IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO.

3, JUNE 2024 727

Received 22 February 2024; revised 5 April 2024; accepted 17 April 2024. Date of publication 22 April 2024; date of current version 27 May 2024.
Digital Object Identifier 10.1109/OJAP.2024.3391835

A Machine Learning Approach to Wireless

Propagation Modeling in Industrial Environment
MOHAMMAD HOSSEIN ZADEH , MARINA BARBIROLI , AND FRANCO FUSCHINI
Department of Electrical, Electronic, and Information Engineering “G. Marconi,” University of Bologna, 40126 Bologna, Italy
CORRESPONDING AUTHOR: M. H. ZADEH (e-mail: [email protected])
This work was supported in part by the European Union under the Italian National Recovery and Resilience Plan (PNRR) of NextGenerationEU, a partnership on
“Telecommunications of the Future” (PE00000014—Program “RESTART”)—Project Industrial Networks, and in part by the EU COST Action INTERACT
(Intelligence-Enabling Radio Communications for Seamless Inclusive Interactions) under Grant CA20120.

ABSTRACT Wireless channel properties in industrial environments can differ from residential or office
settings due to the considerable impact of heavy machinery that triggers intricate multipath propagation
effects and strong blockage effects. Previous investigations on wireless propagation in factories often
consisted of empirical models, that is simple analytical formulas based on measurement data. Unfortunately,
they usually lack in flexibility, since they seldom include geometrical parameters describing the industrial
scenario and therefore turn out reliable only in industrial scenarios sharing the same propagation
characteristics as those where the measurements were performed. In response to this limitation, this article
harnesses the power of Machine Learning to model propagation markers like path loss, shadowing, and
delay spread in the industrial environment. By employing Machine Learning techniques, the objective is to
achieve flexibility and adaptability in modeling, enabling the system to effectively generalize across diverse
industrial scenarios. The proposed model relies on a combination of predictive algorithms, including a
linear regression model and a Multi-Layer Perceptron, working collaboratively to model the relationship
between the considered propagation markers and input features like frequency and machine size, spacing,
and density. Results are in fair overall agreement with previous studies and highlight some trends about
the sensitivity of the propagation parameters to the considered input features.

INDEX TERMS Delay spread, industrial environment, machine learning, path loss, shadowing, wireless
propagation channel.

I. INTRODUCTION properties that make wireless propagation somehow peculiar

W IRELESS communication in industrial manufactur-

ing has advanced in recent years, inspired by the
growing popularity of the smart factory idea, where data
and different from other wireless channels. In general,
industrial environments are larger and/or taller than office
or residential buildings and are divided into large depart-
exchange between machines and controllers (either human ments with high ceilings where the machinery is usually
or unmanned) is continually carried out not only to arranged according to some roughly regular layout, with
improve the whole production process [1], [2] but also to long aisles in between for movement of people and transport
trigger safer work procedures. Wireless solutions can be of materials [4]. Industrial building structures are often
effectively employed for many applications in industrial made of reinforced concrete to safely support vehicles and
environments, like monitoring and surveillance, smart meter- heavy machinery. Industrial equipment is usually made of
ing, cable replacement, remote control, and autonomous metal, i.e., highly reflective, with size, density, and spatial
robotics [1], [3]. distribution that can change significantly case by case. These
The performance of wireless systems in the industrial characteristics of the industrial environment lead to fading
environment depends on equipment design and the properties Path Loss (PL), obstruction level (shadowing), and Delay
of the wireless channel. Industrial scenarios have specific Spread (DS) different from other indoor scenarios [5].


c 2024 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 728

Several empirical models for PL, shadowing, and DS The remainder of this paper is organized as follows:
were achieved based on measurements in various industrial Section II contains recent work on ML-based channel
environments [6], [7], [8], [9], [10], [11], [12], [13], [14], modeling. Section III details the specifications of the
[15]. Empirical models often consist of user-friendly, closed- industrial environment together with the system setup and
form expressions mainly aimed at catching the average industrial wireless channel characterization. In Section IV,
dependence of major propagation markers (like PL) on the machine learning-based approach to channel modeling
some simple link parameters (like distance or frequency). is presented. In Section V numerical results are presented
The measurement campaigns required for deriving these and discussed. Finally, Section VI concludes the paper with
models usually mean a lot of cost in time and manpower. main takeaways and ideas for future work.
Moreover, the nature of empirical models implies that they
are only suitable and fairly reliable for environments sharing
the same propagation characteristics as those where the II. RELATED WORK
measurements were performed, which limits their flexibility Industrial wireless propagation has been addressed in some
and adaptability. previous works, mostly employing experimental investiga-
Ray Tracing (RT) is an alternative channel modeling tech- tions carried out in industrial environments [6], [7], [8],
nique [16], [17] that tracks all potential optical rays between [9], [10], [11], [12], [13], [14], [15]. Measurements are
a transmitter and a receiver for a given number of permitted usually exploited to tune the path-loss exponent (PLE) in
wave-matter interactions and starting from a site-specific simple PL formulas [6], [8], [9], [10], [11], [12], [13], [15]
description of both the environment and the antennas. Then, and/or to estimate other channel parameters like shad-
the computation of the rays’ (EM field) contributions is based owing level [6], [8], [9], [10], [11], [12], [13], [15] or
on geometrical optics (GO), uniform theory of diffraction DS [7], [8], [11], [12], [13], [14], [15]. RT simulations have
for diffraction (UTD), and can also include diffuse scattering been also relied on for similar assessments [27], [28], [29],
to some extent [18]. PL data, large- and small-scale fading, [30]. According to the main outcomes resulting from a
time/frequency/angular dispersion, and optical visibility can literature survey and summed up in Table 1, investigations
be investigated through RT in almost every propagation aimed at industrial propagation modeling have been mostly
scenario and wireless application [19], [20], [21], as RT is carried out in Line of Sight (LOS) conditions at frequencies
a general-purpose approach. However, RT simulations often up to a few tens of GHz. Not surprisingly, in LOS conditions
lead to large, sometimes prohibitive, computational costs and PLE is equal to about 2, or even sometimes lower,1 and (σ )
simulation time [22]. is limited to a few dBs. Both these parameters are likely
There has been increasing interest in Machine Learning to be greater in NLOS cases. DS in general ranges from a
(ML) algorithms in electromagnetic propagation over the few to some tens of nsec. The sensitivity of the industrial
last years [23], [24], [25], [26]. The use of ML techniques channel coefficients to parameters like frequency and clutter
in wireless channel modeling can be extremely beneficial, density does not clearly come out from the survey, as both
as ML-based methods can in principle learn the complex measurements and RT models are inherently case-specific
relationship between the channel parameters and the proper- and the available dataset in Table 1 is likely to be too small
ties of the propagation environment. Moreover, they can be to highlight consolidated trends. The limited consistency
aimed at both average and case-specific evaluations. Another and/or representativeness of results collected from different
significant advantage of ML models is the inference speed: case studies is often a common, general drawback of any
while the training phase may be computationally expensive deterministic, site-specific approach.
(often due to very large datasets), querying the output of Over the last years, channel modeling has been heavily
a trained model is typically computationally light. In this relying on ML techniques. In several cases, ML has been
framework, an ML approach to channel modeling can turn considered for PL prediction, mainly in rural, suburban,
out very flexible, as it can in the end contribute to the and urban outdoor scenarios [31], [32], [33], [34], [35],
achievement of simple, closed-form formulas (like empirical [36], [37], [38], [39], or in other particular cases [40], [41].
models), or result in a black box describing some complex Regarding indoor environments, an ML-based PL regression
input-output relationship (like RT models). model was proposed in [42] for an office scenario, whereas
This paper presents a hybrid scheme integrating RT an ML approach to the classification of indoor spaces inside
simulation tools with ML techniques for predictive chan- a university campus is discussed in [43].
nel modeling in the industrial environment. In particular, To the best of the authors’ knowledge, specific inves-
RT enables the generation of reliable, synthetic, tabular tigations on the industrial wireless channel through ML
propagation data sets that are then used for the train- is basically limited to [44], where ML is fundamentally
ing of different ML models. The ML tools are here exploited to cluster the power-angle-delay profiles collected
conceived for the assessment of PL, shadowing standard
1 PLE values lower than two correspond to some guiding effect experi-
deviation (σ ), and DS based on the value of few, simple
enced by the wireless signal. In industrial environments, this may occur
parameters like communication frequency and machine between the floor and the ceiling of industrial buildings, and/or along the
density. aisles often present between machinery.
729 IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO. 3, JUNE 2024

TABLE 1. Related studies. (a, b, c, d) refer to specific working conditions in each reference.

from a measurement campaign carried out inside an indus- PLE, DS, etc.) and the major features of the industrial
trial plant based on some common feature (like power or scenario, which is instead the main issue tackled in this work.
DS), and to associate then each cluster with a representative According to an assessment framework already experienced
power-delay-profile. Therefore, the goal in [44] is not to in previous investigations [36], [38], RT simulations at
catch the relationship between propagation markers (like different frequencies and on different realizations of the
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 730

FIGURE 1. Two extreme cases of industrial maps - High Density (MS = 4, SP = 2, T = 0.1, MD = 0.4) and Low Density (MS = 2, SP = 4, T = 0.5, MD = 0.05). Blue and Red points
represent the TXs and the RXs, respectively.

industrial environment are leveraged to gather the data Based on the values of MS, SP, and T, the Machine Density
required to train and test artificial neural networks (ANN) (MD) can be also estimated as:
to learn the dependence of PL, shadowing, and DS on input  2
MS
features like the communication frequency and the machine MD  (1 − T) · (1)
density. The trained ML networks are then effectively used to MS + SP
complement and further extend the propagation data achieved B. TRANSMITTERS, RECEIVERS, AND FREQUENCIES
from RT. Some simple, analytical, parametric formulas for In this work, transmitters (TXs) are assumed to lie on
PLE, σ , and DS are finally drawn from such a larger dataset, the shed wall, in four different possible positions (blue
which includes RT simulation data and ML extrapolated data. points in Fig. 1) at a height of 3 meters above the ground
(i.e., slightly higher than MH). For each map, a multitude
III. ASSESSMENT FRAMEWORK of receivers (RXs) have been then spread throughout the
A. INDUSTRIAL ENVIRONMENT REPRESENTATION industrial area at 1 m above ground along the aisles
An industrial shed with dimensions of 100 m × 100 m and between machines, representing the position of automated
a uniform height of 10 m, where machines are all simply guided vehicles (AGVs) often observed in manufacturing
represented as metal boxes with te same machine size (MS), and automation scenarios. Four different frequency values
machine height (MH), is considered to simulate real-world have been explored to be used in RT simulations, namely
situations. Machines were at first deployed according to a 700 MHz, 3.5 GHz, 28 GHz, and 60 GHz. The selection of
perfectly regular layout with constant spacing (SP) between these frequencies is driven by the need to assess wireless
them. To make the digital representation of the industrial propagation characteristics for a broad spectrum of possible
environment more realistic, a fraction T of machines has industrial applications.
been then randomly removed, thus creating some emptier
spaces among them. C. PROPAGATION MARKERS UNDER INVESTIGATION
Multipath propagation in this industrial layout is inves- Metal is commonly present in industrial scenarios: besides
tigated in this study for different values of MS (2, 3, industrial machinery, which is primarily made of metal,
4, 8 m), MH (2 m), SP (2, 3, 4m), and T (0.1, 0.2, metal pipes, shelves, beams, doors, etc. are commonly found
0.35, 0.5). An example of two different realizations of the in factories. From a propagation perspective, the massive
industrial environment is shown in Fig. 1. A total number presence of metal parts corresponds to rich multipath effects,
of 48 scenarios has been then considered, corresponding to i.e., strong reflections and - sometimes - diffractions on metal
the whole set of combinations of these parameters. For each wedges may trigger the existence of a multitude of different
case, five maps have been generated by randomly changing radio paths the wireless signals can follow to effectively
the set of removed machines, for the same T value. The final propagate through the cluttered industrial environment [1],
number of digital maps therefore amounted to 240. [4], [5], [6], [7], [11].
731 IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO. 3, JUNE 2024

TABLE 2. Simulation configuration.

Because of multipath interference, the envelope of the
received signal exhibits fast and - sometimes deep - spatial
fluctuations over distances of the order of the wavelength
(small-scale fading), whereas blockage effects result in addi-
tional oscillations at a slower rate (large-scale fading) [45].
Because of multipath and blockage effects, the received
signal strength decreases with distance only on average.
That is, path loss experienced over a link distance d is
expressed as:
 
d
PL(d) = PL(d0 ) + 10 · n · log10 +χ (2)
d0
where n is the PLE (or propagation factor), PL(d0 ) is the to a final, mean DS somehow characteristic of the whole
PL at a reference distance d0 and χ is a random variable propagation environment.
accounting in general for the spatial fluctuations triggered
by large- and small-scale fading. The reference distance is IV. MACHINE-LEARNING-BASED PATH LOSS,
often set as d0 = 1m and the corresponding loss is then SHADOWING AND DELAY SPREAD PREDICTION
computed through the free space formula, i.e.: ML relies on vast datasets and adaptable model architectures
  to make predictions. In recent times, ML methods have found
4π
PL(d0 ) = 20 · log10 (3) applications in diverse fields such as self-driving cars, data
λ
mining, computer vision, and speech recognition, among oth-
In case the values of PL over a multitude of point- ers [46]. These applications encompass both supervised and
to-point links spread over the propagation scenario are unsupervised learning. In supervised learning, where labeled
collected through measurements or site-specific simulations, data is available, the objective is to deduce a generalized
the corresponding PLE can be easily computed as: and accurate function mapping inputs to outputs. This makes
N supervised learning well-suited for tackling classification
(PLi − PL(d0 )) · log10 (di )
PLE = i=1   2 (4) and regression problems. By contrast, unsupervised learning
10 N i=1 log 10 (di ) algorithms aim at unraveling the concealed structure within
being N the number of considered TX-RX links, PLi and unlabeled data. In essence, PL, shadowing, and DS prediction
di the path-loss value and the TX-RX distance for the i-th represent supervised regression problems, which have been
considered link. here tackled by ANN, and in particular, resorting to Multi-
Finally, it is worth pointing out that the random coefficient Layer Perceptron (MLP) [47]. Other ML methods, like
χ in eq. (2) is commonly limited to large-scale fading Support Vector Regression (SVR) [48], XGBR decision
only [4]. To get rid of small-scale fading effects, PL trees [49], and Random Forest (RF) [50] have been also
measured or simulated values must be spatially averaged considered for the sake of comparison. It has been reported
over a small area having a linear dimension equal to that ML-based models can be more accurate than empirical
several (tens of) λ [45]. Concerning the representation of the models and more computationally efficient than deterministic
industrial scenario considered in this study (Fig. 1), PL point ones [31], [51].
values have been therefore averaged over the 3 × 3 spatial
grids. After the removal of small-scale fading, then χ ∼ A. RAY TRACING SIMULATION
N(0, σ ) [45]. Although attenuation will always represent a RT is a powerful computational technique to model the prop-
crucial issue in wireless communication systems, distortion agation of radio waves in various environments [16], [17].
can also further impair the correct symbol sequence detection By simulating the interaction of radio waves with objects and
at the receiver side. In particular, time dispersion still due to materials in the environment, RT can predict the behavior of
multipath propagation can contribute to the overall degree of wireless communication systems, aiding in their design and
distortion to an extent that depends on the DS experienced optimization. Key parameters in RT simulations include the
over the environment and the symbol time length [45]. If M number of interactions, antenna type, and frequency value.
different propagation paths can be identified between a TX Table 2 summarizes this information for the RT simulations
and RX pair, and Pi and τi represent the intensity and the carried out in this work.
propagation delay of the i-th path, then: RT tools also require the electromagnetic parameters of
 the materials (relative permittivity R and conductivity γ )
M
 Pi as input data. In the considered industrial environments,
DS = · (τi − τ )2 (5)
P two types of materials are predominantly present. The first
i=1
M  type is machinery, which is assumed to be Perfect Electric
where P = i=1 Pi and τ  = M Pi
i=1 P · τi . Averaging the Conductors (PEC). The second type is the shed, which is
DS values over many, different TX, and RX pairs leads assumed made of concrete, with R =5, σ =0.01 S/m at lower
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 732

TABLE 3. Features.
It is worth pointing out that the size of the database
is actually not so large, that in general may not help the
accuracy of the ML training stage. The reason is twofold:
• RT simulations for PL modeling are in general not
computationally light. Although the RT parameters
were set to effectively speed up each single run,2 the
execution of the 960 RT simulations required about two
weeks of computation (on a standard PC).
• Each simulation includes four TXs and many RXs (i.e.,
several hundreds of wireless links) but in the end it
contributes to the computation of a single value of PLE,
σdB and DSmean .

C. MODEL SELECTION
The dataset achieved from RT simulations was finally used
to train and test an MLP network aimed at catching the
relationship between the output label (either PLE, σdB or
DSmean ) and the corresponding features. The learning process
was organized in three different steps:
FIGURE 2. Common PL range dependence in wireless channels. 1) Since the labels included in the final dataset look
somehow linearly dependent on log10 (fGHz ) (Fig. 4),
a simple linear regression was first carried out to
frequencies (0.7 and 3.5 GHz) and σ =0.1 S/m at higher compute the coefficient β and γ describing the average
frequencies (28 and 60 GHz). frequency dependence of the label as::
In this work, several synthetic datasets of the industrial
environment have been generated according to the input file < y >= β · log10 (fGHz ) + γ (6)
format required by the RT tool previously described in [52]. where y indifferently stands for PLE, σdB or DSmean
and · represents the mean value;
B. DATA COLLECTION AND FEATURE EXTRACTION 2) for each yi , i=1,2,.., 192 the residual ri concerning the
The collected data refer to samples obtained from RT regression line has been computed, i.e.:
simulations, and each sample should include the target output
ri = yi − β · log10 (fGHz ) − γ , i=1, 2,. . . , 192 (7)
value such as PL, shadowing, and DS and the corresponding
input features as mentioned in Table 3. 3) MLP has been leveraged to seize the dependence of
In the end, the total number RT simulations amounted to the residuals on the geometrical features (MS, SP,
960 (4 frequencies explored over the whole set of 240 digital and MD).
maps). For each simulation, PL and DS experienced over Of course, three different MLP networks, respectively
all the TX-RX links have been computed assuming isotropic tailored to PLE, σdB or DSmean have been in the end arranged
antenna at both link ends for the sake of simplicity. The RXs in agreement with the outlined procedure.
have been grouped in grids of 3 x 3 locations, with distance
between the grids equal to 10 m and spacing inside each D. HYPERPARAMETER SETTING AND MODEL TRAINING
grid equal to 10λ, being λ the communication wavelength. The primary objective during the training stage of every
In this way, fast fading is expected to independently affect machine learning model is to optimize the parameters, specif-
the different RXs, and therefore has been filtered out from ically the weights (w) and biases (b) of each layer in MLP,
the simulation results by averaging the PL values over each to achieve optimal learning. Following the training stage,
grid. The corresponding PL values collected over the five the validation phase focuses on fine-tuning hyperparameters
simulations sharing the same geometrical parameters have such as the number of hidden layers, the neurons within
been merged and plotted against link distance in logarithmic each hidden layer, and the activation function (represented
scale, for each different frequency (Fig. 2). The final number as ‘f’ in Fig. 5). Table 4 provides a brief summary of the
of collected PLE, σdB , and DSmean values therefore amounted tuned hyperparameters for three distinct models designed for
to 960/5=192 corresponding to the combinations of the predicting PLE, σ , and DSmean . It is noteworthy that ‘ReLU’
features (MS, SP, MD, and frequency). The collected data stands for the Rectifier Linear Unit function, and ‘lbfgs’
have been arranged on the dataset sketched in Fig. 3. serves as an optimizer from the family of quasi-Newton
The values of PLE, σdB , and DSmean corresponding to methods, particularly suitable for small datasets [53], [54].
the 192 combinations of the features (MS, SP, MD, and 2 In particular, the maximum number of bounces permitted for each ray
frequency) have been arranged in a database as sketched in was limited to three, with one diffraction at most. Also, the transmission
Fig. 3. was not enabled, as the machinery was supposed made of metal.
733 IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO. 3, JUNE 2024

FIGURE 3. An example of the dataset considering MS, SP, MD, and frequency as features and PL and Shadowing, and DS as target outputs.

FIGURE 4. Linear relationship between PLE, σdB , and DSmean Vs. log(f[GHz]).

TABLE 4. Major parameters of the neural networks for PLE, σdB and DS regression.

used [31], [51]. In this work RMSE is considered:

  1  2
RMSE y, ŷ = · yi − ŷi (8)
n
where yi is the ground-truth value for input xi and ŷi is the
FIGURE 5. General structure of a Multi-Layer Perceptron. A single hidden layer is
predicted value for input xi and n is the number of samples.
deployed for the sake of simplicity. Generalization property describes the model reusabil-
ity when the deployment concerns new frequency bands
E. MODEL EVALUATION and/or new environment types. The model may have better
In general, the performance of regression models is measured generalization performance with more data collected from
by samples in the test dataset, which do not appear in diverse scenarios, such as different machine densities and
the model training process. The evaluation metrics include frequencies.
prediction accuracy and generalization properties.
In terms of evaluating the accuracy, performance indicators V. RESULTS AND DISCUSSION
like maximum prediction error, mean absolute error, error A. NARROWBAND PARAMETERS ASSESSMENT
standard deviation, correlation factor, root mean square error The accuracy of the ML approach to PLE and σ prediction
(RMSE), and mean absolute percentage error are commonly is displayed in Table 5 in terms of overall loss over the test
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 734

FIGURE 6. PLE sensitivity to MD and frequency for MLP-based model (left) and XGBRegressor (right).

TABLE 5. Model evaluation for PLE and σ and min and max of each case.
generalization skill. Conversely, the XGBRegressor cannot
consistently track the PLE sensitivity to completely new
frequency values (Fig. 6(b)). The reason is in the context
of the training set, where frequency is represented by four
distinct values. Given that XGBRegressor operates on tree-
based structures, the decision points for splits in the tree are
TABLE 6. An example of new dataset.
defined by conditions such as frequency lower and/or greater
than a threshold. With only four unique values for frequency,
the resulting thresholds in the tree are likely to closely
align with these specific values. For instance, consider a
split condition frequency greater than 28 GHz and less than
60 GHz, where observations with frequencies 28 GHz and
40 GHz may be grouped on the same branch. In situations
dataset together with the range of variability of each target where the remaining features exhibit similar characteristics,
label. As the loss turns out quite small compared to the range these two observations are anticipated to receive nearly
of variability for both parameters, the learning task has been identical predictions. In conclusion, the MLP-based model
fairly accomplished. Furthermore, a comparison between turns out as the most reliable and flexible overall and will
different models shows satisfactory overall effectiveness, be therefore referred to in the following.
with the MLP and XGBRegressor slightly outperforming the Fig. 7 shows σ against MD for different frequency values.
others. Similar to the PLE model, they are in clear agreement with
To further investigate and evaluate the models’ perfor- the RT outcome as far as the MLP-based model is concerned.
mances, a generalization perspective is adopted. This entails Prior research suggests that conventional ML models excel
the evaluation of the models’ consistency when challenged in interpolating within the known data range but often exhibit
on new values of the features never seen during the training poor performance when extrapolating beyond this range [55].
stage. To this aim, a fresh dataset with entirely new feature However, the MLP-based model employed in this study is
values was considered according to Table 6. not in agreement with this trend, as it effectively learned
Fig. 6 shows PLE against MD for different frequency val- the correlation between the target output and frequency,
ues for both the MLP-based model (left) and XGBRegressor demonstrating a remarkable ability to predict accurately even
(right). Besides the data returned by RT simulations (dots), for frequency values outside the frequency range explored
new data generated utilizing the trained ML tools related inside the training dataset.
to the new dataset (squares and stars) are added to the In the end, the trained MLP can be exploited to quickly
figures, which also include the corresponding best-fit line. and reliably complement the limited amount of information
Results are in clear agreement with the RT outcome as that has been painstakingly gathered through electromagnetic
far as the MLP-based model is concerned as shown in simulations (or could be provided by channel experimental
Fig. 6(a), thus further corroborating the effectiveness of the sounding). Of course, the availability of a large set of
training but also highlighting a fair robustness in terms of information can help to clearly get a deeper insight into the
735 IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO. 3, JUNE 2024

FIGURE 7. σ sensitivity to MD and frequency for MLP-based model. FIGURE 8. DS sensitivity to MD and frequency for MLP-based model.

TABLE 7. Model evaluation for DSmean and min and max values.

consistent with the RT results. Best-fit lines are also traced

in Fig. 8, aimed at highlighting the average dependence of
DSmean on MD and frequency. The corresponding analytical
formulation is reported in the following eq. (11):
1
DSmean = k5 · (11)
1
k
existing relationship between the output propagation labels MD · fGHz
6

and the considered input features. For instance, Figs. 6(a)

and 7 quite clearly show that both PLE and σ increase with The least square method can be leveraged to compute
machine density and frequency. Simple best-fit lines have the best k5 and k6 values at each different frequency
been then computed to formally describe the highlighted (continuous and dashed lines in Fig. 8. A simple, average
trends, e.g., employing the following expressions: DS prediction formula describing its sensitivity to MD and
frequency can be instead achieved by extending the least
1
4 k square computation all over the whole dataset. In this case,
PLE = 2 + k1 · MD · fGHz
2
(9) the values of the coefficients were found to equal k5 =8.35
√
σ = k3 · MD + k4 · log10 (fGHz ) (10) and k6 =20.35.

The coefficients k1 − k4 can be easily computed for each C. COMPARISON WITH PREVIOUS STUDIES
frequency according to the least square method over the This sub-section discusses the reliability of the proposed ML-
corresponding dataset, leading to the continuous and dashed based approach to wireless propagation modeling in factories
lines in Figs. 6(a) and 7. To get a single, somehow rougher in comparison with the results reported in previous studies
but very simple analytical model, the same coefficients have (Table 1) and technical report [56].
been also optimized across the whole set of frequencies, Experimental and simulation assessments summarized in
corresponding to the final values of k1 = 0.48, k2 = 1.5, Table 1 show that in the presence of LoS, PLE is approxi-
k3 = 10.27, and k4 = 3.45. mately equal to 2 regardless of the frequency (green crosses
in Fig. 9), whereas in NLoS conditions it is increasingly
B. WIDEBAND PARAMETERS ASSESSMENT greater as the NLoS level gets heavier (red dots and purple
Table 7 shows that the learning process can be accomplished stars in Fig. 9). Also, PLE significantly increases with
also for DS prediction. Although the four considered ML frequency in the case of dense industrial clutter (purple stars
methods turn out quite accurate, the MLP-based model in Fig. 9). These trends are actually in contrast with the
outperforms the others. The model flexibility and consistency path-loss models for indoor factories included in [56], where
are again investigated in Fig. 8, where the DSmean values a PLE independent of frequency is instead assumed, and
returned by RT simulations are reported for different MD the value corresponding to sparse clutter is greater than that
and frequency together with the values achieved from the considered for dense clutter. Also, PLE values in [56] for
trained MLP-based model corresponding to the same new the industrial case are just slightly greater than 2 even in
set of features previously referred to (Table 6). NLoS conditions, which sounds somehow unlikely.
As shown in Fig. 8, the outcomes from the MLP for In this framework, the results returned by the proposed
the values of the fresh features are physically sound and MLP-based model look like a sort of trade-off between
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 736

FIGURE 11. Model - literature comparison for DS.

FIGURE 9. Model - literature comparison for path loss exponent.

reported in [56] (σDS ) is equal to about 18 nsec. (for both

LoS and NLoS cases), thus meaning that most of the values
returned by the joint RT and ML approach actually belong to
the ±σDS range around the average values suggested in [56].
It is finally worth noting that although the learning process
was based on data collected from simulations carried out
for the same size of the industrial shed, results discussed
in this section have turned out to be in general agreement
with previous model and experimental investigations clearly
referred to different industrial environments and machinery
layout. This confirms the model is flexible and versatile,
and suggests it has fair generalization skills also in the
environment domain.
FIGURE 10. Model - literature comparison for shadowing level.
VI. CONCLUSION
In this work, narrowband and wideband propagation
previous evaluations, showing a PLE that slightly increases modeling in an industrial environment has been addressed by
with frequency (Fig. 9) and machine density (Fig. 6(a)). employing an ML-based approach including a simple linear
Moreover, since the PLE values referring to RT and ML regression and a multi-layer perceptron network. Training
model in Fig. 9 correspond to LoS and NLoS altogether, it is data have been achieved through ray-tracing simulations
not surprising they are on the average in between PLE values carried out at different frequencies and on several real-
achieved for LoS and (heavy) NLoS conditions separately. izations of the industrial context. Satisfactory performance
Rather similar considerations hold for the assessment of of the ML model has been achieved, both in terms of
shadowing level, as reported in Fig. 10. It is worth reminding interpolation and extrapolation skills. Overall, the results
that the values of σdB corresponding to RT and ML in Fig. 10 are quite in agreement with previous investigations and
again refer to LoS and NLoS altogether, which may explain existing models. Moreover, the trained neural network has
why values greater than those limited to LoS or NLoS only been queried to complement the outcomes of RT simulations,
have been sometimes achieved. therefore achieving a large(r) database accounting for the
Finally, a comparison in terms of DS is shown in Fig. 11. complex relationship between propagation markers like path
Since the values of DS in Table 1 for NLoS conditions are loss exponent, shadowing level, and DS and features like
seldom related to the corresponding clutter level, Fig. 11 frequency and machine size, spacing and density. Based
simply considers LoS and NLoS classes. Results returned by on this extended dataset, simple, closed-form formulas to
RT and ML assessment are in fair agreement with previous estimate the values of path loss exponent, shadowing level,
experimental analyses (green crosses and red dots in the and DS from machine density and communication frequency
figure), suggesting a slight reduction of DS for increasing have been also proposed. Both path-loss exponent and
frequency. This was already clear in Fig. 7, where the shadowing std. deviations increase at a larger frequency and
difference between mean DS at 0.7 GHz and 300 GHz machine density, whereas an opposite trend has been found
amounts to about 7 nsec only. Comparison with the simple for DS. Although the sensitivity of the considered propa-
formulas for mean DS included in [56] looks a bit worse, gation parameters to the height of antennas and machinery
but anyway acceptable. In fact, the DS standard deviation has not been considered to limit the computational effort, it
737 IEEE OPEN JOURNAL OF ANTENNAS AND PROPAGATION, VOL. 5, NO. 3, JUNE 2024

clearly represents an interesting issue that deserves further [21] F. Fuschini, M. Barbiroli, G. Bellanca, G. Calò, J. Nanni, and
investigation as well as the impact of antenna radiation V. Petruzzelli, “A ray tracing tool for propagation modeling in layered
media: A case study at the chip scale,” IEEE Open J. Antennas
patterns on the properties of the industrial wireless channel. Propag., vol. 3, pp. 249–262, 2022, doi: 10.1109/OJAP.2022.3148855.
[22] V. Degli-Esposti, F. Fuschini, E. M. Vitucci, and G. Falciasecca,
“Speed-up techniques for ray tracing field prediction models,” IEEE
REFERENCES Trans. Antennas Propagat., vol. 57, no. 5, pp. 1469–1480, May 2009,
[1] S. Vitturi, F. Tramarin, and L. Seno, “Industrial wireless networks: doi: 10.1109/TAP.2009.2016696.
The significance of timeliness in communication systems,” IEEE Ind. [23] A. Seretis and C. D. Sarris, “An overview of machine learning
Electron. Mag., vol. 7, no. 2, pp. 40–51, Jun. 2013. techniques for radio wave propagation modeling,” IEEE Trans.
[2] V. K. L. Huang, Z. Pang, C.-J. A. Chen, and K. F. Tsang, “New Antennas Propag., vol. 70, no. 6, pp. 3970–3985, Jun. 2022.
trends in the practical deployment of industrial wireless,” IEEE Ind. [24] D. C. Nguyen et al., “Enabling AI in future wireless networks: A data
Electron. Mag., vol. 12, no. 2, pp. 50–58, Jun. 2018. life cycle perspective,” IEEE Commun. Surveys Tuts., vol. 23, no. 1,
[3] M. Cheffena, “Industrial wireless communications over the millimeter pp. 553–595, 1st Quart., 2021, doi: 10.1109/COMST.2020.3024783.
wave spectrum: Opportunities and challenges,” IEEE Commun. Mag., [25] C. Huang et al., “Artificial intelligence enabled radio propagation
vol. 54, no. 9, pp. 66–72, Sep. 2016. for communications—Part I: Channel characterization and antenna-
[4] M. Cheffena, “Propagation channel characteristics of industrial wire- channel optimization,” IEEE Trans. Antennas Propag.. vol. 70, no. 6,
less sensor networks,” IEEE Antennas Propag. Mag., vol. 58, no. 1, pp. 3939–3954, Jun. 2022, doi: 10.1109/TAP.2022.3149663.
pp. 66–73, Feb. 2016. [26] C. Huang et al., “Artificial intelligence enabled radio propa-
[5] P. Stenumgaard, J. Chilo, J. Ferrer-Coll, and P. Angskog, “Challenges gation for communications—Part II: Scenario identification and
and conditions for wireless machine-to-machine communications in channel modeling,” IEEE Trans. Antennas Propag., vol. 70, no. 6,
industrial environments,” IEEE Commun. Mag., vol. 51, no. 6, pp. 3955–3969, Jun. 2022, doi: 10.1109/TAP.2022.3149665.
pp. 187–192, Jun. 2013. [27] D. Solomitckii, A. Orsino, S. Andreev, Y. Koucheryavy, and
[6] E. Tanghe et al., “The industrial indoor channel: Large-scale and M. Valkama, “Characterization of mmWave channel properties at
temporal fading at 900, 2400, and 5200 MHz,” IEEE Trans. Wireless 28 and 60 GHz in factory automation deployments,” in Proc.
Commun., vol. 7, no. 7, pp. 2740–2751, Jul. 2008. IEEE Wireless Commun. Netw. Conf. (WCNC), 2018, pp. 1–6,
[7] J. Karedal, S. Wyne, P. Almers, F. Tufvesson, and A. F. Molisch, doi: 10.1109/WCNC.2018.8377337.
“A measurement-based statistical model for industrial ultra- [28] J. Yang et al., “Ultra-reliable communications for indus-
wideband channels,” IEEE Trans. Wireless Commun., vol. 6, no. 8, trial Internet of Things: Design considerations and channel
pp. 3028–3037, Aug. 2007. modeling,” IEEE Netw., vol. 33, no. 4, pp. 104–111, Jul./Aug. 2019,
[8] Y. Ai, J. B. Andersen, and M. Cheffena, “Path-loss prediction for an doi: 10.1109/MNET.2019.1800455.
industrial indoor environment based on room electromagnetics,” IEEE [29] G. S. Bhatia, Y. Corre, and M. Di Renzo, “Efficient ray-
Trans. Antennas Propag., vol. 65, no. 7, pp. 3664–3674, Jul. 2017. tracing channel emulation in industrial environments: An analysis
[9] B. B. Cebecioglu et al., “Sub-6 GHz channel modeling and eval- of propagation models impact,” in Proc. Joint Eur. Conf. Netw.
uation in indoor industrial environments,” IEEE Access, vol. 10, Commun. 6G Summit (EuCNC/6G Summit), 2023, pp. 180–185,
pp. 127742–127753, 2022. doi: 10.1109/EuCNC/6GSummit58263.2023.10188258.
[10] T. S. Rappaport and C. D. McGillem, “UHF fading in factories,” IEEE [30] M. U. Sheikh, K. Ruttik, R. Jäntti, and J. Hämäläinen, “Blockage
J. Sel. Areas Commun., vol. 7, no. 1, pp. 40–48, Jan. 1989. and ray tracing propagation model in 3GPP specified industrial envi-
[11] Y. Wang, Y. Lv, X. Yin, and J. Duan, “Measurement-based experi- ronment,” in Proc. Int. Conf. Inf. Netw. (ICOIN), 2021, pp. 397–402,
mental statistical modeling of propagation channel in Industrial IoT doi: 10.1109/ICOIN50884.2021.9333909.
scenario,” Radio Sci., vol. 55, no. 9, pp. 1–14, Sep. 2020. [31] E. Ostlin, H.-J. Zepernick, and H. Suzuki, “Macrocell path-
[12] B. Holfeld et al., “Radio channel characterization at 5.85 GHz loss prediction using artificial neural networks,” IEEE Trans.
for wireless M2M communication of industrial robots,” in Veh. Technol., vol. 59, no. 6, pp. 2735–2747, Jul. 2010,
Proc. IEEE Wireless Commun. Netw. Conf., 2016, pp. 1–7, doi: 10.1109/TVT.2010.2050502.
doi: 10.1109/WCNC.2016.7564890. [32] C. A. Oroza, Z. Zhang, T. Watteyne, and S. D. Glaser,
[13] A. Al-Saman, M. Mohamed, M. Cheffena, and A. Moldsvor, “A machine-learning based connectivity model for complex ter-
“Wideband channel characterization for 6G networks in industrial rain large-scale low-power wireless deployments,” IEEE Trans.
environments,” Sensors vol. 21, no. 6, Mar. 2021. Cogn. Commun. Netw., vol. 3, no. 4, pp. 576–584, Dec. 2017,
[14] S. R. Panigrahi, N. Bjorsell, and M. Bengtsson, “Power delay profile doi: 10.1109/TCCN.2017.2741468.
investigation in industrial indoor environments at the 24 GHz ISM [33] I. Popescu, D. Nikitopoulos, P. Constantinou, and I. Nafornita, “ANN
band,” in Proc. IEEE Int. Conf. Ind. Technol. (ICIT), 2022, pp. 1–6, prediction models for outdoor environment,” in Proc. IEEE 17th
doi: 10.1109/ICIT48603.2022.10002732. Int. Symp. Pers., Indoor, Mobile Radio Commun., 2006, pp. 1–5,
[15] D. Dupleich, R. Müller, M. Landmann, J. Luo, G. D. Galdo, and doi: 10.1109/PIMRC.2006.254270.
R. S. Thomä, “Multi-band characterization of propagation in industry [34] M. Ayadi, A. Ben Zineb, and S. Tabbane, “A UHF path loss
scenarios,” in Proc. 14th Eur. Conf. Antennas Propagat., Copenaghen, model using learning machine for heterogeneous networks,” IEEE
2020, pp. 1–5, doi: 10.23919/EuCAP48036.2020.9135630. Trans. Antennas Propag., vol. 65, no. 7, pp. 3675–3683, Jul. 2017,
[16] Z. Yun and M. F. Iskander, “Ray tracing for radio propaga- doi: 10.1109/TAP.2017.2705112.
tion modeling: Principles and applications,” IEEE Access, vol. 3, [35] M. González-Palacio, D. Tobón-Vallejo, L. M. Sepúlveda-Cano,
pp. 1089–1100, 2015, doi: 10.1109/ACCESS.2015.2453991. S. Rúa, and L. B. Le, “Machine-learning-based combined path loss
[17] J. W. McKown and R. L. Hamilton, “Ray tracing as a design tool for and shadowing model in LoRaWAN for energy efficiency enhance-
radio networks,” IEEE Netw., vol. 5, no. 6, pp. 27–30, Nov. 1991. ment,” IEEE Internet Things J., vol. 10, no. 12, pp. 10725–10739,
[18] E. M. Vitucci, N. Cenni, F. Fuschini, and V. Degli-Esposti, “A Jun. 2023, doi: 10.1109/JIOT.2023.3239827.
reciprocal heuristic model for diffuse scattering from walls and sur- [36] I. F. M. Rafie, S. Y. Lim, and M. J. H. Chung, “Path loss prediction
faces,” IEEE Trans. Antennas Propag., vol. 71, no. 7, pp. 6072–6083, in urban areas: A machine learning approach,” IEEE Antennas
Jul. 2023, doi: 10.1109/TAP.2023.3278796. Wireless Propag. Lett., vol. 22, no. 4, pp. 809–813, Apr. 2023,
[19] M. F. Catedra, J. Perez, F. Saez de Adana, and O. Gutierrez, “Efficient doi: 10.1109/LAWP.2022.3225792.
ray-tracing techniques for three-dimensional analyses of propagation [37] A. Gupta, J. Du, D. Chizhik, R. A. Valenzuela, and M. Sellathurai,
in mobile communications: Application to picocell and microcell “Machine learning-based urban canyon path loss prediction
scenarios,” IEEE Antennas Propag. Mag., vol. 40, no. 2, pp. 15–28, using 28 GHz manhattan measurements,” IEEE Trans.
Apr. 1998. Antennas Propag., vol. 70, no. 6, pp. 4096–4111, Jun. 2022,
[20] Y. Hwang, Y. P. Zhang, and R. G. Kouyoumjian, “Ray-optical doi: 10.1109/TAP.2022.3152776.
prediction of radio-wave propagation characteristics in tunnel environ- [38] K. Mao et al., “Machine-learning-based 3D channel modeling for U2V
ments. 1. Theory,” IEEE Trans. Antennas Propagat., vol. 46, no. 9, mmWave communications,” IEEE Internet Things J., vol. 9, no. 8,
pp. 1328–1336, Sep. 1998. pp. 17590–17607, Sep. 2022, doi: 10.1109/JIOT.2022.3155773.
ZADEH et al.: ML APPROACH TO WIRELESS PROPAGATION MODELING 738

[39] Y. Zhang, J. Wen, G. Yang, and Z. He, “Air-to-air path loss prediction MOHAMMAD HOSSEIN ZADEH received the M.Sc.
based on machine learning methods in urban environments,” Wireless degree in telecommunications engineering from
Commun. Mobile Comput., vol. 2018, Jun. 2018, Art. no. 8489326. the University of Bologna, where he is currently
[40] A. Seretis, V. Jevremovic, A. Jemmali, and C. D. Sarris, “Generalizable pursuing the Ph.D. degree in telecommunica-
machine-learning-based modeling of radio wave propagation in stadi- tions engineering. His activity is focused on the
ums,” IEEE Open J. Antennas Propag., vol. 4, pp. 1116–1128, 2023, research topic of Assessment and Development
doi: 10.1109/OJAP.2023.3310382. of Channel Modeling by Machine Learning.
[41] D. Wu, G. Zhu, and B. Ai, “Application of artificial neural networks His research activity includes participation in
for path loss prediction in railway environments,” in Proc. 5th Int. European research and cooperation programs
ICST Conf. Commun. Netw., 2010, pp. 1–5. (COST INTERACT).
[42] T. Popescu, D. Nikitopoulos, I. Nafornita, and P. Constantinou, “ANN
prediction models for indoor environment,” in Proc. IEEE Int. Conf.
Wireless Mobile Comput., Netw. Commun., 2006, pp. 366–371.
[43] M. I. AlHajri, N. T. Ali, and R. M. Shubair, “Classification
of indoor environments for IoT applications: A machine learn-
ing approach,” IEEE Antennas Wireless Propag. Lett., vol. 17,
pp. 2164–2168, 2018, doi: 10.1109/LAWP.2018.2869548. MARINA BARBIROLI received the Laurea degree
[44] M. Kashef, P. Vouras, R. D. Jones, R. Candell, and K. A. Remley, “A in electronic engineering and the Ph.D. degree
machine-learning approach for the exemplar extraction of mmWave in computer science and electronic engineer-
industrial wireless channels,” IEEE Open J. Instrum. Meas., vol. 1, ing from the University of Bologna in 1995
pp. 1–15, Jun. 2022, doi: 10.1109/OJIM.2022.3181309. and 2000, respectively, where she is currently
[45] H. L. Bertoni, “Survey of observed characteristics of the propaga- an Associate Professor with the Department of
tion channel,” in Radio Propagation for Modern Wireless Systems. Electrical, Electronic and Information Engineering
Hoboken, NJ, USA: Prentice Hall, 2000. “G. Marconi.” Her research interests are on
[46] Y. Zhang, J. Wen, G. Yang, Z. He, and J. Wang, “Path loss prediction propagation models for mobile communications
based on machine learning: Principle, method, and data expan- systems, with a focus on wideband channel
sion,” Appl. Sci., vol. 9, p. 1908, May 2019, doi: 10.3390/app9091908. modeling for 5G systems and beyond. Her research
[47] M. Popescu, V. Balas, L. Perescu-Popescu, and N. Mastorakis, activities include investigation of planning strategies for mobile systems,
“Multilayer perceptron and neural networks,” WSEAS Trans. Circuits broadcast systems, and broadband wireless access systems, analysis of
Syst., vol. 8, no. 7, pp. 579–588, 2009. exposure levels generated by all wireless systems and for increasing
[48] A.J. Smola and B. Schölkopf, “A tutorial on support vector spectrum efficiency. Her research activity includes participation in European
regression,” Stat. Comput. vol. 14, pp. 199–222, Aug. 2004, research and cooperation programs (COST 259, COST 273 COST2100,
doi: 10.1023/B:STCO.0000035301.49549.88. COST IC1004, COST IRACON, and COST INTERACT) and the European
[49] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting Networks of Excellence FP6-NEWCOM and FP7-NEWCOM++.
system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data
Min., 2016, pp. 785–794, doi: 10.1145/2939672.2939785.
[50] A. Liaw and M. Wiener, “Classification and regression by random-
Forest,” R News, vol. 2, no. 3, pp. 18–22, 2002.
[51] J. Isabona and V. M. Srivastava, “Hybrid neural network approach
for predicting signal propagation loss in urban microcells,” in Proc.
IEEE Region 10 Humanit. Technol. Conf. (R10-HTC), 2016, pp. 1–5. FRANCO FUSCHINI received the M.Sc. degree
[52] E. M. Vitucci et al., “Ray tracing RF field prediction: An unforgiving in telecommunication engineering and the Ph.D.
validation,” Int. J. Antennas Propag., vol. 2015, pp. 1–11, Aug. 2015. degree in electronics and computer science from
[53] A. F. Agarap, “Deep learning using rectified linear units the University of Bologna, Italy, in March 1999
(ReLU),” 2019, arXiv:1803.08375. and July 2003, respectively, where he is currently
[54] A. S. Berahas, M. Jahani, P. Richtárik, and M. Takáč, “Quasi- an Associate Professor with the Department of
newton methods for machine learning: Forget the past, just Electrical, Electronic and Information Engineering
sample,” Optim. Methods Softw., vol. 37, no. 5, pp. 1668–1704, 2022, “Guglielmo Marconi.” He is the author or a coau-
doi: 10.1080/10556788.2021.1977806. thor of about 50 journal articles on electromagnetic
[55] G. J. Hahn, “The hazards of extrapolation in regression anal- wave propagation and wireless systems design.
ysis,” J. Qual. Technol., vol. 9, no.4, pp. 159–165, 2018, His main research interests are in the areas of
doi: 10.1080/00224065.1977.11980791. radio systems technologies and radio propagation theoretical modeling and
[56] “Study on channel model for frequencies from 0.5 to 100 GHz; experimental investigation. He received the “Marconi Foundation Young
(Release 16), Version 16.1.0,” ETSI Sophia Antipolis, France, Rep. TR Scientist Prize” in the context of the XXV Marconi International Fellowship
38.901, 2020. Award in April 1999.

Open Access funding provided by ’Alma Mater Studiorum - Università di Bologna’ within the CRUI CARE Agreement