes the total time an algorithm takes on a specific proces-
sor or platform, and CPU utilization denotes the percentage
of total available processor time available.
We benchmark our EKF and CKF-based models with re-
spect to a baseline model that uses only the unfiltered sensor
data. Performance of our proposed models with data corruption
of 080% with a 10% increment per step is shown in Table II.
For each model, we report the mean and standard deviation
of the RMSE between the actual states and the filter means
from 10 independent runs. The trends in RMSE Mean ±
Standard Deviation (SD) for proposed models and the baseline
is depicted in Figure 1. The observed trends are indicative of
the fact that even at 80% data loss, the RMSE Mean and
Standard Deviations for the proposed models remain lower
than the baseline. Thus, both our models mitigate the effect
of corruption significantly better than the baseline model.
We also tested the proposed models by varying the an-
gular velocity, acceleration, and parameters of process noise
covariance. We observed the effects of these variations on the
Baseline and the proposed EKF and CKF-based models. In
particular, we are interested in the RMSE Mean ±SD plots
for each model upon changing the aforementioned parameters
within the corresponding acceptable range, as illustrated by the
error plots in Figure 2. For the simulations, we maintain the
reference conditions listed in Table III. To study the effects
of varying a single parameter, the value of every parameter
excluding the one is set as stated the table.
The error plots in Figure 2 suggest that for every parameter
varied within their admissible limits, the proposed models
perform approximately five times better than the baseline
and remain constant on expectation in most cases. Although
increasing process noise parameters σxand σycause a gradual
TABLE I
COMPARISON OF EXECUTION LATENCY, CPU UTILIZATION AND CURRENT DRAW FOR THE PROPOSED ALGORITHMS.
Platform Operating Algorithm Execution Latency CPU Utilization Avg. Current
/ CPU System Based on (ms) (%) (A)
Min. Max. Avg. Min. Max. Avg. Idle Active
Raspberry Pi 3B+ Raspbian Buster EKF 1.159 1.194 1.167 25.00 25.80 25.36 0.49 0.71
(ARMv8 Cortex-A53) CKF 2.196 2.214 2.203 21.30 25.50 25.16 0.49 0.66
Raspberry Pi 4B Raspbian Buster EKF 0.875 0.889 0.880 24.90 25.80 25.27 0.55 0.57
(ARMv8 Cortex-A72) CKF 1.653 1.709 1.675 24.80 25.60 25.08 0.55 0.59
Aaeon UP-GWS01 Ubuntu 16.04 LTS EKF 2.545 2.604 2.572 34.60 49.10 39.20 0.78 0.85
(Intel® x5-Z8350) CKF 5.598 5.849 5.804 26.20 52.80 47.40 0.78 0.87
Core i5-5200U Windows 10 Pro EKF 0.312 0.390 0.350 23.60 46.70 25.64 Untested
CKF 0.531 0.609 0.576 23.00 26.50 24.77
Core i7-4770 Windows 10 Pro EKF 0.203 0.265 0.250 10.10 14.60 12.09 Untested
CKF 0.875 0.889 0.880 24.90 25.80 25.27
TABLE II
RMSE OF DATA ESTIMATION MODELS DUE TO CORRUPTION. THE FI RST
ROW WI TH 0 % ERROR INDICATES ORIGINAL,UNCO RRU PT ED DATA.
Mean ±SD
Corrupted Baseline EKF-based CKF-based
Data (%) Model Model Model
0 0.228 ±0.005 0.037 ±0.003 0.038 ±0.003
10 0.226 ±0.006 0.041 ±0.004 0.041 ±0.004
20 0.226 ±0.005 0.047 ±0.003 0.048 ±0.004
30 0.226 ±0.006 0.054 ±0.007 0.055 ±0.008
40 0.225 ±0.007 0.056 ±0.004 0.057 ±0.006
50 0.226 ±0.006 0.069 ±0.008 0.071 ±0.009
60 0.226 ±0.004 0.079 ±0.013 0.081 ±0.014
70 0.228 ±0.005 0.104 ±0.021 0.110 ±0.023
80 0.224 ±0.006 0.122 ±0.034 0.124 ±0.034
rise in RMSE trends for the proposed models, the overall
curves for EKF and CKF-based models still remain well
below the baseline curve, as seen in Figure 2c and Figure 2d,
therefore outperforming the baseline in each case.
The estimated trajectories for the proposed and baseline
models are shown in Figure 3 for uncorrupted data (prior
to forced manual corruption) and with heavily corrupted
Fig. 1. Mean RMSE (bold line) ±standard deviation (shaded area) compared
against baseline for increasing % of missing data.
TABLE III
REFERENCE CONDITIONS FOR SIMULATIONS IN FIGURE 2. WHEN
VARYIN G A PARTI CU LAR PA RAM ET ER,E VE RY OTHE R PARA ME TER
MAINTAINS THE VALUE LISTED IN THE TABL E
Corrupted a ω σxσyσφσvσrx σry
Data (%)
30 0.2 1.3 0.06 0.1 0.2 0.1 0.45 0.50
data (forced 80% data loss). Figure 3 clearly shows that
the proposed EKF and CKF-based data recovery models
outperform the baseline throughout the experiment. At first
the proposed models perform approximately eight times better
than the baseline. Although a rising trend is observed for
increasing corruption of sensor data, the proposed models still
perform about 1.7 times better than the baseline at maximum
corruption, when 80% of incoming data is distorted. It is
important to note that even when the RMSE for the proposed
data recovery models is calculated on the most corrupt data set,
the RMSE for the baseline is worse despite being calculated
on the original data where the observations are intact. Thus,
the proposed models exhibit robust performance in the context
of missing data estimation at the Edge.
V. CONCLUSION AND FUTURE WORK
With rapid progress in electronics, sensor technology and
embedded systems, automation has grown significantly owing
to increased productivity, improved robustness and consistent
progress in the field. Autonomous mobile units are capable of
performing numerous tasks with little to no human supervision
using a variety of sensors for contextual awareness, naviga-
tion and interaction. However, as many robots are resource-
constrained, Edge computing can play an important role by
moving the computationally heavy processing from a robot to
the Edge. This results in transferring a large amount of data
between the mobile robots and Edge which is susceptible to
data corruption or loss during transmission due to interference,
congestion and unstable network.
Erroneous data acquisition can lead to erroneous au-
tonomous operation and cause personal injury or damage to
property. In this paper, we proposed a novel method for esti-
mating sensor data in real-time with reasonable operational ac-
(a) Error bars when varying ω(b) Error bars when varying a
(c) Error bars when varying process noise covariance σx(d) Error bars when varying process noise covariance σy
(e) Error bars when varying process noise covariance σφ(f) Error bars when varying process noise covariance σv
(g) Error bars when varying observation noise covariance σrx (h) Error bars when varying observation noice covariance σry
Fig. 2. Standard Error Representations when varying individual parameters of proposed models.
(a) With no corruption.
(b) With 80% corruption.
Fig. 3. Estimated trajectories for observed data points.
curacy. We presented two different models based on Recursive
Bayesian filtering. We experimented the proposed methods on
sensor data streams by introducing varying amounts of noise
and erasure to mimic data corruption and loss, respectively.
We used localization coordinates calculated using data from
a LIDAR, GPS module and orientation information provided
by a gyroscope. The obtained results clearly suggest that our
proposed method works well for estimating corrupted data,
even when up to 80 % of incoming data is deemed unusable.
Furthermore, experiments on different platforms achieved low
execution latency, consumed minimal CPU resources and drew
moderate current, yielding strong candidacy for sensor data
estimation on computationally limited units. In future, we will
take a closer look at multi-sensor systems and coordinated
prediction methodologies, both on-board robots and at the
Edge, to improve data prediction in more compromised cir-
cumstances. Furthermore, use of Monte Carlo methods such
as particle filtering as a Bayesian statistical inference tool
would be an intuitive approach for missing data imputation
in resource-constrained devices.
ACKNOWLEDGMENT
This research work is supported by Academy of Finland
(Grant No. 328755).
REFERENCES
[1] M. A. K. Bahrin, M. F. Othman, N. H. N. Azli, and M. F. Talib. Industry
4.0: A Review on Industrial Automation and Robotic. Jurnal Teknologi,
78, 2016.
[2] C. Li, Y. Ma, S. Wang, and F. Qiao. Novel industrial robot sorting
technology based on machine vision. In 9th International Conference
on Modelling, Identification and Control (ICMIC), pages 902–907, 2017.
[3] A. M. Kabir, K. N. Kaipa, J. Marvel, and S. K. Gupta. Automated Plan-
ning for Robotic Cleaning Using Multiple Setups and Oscillatory Tool
Motions. IEEE Transactions on Automation Science and Engineering,
14(3):1364–1377, 2017.
[4] A. Vick, D. Surdilovic, A. K. Dr¨
ager, and J. Kr¨
uger. The Industrial
Robot as Intelligent Tool Carrier for Human-Robot Interactive Artwork.
In The 23rd IEEE International Symposium on Robot and Human
Interactive Communication, pages 880–885, 2014.
[5] Y. Shao and Z. Chen. Reconstruction of big sensor data. In 2nd IEEE
International Conference on Computer and Communications (ICCC),
pages 1–6, 2017.
[6] H. Kang. The Prevention and Handling of The Missing Data. Korean
J Anesthesiol, page 402–406, 2013.
[7] V. K. Sarker, J. P. Queralta, T. N. Gia, H. Tenhunen, and T. Westerlund.
Offloading SLAM for Indoor Mobile Robots with Edge-Fog-Cloud
Computing. In 1st International Conference on Advances in Science,
Engineering and Robotics Technology, pages 1–6, May 2019.
[8] I. Izonin, N. Kryvinska, R. Tkachenko, and K. Zub. An Approach
towards Missing Data Recovery within IoT Smart System. Procedia
Computer Science, 155:11–18, 2019.
[9] I. Ullah, S. Qian, Z. Deng, and J. Lee. Extended Kalman Filter-
based Localization Algorithm by Edge Computing in Wireless Sensor
Networks. Digital Communications and Networks, 2020.
[10] B. Fekade, T. Maksymyuk, M. Kyryk, and M. Jo. Probabilistic Recovery
of Incomplete Sensed Data in IoT. IEEE IoT Journal, PP:1–1, 2017.
[11] T. Peng, S. Sellami, and O. Boucelma. IoT Data Imputation with
Incremental Multiple Linear Regression. Open Journal of Internet of
Things (OJIOT), 2019.
[12] I. Lujic, V. De Maio, and I. Brandic. Adaptive Recovery of Incomplete
Datasets for Edge Analytics. In IEEE 2nd International Conference on
Fog and Edge Computing (ICFEC), pages 1–10, 2018.
[13] M. Velas, M. Spanel, M. Hradis, and A. Herout. CNN for IMU Assisted
Odometry Estimation Using Velodyne LiDAR. In IEEE International
Conference on Autonomous Robot Systems and Competitions, 2018.
[14] C. Sazara, R. V. Sazara, and M. Cetin. Offline reconstruction of missing
vehicle trajectory data from 3D LIDAR. CoRR, 2017.
[15] N. Vijayakumar and B. Plale. Prediction of Missing Events in Sensor
Data Streams Using Kalman Filters. In 1st Int’l Workshop on Knowledge
Discovery from Sensor Data, pages 1–9, 2008.
[16] G. Premsankar, M. Di Francesco, and T. Taleb. Edge computing for
the internet of things: A case study. IEEE Internet of Things Journal,
5(2):1275–1284, 2018.
[17] J. Zhang and S. Singh. LOAM: Lidar Odometry and Mapping in Real-
time. In Robotics: Science and Systems X, 2014.
[18] R. E. Kalman. A New Approach to Linear Filtering and Prediction
Problems. Transactions of the ASME–Journal of Basic Engineering,
82(Series D):35–45, 1960.
[19] J. Durbin and S. J.Koopman. Time Series Analysis by State Space
Methods. Oxford University Press, 2 edition, 2012.
[20] Simo Srkk. Bayesian Filtering and Smoothing. Cambridge University
Press, USA, 2013.
[21] Haran Arasaratnam and Simon Haykin. Cubature Kalman Filters.
Automatic Control, IEEE Transactions on, 54:1254 – 1269, 07 2009.
[22] Raspberry Pi Foundation. Products. [Online] Available: https://www.
raspberrypi.org/products/. Accessed: Mar. 20, 2020.
[23] Aaeon. UP-GWS01. [Online] Available: https://www.aaeon.com/en/p/
tiny-gateway-system- with-upboard. Accessed: Apr. 5, 2020.
... Active research areas in TIERS include multi-robot coordination [1], [2], [3], [4], [5], swarm design [6], [7], [8], [9], UWB-based localization [10], [11], [12], [13], [14], [15], localization and navigation in unstructured environments [16], [17], [18], lightweight AI at the edge [19], [20], [21], [22], [23], distributed ledger technologies at the edge [24], [25], [26], [27], [28], [29], edge architectures [30], [31], [32], [33], [34], [35], offloading for mobile robots [36], [37], [38], [39], [40], [41], [42], LPWAN networks [43], [44], [45], [46], sensor fusion algorithms [47], [48], [49], and reinforcement and federated learning for multi-robot systems [50], [51], [52], [53]. ...
Article
Full-text available
The Extended Kalman Filter (EKF) has received abundant attention with the growing demands for robotic localization. The EKF algorithm is more realistic in non-linear systems which has an autonomous white noise in both the system and the estimation model. Also, in the field of engineering, most systems are non-linear. Therefore, the EKF attracts more attention than the Kalman Filter (KF). In this paper, we propose an EKF-based localization algorithm by edge computing, and a mobile robot is used to update its location concerning the landmark. This localization algorithm aims to achieve a high level of accuracy and wider coverage. The proposed algorithm is helpful for the research related to the use of EKF localization algorithms. Simulation results demonstrate that, under the situations presented in the paper, the proposed localization algorithm is more accurate compared with the current state-of-the-art localization algorithms.
Article
Full-text available
Today, the fast development of the hardware for the Internet of things systems creates conditions for the development of IoT based Services of various purposes. The imperfect systems of collecting, aggregation and the transmission of large volumes of various types of data, fixed by sensors of IoT devices, as well as possible failures of the latter, cause the occurrence of missing data problems. The paper proposes a regression approach to solving the task of missed data recovery. The authors have developed a composition of the method of the missing data recovery for IoT systems based on the use of the Ito decomposition and the AdaBoost algorithm. We transform each data vectors by using Ito decomposition, and searching the coefficients of this decomposition scheme using AdaBoost algorithm. Increasing the dimensionality of the input space due to the use of the second-degree Ito decomposition scheme, as well as its high approximation properties, allowed to increase the accuracy of filling the missed values by the AdaBoost regressor at more than 6% (MAPE). It has been established that the developed method provides the highest accuracy of filling missed data based on all other indicators (MAE, RMSE, SMAPE) among the considered regression methods.
Conference Paper
Full-text available
Indoor mobile robots are widely used in industrial environments such as large logistic warehouses. They are often in charge of collecting or sorting products. For such robots, computation-intensive operations account for a significant percentage of the total energy consumption and consequently affect battery life. Besides, in order to keep both the power consumption and hardware complexity low, simple micro-controllers or single-board computers are used as onboard local control units. This limits the computational capabilities of robots and consequently their performance. Offloading heavy computation to Cloud servers has been a widely used approach to solve this problem for cases where large amounts of sensor data such as real-time video feeds need to be analyzed. More recently, Fog and Edge computing are being leveraged for offloading tasks such as image processing and complex navigation algorithms involving non-linear mathematical operations. In this paper, we present a system architecture for offloading computationally expensive localization and mapping tasks to smart Edge gateways which use Fog services. We show how Edge computing brings computational capabilities of the Cloud to the robot environment without compromising operational reliability due to connection issues. Furthermore, we analyze the power consumption of a prototype robot vehicle in different modes and show how battery life can be significantly improved by moving the processing of data to the Edge layer.
Conference Paper
The Internet of Things (IoT) has attracted significant attention from both academia and industry, thanks to applications such as smart cities, smart buildings and intelligent traffic management. These systems rely on data, collected from IoT devices, that are sent to the cloud for analytics. Data are either used for near real-time decisions or stored for long-term analysis. However, in highly distributed IoT systems, missing or invalid data may appear because of different reasons including sensor failures, monitoring system failures and network failures. Analyzing incomplete data sets can lead to inaccurate results and imprecise decisions, with negative effects on the target systems. Also, due to the increasing size of such systems and the consequently increasing amount of data generated from sensors,recovery of incomplete datasets for analytics on the cloud is often infeasible, due to the limited bandwidth available and the strict latency constraints of IoT applications. We propose a novel semi-automatic recursive mechanism for recovery of incomplete datasets on the edge that is closer to the source of data. This mechanism enables efficient recovery of incomplete datasets employing different forecasting techniques for multiple gaps, based on user specifications. We evaluate our approach on datasets coming from the context of smart buildings and smart homes. The experimental results show that our approach is able to identify multiple gaps, then recover incomplete datasets, decreasing forecasting error by up to 82.68%, and reducing running time by up to 52.38%.
Article
The amount of data generated by sensors, actuators and other devices in the Internet of Things (IoT) has substantially increased in the last few years. IoT data are currently processed in the cloud, mostly through computing resources located in distant data centers. As a consequence, network bandwidth and communication latency become serious bottlenecks. This article advocates edge computing for emerging IoT applications that leverage sensor streams to augment interactive applications. First, we classify and survey current edge computing architectures and platforms, then describe key IoT application scenarios that benefit from edge computing. Second, we carry out an experimental evaluation of edge computing and its enabling technologies in a selected use case represented by mobile gaming. To this end, we consider a resource-intensive 3D application as a paradigmatic example and evaluate the response delay in different deployment scenarios. Our experimental results show that edge computing is necessary to meet the latency requirements of applications involving virtual and augmented reality. We conclude by discussing what can be achieved with current edge computing platforms and how emerging technologies will impact on the deployment of future IoT applications.
Article
Reliable data delivery in the Internet of Things (IoT) is very important in order to provide IoT-based services with the required quality. However, IoT data delivery may not be successful for different reasons, such as connection errors, external attacks, or sensing errors. This results in data incompleteness, which decreases the performance of IoT applications. In particular, the recovery of missing data among the massive sensed data of the IoT is so important that it should be solved. In this paper, we propose a probabilistic method to recover missing (incomplete) data from IoT sensors by utilizing data from related sensors. The main idea of the proposed method is to perform probabilistic matrix factorization (PMF) within the preliminary assigned group of sensors. Unlike previous PMF approaches, the proposed model measures the similarity in data among neighboring sensors and splits them into different clusters with a K-means algorithm. Simulation results show that the proposed PMF model with clustering outperforms support vector machine (SVM) and deep neural network (DNN) algorithms in terms of accuracy and root mean square error. By using normalized datasets, PMF shows faster execution time than SVM, and almost the same execution time as the DNN method. This proposed incomplete data-recovery approach is a promising alternative to traditional DNN and SVM methods for IoT telemetry applications.