IBRBNC

TYPE Original Research
PUBLISHED 04 June 2024

DOI 10.3389/fphy.2024.1374138
Intelligent Bayesian regularization

OPEN ACCESS backpropagation neuro
EDITED BY
Andaç Batur Çolak,
Istanbul Commerce University, Türkiye
computing paradigm for state
REVIEWED BY
Yuxing Li,
features estimation of underwater
Xi’an University of Technology, China
Khawaja Fahad Masood,
University of Technology Sydney, Australia
passive object
*CORRESPONDENCE
Wasiq Ali, Wasiq Ali 1,2,3*, Muhammad Bilal 1,2,3, Ayman Alharbi 4, Amar Jaffar 4,
wasiqali@hrbeu.edu.cn
Abdulaziz Miyajan 4 and Syed Agha Hassnain Mohsan 5
RECEIVED 20 February 2024 1
ACCEPTED 22 April 2024 National Key Labortary of Underwater Acoustic Technology, Harbin Engineering University, Harbin,
PUBLISHED 04 June 2024
China, 2Key Laboratory of Marine Information Acquisition and Security (Harbin Engineering University),
Ministry of Industry and Information Technology, Harbin, China, 3College of Underwater Acoustic
CITATION Engineering, Harbin Engineering University, Harbin, China, 4Computer and Network Engineering
Ali W, Bilal M, Alharbi A, Jaffar A, Miyajan A and Department, College of Computing, Umm Al-Qura University, Mecca, Saudi Arabia, 5Eastern Institute for
Hassnain Mohsan SA (2024), Intelligent Advanced Study, College of Information Science and Technology, Ningbo, China
Bayesian regularization backpropagation neuro
computing paradigm for state features
estimation of underwater passive object.
Front. Phys. 12:1374138.
doi: 10.3389/fphy.2024.1374138 In underwater environments, the accurate estimation of state features for passive
COPYRIGHT object is a critical aspect of various applications, including underwater robotics,
© 2024 Ali, Bilal, Alharbi, Jaffar, Miyajan and surveillance, and environmental monitoring. This study presents an innovative neuro
Hassnain Mohsan. This is an open-access article
distributed under the terms of the Creative
computing approach for instantaneous state features reckoning of passive marine
Commons Attribution License (CC BY). The use, object following dynamic Markov chains. This paper introduces the potential of
distribution or reproduction in other forums is intelligent Bayesian regularization backpropagation neuro computing (IBRBNC) for
permitted, provided the original author(s) and
the copyright owner(s) are credited and that the
the precise estimation of state features of underwater passive object. The proposed
original publication in this journal is cited, in paradigm combines the power of artificial neural network with Bayesian
accordance with accepted academic practice. regularization technique to address the challenges associated with noisy and
No use, distribution or reproduction is
permitted which does not comply with these
limited underwater sensor data. The IBRBNC paradigm leverages deep neural
terms. networks with a focus on backpropagation to model complex relationships in the
underwater environment. Furthermore, Bayesian regularization is introduced to
incorporate prior knowledge and mitigate overfitting, enhancing the model’s
robustness and generalization capabilities. This dual approach results in a highly
adaptive and intelligent system capable of accurately estimating the state features of
passive object in real-time. To evaluate the efficacy of this intelligent computing
approach, a controlled supervised maneuvering trajectory for underwater passive
object is constructed. Real-time estimations of location, velocity, and turn rate for
dynamic target are scrutinized across five distinct scenarios by varying the Gaussian
observed noise’s standard deviation, aiming to minimize mean square errors (MSEs)
between real and estimated values. The effectiveness of the proposed IBRBNC
paradigm is demonstrated through extensive simulations and experimental trials.
Results showcase its superiority over traditional nonlinear filtering methods like
interacting multiple model extended Kalman filter (IMMEKF) and interacting
multiple model unscented Kalman filter (IMMUKF), especially in the presence of
noise, incomplete measurements and sparse data.
KEYWORDS
neuro computing, Markov chain, state features, sensor data, robustness, Bayesian
regularization, maneuvering trajectory
Frontiers in Physics 01 frontiersin.org

Ali et al. 10.3389/fphy.2024.1374138
1 Introduction step. The aforementioned capacity enables them to effectively

manage a diverse set of underwater target motion models and
The estimation of state elements for passive vehicles in aquatic measurement expressions, hence improving the accuracy of state
environments poses a significant challenge and is of utmost feature estimation [15]. Underwater circumstances exhibit dynamic
importance for a wide range of applications [1]. The usages of behaviors, characterized by the presence of quickly fluctuating
these technologies incorporate several fields, including underwater conditions, including underwater currents, tides, and object
robotics, where accurate knowledge of a vehicle’s trajectory and maneuverings. In this situation, the adaptable nature of KF and
orientation is essential. As well as surveillance, which requires the its nonlinear versions is essential [16]. One of the primary
tracking of targets of interest, and atmosphere monitoring, where constraints associated with the EKF and UKF methodologies is
identifying the complexities of marine ecosystems is of extreme their underlying assumption of linearity. There are a lot of
importance [2, 3]. Accurately forecasting the state characteristics of underwater target tracking circumstances where state variables
passive object is a crucial factor in various circumstances, as it and measurements have nonlinear associations. In cases, where
directly impacts its efficiency and safety of operations [4]. One of the the system exhibits substantial nonlinearity, the application of
main challenges in estimating the state features of submerged object these filters for linearization purposes may result in inaccuracies,
is the inherent noise and limited availability of sensor data in aquatic resulting in imprecise state feature estimations. Accurately
scenarios [5]. In the domain of underwater sensing, many types of simulating these processes can be challenging in underwater
sensors, including sonar, acoustic, and optical sensors, frequently environments, where the dynamics can be complicated and
encounter a wide range of noise sources arising from phenomena unpredictable [17].
such as water turbulence, ambient noise, and signal compression [6]. The utilization of interacting multiple model (IMM) Kalman
These elements have a tendency to provide measurement errors, filtering is of great importance in the assessment of the state of
hence posing challenges in the extraction of useful information for underwater targets, especially in situations where the dynamics of
the state estimation mechanism. In addition, underwater processes the target or the underwater environment are prone to frequent
often face situations in which there may be a lack of sensor coverage, alterations or uncertainty [18]. This modified filtering strategy
demanding the development of creative approaches to address data aims to mitigate certain drawbacks inherent in conventional
deficiencies [7]. State estimation is a crucial aspect within the field of Kalman filtering approaches, hence providing numerous
underwater robotics as it facilitates the ability of autonomous benefits [19]. The IMM Kalman filtering technique enables the
underwater vehicles (AUVs) to navigate, execute tasks, and representation of numerous motion models or modes, each
successfully engage with their surrounding water [8]. Precise characterizing a distinct target behavior. IMM Kalman filtering
assessments of a vehicle’s physical coordinates, speed, and demonstrates exceptional performance in tracking dynamic
angular orientation are crucial for various activities, including behaviors shown by underwater targets that exhibit motion
subaquatic mapping, investigation, and even archaeological pattern variations, such as dodging maneuvers or alterations in
research. The capability of accurately evaluating these state depth [20]. The technique has the capability to smoothly transition
characteristics in real-time and adapt to varying underwater between various motion models in order to uphold precise state
conditions is a vital requirement for the effective execution of estimation. The selection of motion models in the context of IMM
underwater robotic activities [9]. Kalman filtering is often a complex decision-making process [21].
Researchers have been actively investigating novel methods for The task of choosing an appropriate combination of models and
estimating state features in order to address the challenges presented their associated transition probabilities might present difficulties,
by complex underwater mediums [10]. The utilization of well- as it relies on the particular behavior of the undersea target, which
known Kalman filtering and its nonlinear versions holds may not always be fully known or readily described [22]. The
considerable importance for estimating the current state of selection of inappropriate models can result in poor results, filter
underwater objects [11]. The employing of filtering techniques is divergence, and overfitting. Moreover, the inclusion of a large
vital in the process of modeling and estimating the state features of number of models may lead to the formation of a system that is
passive targets within underwater environments. The Kalman filter overly complicated without necessarily enhancing the accuracy of
(KF), along with its derivative methods such as the extended Kalman state features estimation [23]. When dealing with instances where
filter (EKF) and unscented Kalman filter (UKF) demonstrates the undersea target exhibits extreme, sudden, or erratic
exceptional proficiency in state estimation through effective fluctuations in behavior, IMM Kalman filtering may encounter
handling of measurements affected by noise. These filtering challenges in fast transitioning between models or precisely
algorithms deploy probabilistic models in order to make adjusting to these changes [24]. The integration of neural
estimations about the condition of submerged targets, thereby networks and deep learning has demonstrated potential for
mitigating the impact of noise and yielding more precise improving the precision and flexibility of state estimation
estimations [12]. Most state feature estimation problems in mechanisms in the underwater domain [25]. The use of deep
underwater environments consist of non-linear correlations learning techniques in state feature estimation of underwater
among state parameters and measurements [13]. In instances of objects has experienced a significant rise, bringing about an
this nature, conventional linear KFs are not properly viable. The evolutionary effect. This is primarily attributed to deep
EKF and UKF are algorithms that have been developed with the learning’s capacity to effectively process complicated data,
explicit purpose of mitigating the effects of nonlinearity in system dynamically adjust to varying underwater conditions, and
models [14]. This is achieved by approximating the nonlinear deliver precise estimations [26]. Deep learning models,
system dynamics through linearization at each discrete time specifically convolutional neural networks (CNNs) and

Ali et al. 10.3389/fphy.2024.1374138
recurrent neural networks (RNNs), have exceptional proficiency in and monitored itinerary for underwater target. The real-time
independently extracting relevant information from raw sensor approximation of the key features, including the position,
input, including acoustic sonar, and visualizing data [27, 28]. The velocity, and rate of change in path for a moving target are
collection of these learned parameters is crucial in the description thoroughly examined in five different scenarios. Statistical
of the submerged surroundings and the movement of objects, variants of observed Gaussian noise are applied to generate
hence enhancing the precision of state feature estimates. Often, in various underwater configurations for the purpose of
underwater environments, the relationships between sensor comparing the proposed paradigm with conventional Kalman
readings and target states are highly nonlinear. Deep learning filtering methods. Figure 1 presents a comprehensive and brief
models possess a built-in capacity to effectively capture and visual representation of the designed research approach. The
represent nonlinearities, hence enabling more precise state subsequent points outline the key findings of the
feature estimation in comparison to conventional nonlinear performed study.
variants of KF [29]. Due to several reasons such as obstacles,
noise, or constraints in the sensors, underwater data may be sparse • The work discusses the fundamental demand for precise
and unstable. Deep learning models have the ability to efficiently estimate of state features of passive dynamic object in
address the challenges posed by missing data points and underwater scenarios.
inconsistent measurements through the utilization of time series • The study presents a novel neuro computing methodology
and historical data [30]. These methods provide the capability to designed for estimating state features, such as position,
dynamically adjust the new data in real-time, enabling them to velocity, and rotation rate, across the x and y-axes.
effectively handle unknown variations in the undersea • The IBRBNC paradigm utilizes deep neural systems,
environment or the behavior of the target [31]. This emphasizing backpropagation to effectively represent
adaptability is of great significance in order to ensure precise complex connections within the undersea surroundings.
state estimations while circumstances undergo changes. • The integration of prior data and the mitigation of overfitting
Intelligent Bayesian regularization backpropagation neuro are achieved by the introduction of Bayesian regularization,
computing (IBRBNC) is a category of artificial neural networks that acts as a fundamental element of IBRBNC.
(ANNs) that integrate Bayesian methods for regularization [32]. It • The IBRBNC based dual approach is utilized to create a highly
has proven useful in a wide range of situations, especially when adaptable and smart system that can correctly guess the state
working with sparse data or in a noisy environment. The IBRBNC features of passive object in real time.
method offers a probabilistic structure that allows the modeling of • This study performs an in-depth review of the IBRBNC
uncertainties in predictions made by neural network [33]. These paradigm through a comparative analysis with conventional
systems possess the ability to adapt their complexity according to the nonlinear filtering techniques, IMMEKF and IMMUKF.
quantity of accessible data, making them highly suitable for various • The findings demonstrate the exceptional performance of
applications that exhibit dynamic data features. The effectiveness of IBRBNC, particularly in challenging underwater
IBRBNC has been found in various financial applications, including environments.
but not limited to stock price forecasting, risk evaluation, and
portfolio optimization [34]. This innovation has the potential to The subsequent portions of the paper are organized in the
facilitate disease diagnosis, evaluate patient risk levels, and provide following manner: Section 2 of this paper outlines the
recommendations for therapy treatments. The capacity to offer methodology for developing a maneuvering state features
estimations of uncertainty can aid medical professionals in estimation model in 2-dimensional rectangular coordinates. This
making well-informed judgments [35]. The IBRBNC extends section further elaborates on the comprehensive mathematical
across many domains, encompassing cybersecurity, network modeling of continuous routing object. Section 3 provides a
surveillance, and manufacturing quality control [36]. This soft comprehensive overview of the establishment and behavior of the
computing has been employed in several natural language IBRBNC network, containing a detailed examination of the training,
processing applications, including sentiment detection, text testing, and validation processes. In Section 4, we address the
sorting, and machine translation [37]. It also works well for estimation results and the least mean squared error of the
image analysis tasks like medical image analysis, object mentioned techniques. The final portion of the proposed work
recognition, and picture segmentation. Probabilistic estimations explains the notable achievements and outlines further research
of object locations and properties can be provided by this directions.
methodology [38, 39]. In addition, it has applications in
meteorology [40], air quality forecasting [41], climate change
[42], astronomy [43], and astrophysics [44]. Furthermore, it plays 2 System modeling of passive
a significant role in the field of robotics, including many applications underwater object
such as autonomous navigation, routing, and the supervision of
robots [45, 46]. In this portion of the study, the modeling of Markov chain
Motivated by the aforementioned applications, the current moving object is designed using a bilateral state feature estimation
study aims to explore a robust neuro computing methodology approach in angular dimensions. This methodology incorporates
with the objective of improving the real-time estimates of state the state space-based bearings only tracking (BOT) technique to
features for underwater passive maneuvering object. In order to precisely estimate the state features of a constantly rotating object in
evaluate the ability of this computing, we developed a regulated a complex and difficult marine atmosphere.In order to gather the

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 1
Graphical description of proposed paradigm.
collation of passive measurements, a total of eight monitoring bases of a navigational target and its mechanism for estimating state
are strategically positioned at equal intervals. The term passive features. Several systems in real life exhibit dynamic modeling
measurements of the moving object refer to the nonlinear and variables. The description of these diverse system variables is
intricate data obtained by hydrophones. It is assumed that the unlikely to be addressed by a singular model. In the context of
positions of the observers are known beforehand. The sole real-time state feature estimating applications, it is probable for the
approach to get the bearings of the dynamic vehicle by passive modeling values to experience variations during the estimation
acquisition from surveillance units. These bearings are relying on the phase. These mechanisms are commonly referred to as Markov
angular position and placement of each individual sensor module. chains or multi models. In these situations, the whole design could
The proposed state feature estimation structure aims to observe diverge if one particular system model is selected. Consequently, the
target mobility in the far field zone. This observation is based on the development of a generic movement model for dynamic object that
prediction of a consistent turning course, which is tracked through will regulate numerous system models is critical. Coordinated Turn
the adoption of nonlinear multi model Kalman filters and neuro (CT) and Wiener process velocity (WPV) models are applied in this
computing techniques. Figure 2 illustrates the maneuvering trends study to explain the underwater navigation object’s kinematics.

Ali et al. 10.3389/fphy.2024.1374138
calculation for spread of the dynamic shifting matrix Xτ with

respect to the sampling interval is described in Eq. 5 as:
1 0 ∂τ 0
⎢
⎡
⎢
⎢ ∂τ ⎤⎥⎥⎥⎥⎥
⎢
⎢
Xτ ⎢
0 1 0
⎥,
⎢
⎢0
⎣ 0 1 0 ⎥⎥⎦
(5)
0 0 0 1
while the sampling interval ∂τ is given in Eq. 6 as:

∂τ [(τ + 1) − τ ]. (6)
To ensure precise estimation of state features through the IBRBNC

paradigm, it is necessary to transform the state space model outlined
in Eq. 4 into discrete time notation. The discrete-time regressive
model is adopted because of its potential to achieve more precise
assessment of the system’s behavior at time instances τ that are
combination of the sampling interval ∂τ. The discrete time state
FIGURE 2 computation has been transformed with its required elements in
State feature estimation architecture. Eq. 7, resulting in the following form:
1 0 ∂τ 0 xτ
⎢
⎡
⎢ ∂τ ⎤⎥⎥⎥⎥⎥ ⎢
⎡
⎢
⎢ ⎤⎥⎥
⎢
⎢ 0 1 0 ⎢
⎢ yτ ⎥ ⎥⎥⎥ +α .
2.1 Wiener process velocity (WPV) Hτ+1 ⎢
⎢ ⎥ ⎢
⎢
⎢ 0 ⎥⎥⎦ ⎣ xτ′ ⎥⎥⎦ τ
⎢ ⎢ (7)
⎣0 0 1
dynamic modeling 0 0 0 1 yτ′

Xτ Hτ
The state vector Hmτ given in Eq. 1 illustrates the real-time
features of a maneuvering object at a given time interval τ, by using The process noise ατ in the Gaussian spectrum is characterized by its
x covariance ℓτ in the following way:
bidirectional angular axes to represent its position τ and
yτ
x′ ατ ≈ N(0, ℓ τ ). (8)
velocity τ as:
yτ′
In experimental tests, the Gaussian noise variance is typically
xm
⎡⎢⎢⎢ yτm ⎤⎥⎥⎥ assigned an integer value of 0.05 in order to generate target
Hm ⎢⎢⎢ τm ⎥⎥⎥⎥.
τ ⎢ ⎢⎣ x′τ ⎥⎦ (1) trajectory with minimal curves. In contrast,
m
y′τ ℓ τ Eατ αTτ . (9)
Concurrently, the state vector at the monitoring unit in angular axes In order to achieve a discrete-time state expression, it is necessary to
can be represented in Eq. 2 as follows: distinguish the process Gaussian noise in the WPV mathematical
xn design. It can be adopted to accurately incorporate the model’s
⎡⎢⎢⎢ yτn ⎤⎥⎥⎥ parameters over time intervals that are successive integers of ∂τ.
⎢
Hnτ ⎢⎢⎢⎢ τn ⎥⎥⎥⎥⎥. (2)
⎣ x′τ ⎦ Following this methodology, the aforementioned Eq. 9 is modified to
n
y′τ represent a covariance matrix as:
The state formulation that relates the moving object and the 0.3∂τ 3 0 0.5∂τ 2 0
observation station is listed in Eq. 3 below: ⎢
⎡
⎢
⎢ 3 ⎥⎥⎥
2⎤
⎢
ℓτ ⎢
⎢
0 0.3∂τ 0 0.5∂τ ⎥⎥⎥ω,
⎢
⎢
⎣ 0.5∂τ 2
0 ∂τ 0 ⎥⎥⎦
(10)
xτ
⎡⎢⎢⎢ y ⎤⎥⎥⎥ 0 0.5∂τ 2 0 ∂τ
m n ⎢
Hτ Hτ − Hτ ⎢⎢⎢ τ ⎥⎥⎥⎥⎥.
⎢ (3)
⎣ xτ′ ⎦
meanwhile the constant ω in Eq. 10 defines the spectral magnitude
yτ′
of the Gaussian noise.
The discrete-time WPV approach is applied to construct the
momentum of the maneuvering object under the framework of
state-space science. This design is implemented to define the state 2.2 Coordinated turn (CT) modeling
expression in the following way:
Hτ+1 Xτ Hτ + ατ . (4) The CT design is a commonly used approach for modeling the
kinetic properties of a passive target that undergoes continuous axial
The state model illustrated above defines the dynamic shifting motion. In this kinetic system modeling, rotation rate is an
matrix Xτ, which consists of elements ι × ι. The dynamic shifting additional feature in the state vector. Here, the location, velocity,
matrix reflects the response of the WPV model in the context of and spin rate of the maneuvering vehicle are represented by its state
state-space strategy. It is assumed that process noise ατ in this model features vector in angular dimensions. Through the CT model, it can
follows a Gaussian spectrum having its mean near zero. The be defined in Eq. 11 as:

Ali et al. 10.3389/fphy.2024.1374138
xm
τ noise covariance for the rotational feature is set at a particular
⎡⎢⎢⎢ ym ⎤⎥⎥⎥
⎢⎢⎢ τm ⎥⎥⎥ integer value of 0.15.
Hm ⎢⎢ x′ ⎥⎥⎥.
τ ⎢ ⎢⎢⎢ τm ⎥⎥⎥ (11)
⎢⎢⎣ y′ ⎥⎥⎦
τ
ωmτ 2.3 Measurement modeling
As well as, the formulation of the state feature vector at the
observation point can be done in Eq. 12 in the following manner: In state feature estimation mathematical strategy, each WPV
and CT framework corresponds to the same measurement modeling
xnτ
⎡⎢⎢⎢ yn ⎤⎥⎥⎥ that is developed as well, utilizing the concept of state-space
⎢⎢⎢ τn ⎥⎥⎥
Hnτ ⎢⎢⎢⎢ x′τ ⎥⎥⎥⎥. (12) approach. The computational representation of the measurement
⎢⎢⎢ n ⎥⎥⎥
⎣ y′τ ⎦ modeling can be termed as:
ωnτ
Yτ+1 RHτ+1 , βτ+1 . (22)
The corresponding state vector is presented to explain the relationship
across the monitoring unit and the maneuvering object as: Throughout the time intervals τ, the simultaneous passive
xτ bearings obtained from maneuvering object are compiled in a
⎡⎢⎢⎢ y ⎤⎥⎥⎥ matrix R (.), which is commonly known as measurement matrix.
⎢⎢⎢⎢⎢ τ ⎥⎥⎥⎥⎥
Hτ Hm n
τ − Hτ ⎢ ⎢⎢ xτ′ ⎥⎥⎥. (13) It integrates complex passive bearings that are produced using
⎢⎢⎣ y ′ ⎥⎥⎦
τ the point-slope tangent interaction algebraic technique. The
ωτ parameter β in the aforementioned measurement formula
represents the detected noise at time interval τ, which follows
The governing state formulation of the CT framework is presented
a distinct Gaussian spectrum. The subsequent computation is
in Eq. 14 as:
carried out to develop passive bearings, which are observed
Hτ+1 Xτ Hτ + ϒατ . (14) acoustically based on the actual movement of the moving
object and the placement of monitoring units.
The state formulation in the CT framework closely resembles the
configuration described in Eq. 4 of the Wiener process model with a yτ − Φiy
supplementary variable ϒ, that determines the spatial pattern of R(Hτ+1 ) atan . (23)
xτ − Φix
Gaussian noise ατ. The discrete-time sequence of state equation for
Passive Bearings
the aforementioned framework is derived from an approach
identical to the WPV modeling in Eq. 15 as: The given measurement equation represents the instantaneous
sin(ωτ ℓτ ) cos(ωτ ℓτ )−1 motion of the moving object (yτ, xτ) through bilateral angular
1 0 0
⎡⎢⎢⎢ ωτ ωτ
⎤⎥⎥⎥ axes. As well as the placement of hearing observers, symbolized
⎢⎢⎢ ⎥⎥ 0
⎢⎢⎢ 0 1 1−cos(ωτ ℓτ )
0 ⎥⎥⎥⎥
sin(ωτ ℓτ )
⎡⎢⎢⎢ 0 ⎤⎥⎥⎥ as (Φiy , Φix ), can be seen in this computation. For the monitoring
⎢⎢⎢ ωτ ωτ ⎥⎥⎥ ⎢⎢⎢ ⎥⎥⎥
Hτ+1 ⎢⎢⎢⎢⎢ 0 ⎥ ⎢ ⎥
0 cos(ωτ ℓτ ) −sin(ωτ ℓτ ) 0 ⎥⎥⎥⎥⎥ Hτ + ⎢⎢⎢⎢⎢ 0 ⎥⎥⎥⎥ ατ . (15) unit i at time interval τ, the measurement framework Y, as specified
⎢⎢⎢ ⎥⎥⎥ ⎢⎢⎣ 0 ⎥⎥⎦
⎢⎢⎢ ⎥ in Eq. 22, can be formulated in revised form as:
⎢⎢⎢ 0 0 sin(ωτ ℓτ ) cos(ωτ ℓτ ) 0 ⎥⎥⎥⎥
⎣ ⎦ 1
0 0 0 0 1 ϒ yτ − Φiy
Yiτ atan + βiτ . (24)
Xτ xτ − Φix
The CT paradigm additionally incorporates Gaussian dynamical In the above-mentioned expression, the observed noise βiτ possesses
noise given in Eq. 16, consisting of a zero-valued mean and a a zero mean. Whereas the calculation of its covariance ℵτ is as
covariance equivalent to that specified in Eq. 8. follows in Eq. 25.
ατ ≈ N(0, ℓ τ ). (16)
βiτ N(0, ℵτ ), (25)
The computational analysis of the CT model is displayed in matrix
ℵτ diag(σ 2Y ). (26)
format earlier, even though it has nonlinear dynamics. Consequently,
the CT system can be formulated by employing Eqs 17-21 as: The observed noise standard deviation given in Eq. 26, denoted by
the sign σ, is represented in the above equation. It plays an important
sin(ωτ ℓτ ) cos(ωτ ℓτ ) − ℓτ
xτ+1 xτ + xτ′ + yτ′, (17) role in analyzing the performance of state feature estimation
ωτ ωτ
methodologies in the context of underwater object navigation.
1 − cos(ωτ ℓτ ) sin(ωτ ℓτ )
yτ+1 yτ + xτ′ + yτ′, (18) The dynamics of the subaquatic environment are characterized
ωτ ωτ by the unpredictable nature of the observed standard deviation of
xτ+1
′ cos(ωτ ℓτ )xτ′ − sin(ωτ ℓτ )yτ′, (19) acoustic noise. Different metrics for the observed noise standard
yτ+1
′ sin(ωτ ℓτ )xτ′ + cos(ωτ ℓτ )yτ′, (20) deviation are adopted in our investigation in order to analyze the
ωτ+1 ωτ + ατ . (21) robustness and accuracy of the neuro computing and traditional
Bayesian approaches. By adhering to the prescribed sequence of
In the context of experiments, when anticipating significant maneuvers, a synchronized rotational track can be established for
maneuverability of underwater object, the Gaussian dynamical the underwater object.

Ali et al. 10.3389/fphy.2024.1374138
• The dynamic object commences its motion at a position of The ability to distinguish actual features from erratic noise enhances
x 0 x′ the overall robustness of the algorithm. IBRBNC possesses the
τ and maintaining a constant velocity τ
yτ 0 yτ′ ability to effectively adjust to different degrees of complexity in
1 the network. The proposed deep learning framework successfully
across the x-axis, sets up its moves from the point integrates historical information in order to forecast the future
0
of origin. outcomes of state features associated with submerged passive
• Following a time interval of 4 s, the object undergoes a right navigating object. Within the nonlinear environment of the
twist having a rotation rate specified as ωτ = −1. IBRBNC paradigm, external input and consequent output are
• After completing a duration of 9 s, the object ceases its employed to forecast the future trends of state features. In this
rightward rotation and proceeds to move along normal regard, IBRBNC adopts an extremely efficient multi-layer design
trajectory, maintaining a consistent velocity of 1 unit. consisting of an input layer, an embedded layer, a hold layer, and an
• The aquatic object exhibits leftward movement with a outcome layer, as depicted in Figure 3. In this particular
rotational parameter ωτ = 1, which occurs at the 11 s configuration, the measurement function Y(τ) given in Eq. 24 is
interval over the entire period. applied to the IBRBNC network as input to create estimations
• Around 16 s, the vehicle discontinues its leftward turn and regarding the passive object state vector H(τ) given in Eq. 13.
proceeds to travel straight ahead at a constant velocity for the The most difficult aspect of applying the Bayesian regularization
duration of 4 s. approach is determining the appropriate setting for the desired
function coefficients. The Bayesian theory implemented in neural
systems is founded upon the likelihood analysis of variables inside
3 Neuro computing designing the network. In comparison with the standard procedure for
network training, which selects the best combination of weights
In this portion of the study, a mathematical framework using by reducing divergence, the Bayesian strategy incorporates a
smart neuro computing is developed for state feature estimation of probabilistic spectrum of network weights. Consequently, the
underwater moving object. The link between Bayesian outputs of the network can be characterized by spectrum of
normalization deep learning, and underwater localization is their likelihoods. Let us suppose a Bayesian neuro computing
mutual reliance on Bayesian theories for the purpose of managing architecture that uses a learning data set S, which comprises z
uncertainties. The primary objective of target localization is to numbers of input and target matrix pairs. These pairings are
estimate the motion features of a dynamic object. In contrast, used to train the neural model.
Bayesian regularization neuro computing is primarily concerned S {(Y1 , H1 ), (Y2 , H2 ), . . . , (Yz , Hz )}. (27)
with modeling the ambiguity that arises from neural network
During the training stage, it is preferred to determine a standard
variables. The performance of the estimation approach in state
evaluation metric for calculating the difference among actual and
feature applications is significantly influenced by the presence of
estimated data. The aforementioned metric can be mathematically
noisy bearings. Therefore, the basic knowledge of complicated noisy
represented in the following manner:
bearings or prior information can be employed for the purpose of
modeling IBRBNC in order to obtain the ideal performance of
1 C ^ 2
state features. A ES (S|e, F) Hj − Hj . (28)
C j1
In Eq. 28, ES represents the average sum of squares of the model

3.1 Intelligent Bayesian regularization loss, it is also a measure for prior terminating, which is employed
backpropagation neuro in several computational methods as a means to prevent excessive
computing (IBRBNC) fitting. S denotes the learning data set, which consists of input-
target combos, as defined in Eq. 27. The F is neural network
The IBRBNC is a particular form of neural network model that
layout, which is characterized by its configuration, that includes
implements Bayesian concept to make its training operation
the quantity of layers, the size of elements inside each layer, and
smoother. The basic objective of Bayesian regularization is to
the particular trigger function employed by each element. In an
establish a statistical structure for representing the
IBRBNC model, the process of normalization includes a
unpredictability associated with the key variables of a network,
supplementary parameter in the desired function. This
namely, the weights and biases.
parameter is used to minimize the presence of massive
weights, which can potentially result in more consistent
3.1.1 Mathematical modeling of IBRBNC mapping. In the present scenario, it is appropriate to use the
The IBRBNC framework employs Bayesian approaches, namely,
gradient-based optimization method in order to effectively
weight priors and posterior variations, in order to incorporate
minimize the objective function.
regularization into the system. Regularization plays a crucial role
in mitigating overfitting, hence enhancing the effectiveness of A ηES (S|e, F) + κEweight (e|F). (29)
IBRBNC in scenarios such as limited access to data or noisy
training samples. Taking advantage of the Bayesian approach in The term Eweight (e|F) in Eq. 29 represents the sum of squares of
the context of IBRBNC enables better management of noisy data. model weights, described in Eq. 30 as:

Ali et al. 10.3389/fphy.2024.1374138
1 C 2 X JJ’. (37)
Eweight e . (30)
C j1 j
Here letter J in Eq. 37 represents the Jacobian matrix, which
comprises the partial derivatives of the network failures relative
A pair of hyperparameters η and κ are needed to be computed as
to the network variables. While J can be defined in Eq. 38 as:
coefficients of the desired function. The closing term denoted as
κEweight (e|F), is commonly referred to as weight decay, while κ is ∂μ1 (e) ∂μ1 (e) ∂μ1 (e)
⎢
⎡ / ⎥
alternatively recognized as the decay ratio. If the value of κ is ⎢
⎢
⎢
⎢
⎢ ∂e1 ∂e2 ∂eq ⎤⎥⎥⎥⎥⎥
significantly smaller than η, the neuro computing network will ⎢
⎢
⎢ ⎥⎥⎥
⎢
⎢
⎢ ∂μ2 (e) ∂μ2 (e) ∂μ2 (e) ⎥⎥⎥⎥⎥
decrease mistakes. When κ is significantly larger than η, the ⎢
⎢ / ⎥
⎢
⎢ ∂e
⎢
⎢ ∂e2 ∂eq ⎥⎥⎥⎥⎥
training process will prioritize reducing the size of the weights, J ⎢
⎢
⎢
⎢
⎢
1
⎥⎥⎥. (38)
⎢ ⎥
even if it leads to an increase in network mistakes. As a result, the ⎢
⎢
⎢
⎢ .. .. .. ⎥⎥⎥⎥⎥
⎢
⎢ . . 1 . ⎥⎥⎥
⎢
⎢
⎢ ⎥⎥⎥
network’s outcome will become more convenient. Once the data has ⎢
⎢
⎢
⎣ ∂μQ (e) ∂μQ (e)
⎢ ∂μQ (e) ⎥⎥⎥⎦
been acquired, assuming Gaussian noise in the target samples, the /
subsequent spectrum of the weights in the neural network can be ∂e1 ∂e2 ∂eq
adjusted by applying Bayes’ rule.
The selection between the Gauss-Newton (GN) estimation
PS|e, η, F. P(e|κ, F) approach and the Hessian matrix is an essential factor when
Pe|S, κ, η, F . (31)
PS|κ, η, F implementing the Levenberg-Marquardt (LM) training method
for the optimization of the objective function A. In the context of
Hence, Bayesian regularization incorporates a likelihood the LM algorithm, the variables at the rth loop are modified in
spectrum of network weights, defining the system architecture Eq. 39 as:
as a stochastic platform. In Eq. 31, the variable S represents the
1
training sample, whereas the prior allocation of weights is er+1 er − T JT μ. (39)
J J + εI
formalized in Eq. 32 as:
κ u/2 κ Whereas ε represents the Levenberg attenuation coefficient. The
P(e|κ, F) exp− ee′. (32) variable ε can be adjusted with every repetition, which improves the
2π 2
process of optimization. It is widely adopted as an alternative to the
The neural network layer design is denoted as F, while the vector e GN approach for the purpose of identifying the lowest value of the
represents the weights associated with the architecture. The term given function.
P(e|κ, F) denotes understanding of weights prior to the acquisition
of any dataset. The probability function, indicated as P(S|e, η, F), 3.1.2 Working principle of IBRBNC
represents the likelihood of the happening of dataset pairs, The operation and work flow of the IBRBNC can be summarized
specifically in relation to the design weights. Within the as follows: Begin.
constraints of this Bayesian scheme, the most desirable weights
are those that increase the subsequent likelihood, denoted as 1. Preliminary processing of data:
P(e|S, κ, η, F). The process of improving the subsequent • Regularization of given data
probability of weights e can be regarded as similar to the process • Dividing data into training set, validation set, and test set
of reducing the normalized desired function A = ηES + κEweight. Let 2. Initialization phase:
us assume the combined posterior density in Eq. 33 as: • Layout of the network configured by choosing the number of
PS|κ, η, F. Pκ, η|F neurons, layers and delays
Pκ, η|S, F . (33) • Specify the weights and biases for the pre dispersion values
P(S|F)
3. Training:
It can be observed in Eq. 34 as: • Regarding every training instance:
PS|e, η, FP(e|κ, F) • Feed forward cycle:
PS|κ, η, F , (34) • Calculate the system outcome
Pe|S, κ, η, F
• Determine Probability:
∇A κ, η
PS|κ, η, F v/2 . (35) • Use training data alongside the network outcome to figure
πη (π /κ)u/2 out the probability value
• Figure out preceding:
Whereas v represents the sum of the findings and u indicates the entire
• Using the previous dispersion variables, compute the
set of network coefficients. The Laplace approximation, represented in
preceding value
Eq. 35, yields the subsequent mathematical expression:
• Calculate the posterior:
1 • Use Bayes’ formula to figure out the posterior dispersion
∇A κ, η ∝ exp−AeBAP . (36)
XBAP • Regularization:
• Apply regularization value to adjust weights
The Hessian matrix of the desired function is denoted by XBAP in • Backpropagation:
Eq. 36, while BAP is an acronym that refers to best a posteriori. • Determine weights and biases variations
The Hessian matrix may be estimated as: • Modify variables:

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 3
Architecture of Bayesian regularization neural network.
• Through gradient descent optimization approach, Eq. 13 in CT model represents real state vector that consists of
improve weights and biases five different features: the position along the x-axis, the position
4. Validation: along the y-axis, the velocity along the x-axis, the velocity along the
• Measure effectiveness of the network by y-axis, and a variable related to turning around. The state vector is
validation sample incorporated into the measurement framework, as depicted in Eq.
• Examine for overestimation or convergence 22, in order to calculate passive bearings. The passive bearings
• If required, modify the normalization constants obtained from multiple acoustic hydrophones are processed as
5. Testing: input for the IBRBNC infrastructure. This input serves to estimate
• Utilize the test sample to generate estimations using the the state vector, which consists of five state features. In
trained IBRBNC experiments, a concealed layer comprising 25 neurons is
6. Possible Inference: deployed, wherein the triggering of these neurons is assisted by
• If complete Bayesian inference is needed, then: a sigmoid function. The process of training weights is achieved by
• To estimate the subsequent probability over network’s employing Bayesian regularization-based training methodology,
variables, use Bayesian inference approach (Markov which incorporates the backpropagation through time (BPTT)
Chain Monte Carlo) mechanism. Similarly, the application of the epoch format is
7. Evaluation: carried out throughout the training stage of the neuro
• Evaluate the IBRBNC success using test sample computing paradigm. Through the process of simulation
• Calculate the estimation instability, if necessary experiments using the IBRBNC paradigm, the data is divided
into three parts. Specifically, 70% of the complete dataset is
Finish. assigned to the training stage, while the remaining 30% is
The overall operational framework of the IBRBNC is depicted evenly distributed among the testing and validating processes.
in Figure 4. This allocation allows the evaluation of the outputs generated by
The data that is fed to the IBRBNC design, as shown in Figure 5, the neuro computing network.
consists of the passive bearings Y(τ) acquired from eight acoustical
hydrophones. The bearings information provided to the IBRBNC
network for the purpose of computing the estimated state vector 3.2 Assessment metrics
H(τ) of five state features. This vector is also depicted as the outcome
of the neural system. The layout of the IBRBNC model consists of The evaluation standard for the approach of deep neuro
three layers, namely, the input layer, hidden layer, and output layer, computing refers to the development of a reduced MSE between
as illustrated in Figure 5. the actual and forecasted state features of maneuvering object at

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 5
FIGURE 4 Neural network toolbox flow diagram for IBRBNC.
Flow chart of the IBRBNC paradigm.
4 Discussion on simulation results

every single moment τ. This research provides an investigation of
the accuracy and robustness demonstrated by the neuro computing This segment of the paper presents a brief discussion of the
paradigm. Consequently, the MSE metric to analyze the simulation outcomes for the proposed deep learning approach based
performance of IBRBNC, IMMEKF, and IMMUKF is generated on IBRBNC. The findings comprise on estimates of state features in
independently during every Monte Carlo simulation as: real-time, errors in object’s position on the x and y-axes, divergence
in object’s velocity on the x and y-axes, estimates of its rotation, a
j histogram of inaccuracies and analysis on regression. Five particular
1 2
Error(τ ) HReal − HEstimates . (40)
j τ 1 τ τ circumstances are simulated, and the assessment metrics is standard
deviation of observed Gaussian noise. The level of this metrics is
The moving object’s real state features in the described error Eq. 40 consistently adjusted within a range of 0.01–1 radian. The observed
are represented by the variable HReal τ . In contrast, HEstimates
τ noise demonstrates a complicated maritime atmosphere when
expresses the predicted state features of the object, which are magnitude reaches its maximum, which is 1 radian. The optimal
computed using IBRBNC and Bayesian filtering methodologies. environment, on the other hand, is characterized by the lowest value,
The entire quantity of data samples is represented by the variable j, specifically 0.01 radians. In experiments, it is crucial to accurately
which has an amount of 200 in trials. Additionally, the initial data manipulate the numerical factors of the state estimation model for
sample is given by τ = 1. The calculation of deviations in state better state feature estimation and to obtain the required outcomes.
features for the bending trajectory is performed on every Table 1 presents the optimal setting of the parameters used in state
sample of time. estimation modeling.

Ali et al. 10.3389/fphy.2024.1374138
TABLE 1 Establishing various parameters of state features estimation yτ − Φiy

model. Yiτ atan + βiτ . (43)
xτ − Φix
Parameter Optimal setting
The IBRBNC paradigm takes the above stated measurement data as
The object’s preliminary features H0 = [0 0 1 0 0]T
its input. The true state vector, as explained in Eq. 44, is used in the
Localization function for acoustic observers (Φix , Φiy ) measurement model, as shown in Eq. 22, to find out expected
state features.
Observer units i=8
xτ
Distance among observer units 0.5 ⎢
⎡
⎢
⎢ ⎤⎥⎥
⎢
⎢ y τ ⎥ ⎥⎥⎥
⎢
⎢
Variation in observed noise σ = 0.01 → 1 radians m n ⎢
Hτ Hτ − Hτ ⎢
⎢ xτ′ ⎥⎥⎥⎥. (44)
⎢
⎢ ⎥
⎣ yτ′ ⎥⎥⎦
⎢
Model noise variance for WPV design ℓτ = 0.05
ωτ
Model noise variance for CT design ℓτ = 0.15
The first scenario demonstrates real-time state approximations, real
Segment space ∂τ = 0.1
and observed rotation rates, errors in x-axis and y-axis location,
Segments of trajectory 200 divergences in x-axis and y-axis velocity, an error histogram, and a
regression study. All findings compare the true and estimated state
Neurons in network 25
features for σ = 0.01 radians observed noise standard deviation. For
Total instances in IBRBNC network 1,000 this scenario, we assume that the marine environment is steady by
setting a low integer level for the detected noise.
4.1 Assessment of navigating object state • In Figure 6A, a comparison is made between the ability of the
features with different levels of observed IBRBNC network and filtering techniques IMMEKF and
noise standard deviations IMMUKF in the prediction of state features for a
navigating object throughout the identification of the
This part provides a comprehensive analysis of the simulation turning route. It is worth mentioning that IBRBNC
findings and a detailed discussion on the real-time estimates of state effectively tracks the exact trajectory of the maneuvering
features for a dynamic object. The study implements Bayesian object, hence demonstrating its superior precision when
estimation approaches, including IMMEKF, IMMUKF, and compared to the other two approaches.
IBRBNC. This investigation is aimed to compare the predictability • Figure 6B depicts the rotational component estimates of the
and resolution efficiency of the IBRBNC strategy with IMM Kalman IMMEKF, IMMUKF, and IBRBNC strategies. The better
filters. The investigation mainly emphasizes on five distinct levels of results of the IBRBNC method compared to the IMM
recorded noise standard deviation. The statistics from Figures 6–15 predictors are consistently observed throughout the process,
exhibit diverse findings obtained from the filtering algorithms and as evidenced by the precise estimation of the turning feature
IBRBNC paradigm. These findings represent state estimates, x-axis for all data points.
position error, y-axis position error, x-axis velocity error, y-axis velocity • The analysis of the x-axis position inaccuracy is presented in
error, rotation forecasts, error histogram, and regression study. The Figure 6C using the mean square approach. Notably, the IBRBNC
subsequent parts provide an assessment of five independent scenarios algorithm exhibits minimal average error when compared to
associated with algebraic formulations and simulation outcomes. other approaches, except for a single conspicuous spike.
• The simulation results for the y-axis position error of the
IMMEKF, IMMUKF, and IBRBNC approaches are presented
4.1.1 Condition 1: The observed noise in Figure 6D. It is obvious that IBRBNC displays occasional
standard deviation σ = 0.01 radians spikes while also showcasing a higher average performance
when compared to the other approaches. This highlights the
Gaussian noise βiτ is quantified by 0.01 radians standard effectiveness of IBRBNC in reducing errors in y-axis
deviation in the first condition. This level indicates a nearly ideal positioning, demonstrating its technical superiority.
oceanic atmosphere with consistently calm circumstances. In this • Figure 6E illustrates the disparity between the true velocity and
specific instance, the derivation of covariance which contains the estimated velocity along the x-axis, which is measured in MSE
standard deviation of observed noise is performed in Eq. 41 as: context. The deep learning mechanism developed by the
IBRBNC demonstrates significant computational efficiency in
ℵτ diag(σ 2Y ). (41) estimating the velocity along the x-axis. It surpasses the
The modeling of the observed noise at particular time interval τ for i performance of nonlinear Kalman estimators when evaluated
sensor is established by employing covariance as: using 200 samples. Nevertheless, IBRBNC exhibits small number
of spikes that occur during turns of trajectory.
βiτ N(0, ℵτ ). (42) • The simulation results for the y-axis velocity error of the
IMMEKF, IMMUKF, and IBRBNC approaches are depicted in
The above stated observed noise in Eq. 42 is adding in measurement Figure 6F. It is worth noting that IBRBNC demonstrates
expression Y as defined for entire passive bearings in Eq. 43 as: limited sharp climbs while consistently displaying

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 6
The outcomes of IMMEKF, IMMUKF, and IBRBNC for σ = 0.01 radians.
commendable overall performance in comparison to approximated output HEstimates

τ of the IBRBNC network.
alternative methodologies. The deviation histogram consists of a set of divergence
• Figure 6G displays the deviation histogram that represents the points, encompassing both negative and positive samples.
discrepancies among the target data HReal τ , and the The error values enable to quantify the discrepancy

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 7
The IMMEKF, IMMUKF, and IBRBNC’s MSEs for estimating bidirectional position and velocity in condition 1.
between the input dataset and the computed dataset. In this deliberately apply a certain amount of observed noise to the
experiment, the total error of the IBRBNC network is divided computational process. In this specific situation, the calculation
into 20 bins, which are graphically depicted as perpendicular of covariance is obtained in Eq. 45 by utilizing the standard
segments. The histogram displays an obvious disparity in the deviation of observed noise as:
central box, as it demonstrates a discrepancy of 0.001122 with
samples of 600. This observation implies that several data
points display errors that lie within a specific zone. The zero- The noise reported for time period τ at hydrophone i belongs to the
error line is also located within this box, indicating the neural Gaussian spectrum, with its parameters given by the formerly
network’s zero error. computed covariance in Eq. 46 as:
• Figure 6H depicts the regression investigation of the IBRBNC
method over the training, validation, and testing phases. The βiτ N(0, ℵτ ). (46)
neuro computing platform employs a partitioning strategy to
The measurement mechanism of sensor i includes observed noise at
divide the input samples into three distinct subsets: training,
every point in time τ as:
validation, and testing. The aforementioned subsets are
assigned proportions of 70%, 15%, and 15% respectively. yτ − Φiy
This inspection utilizes probabilistic elements to illustrate xτ − Φix
the relationship between the findings of state features
HEstimates
τ and the corresponding actual state features HReal τ .
Likewise, the neural paradigm integrates this designed measuring
The regression modeling reveals a substantial level of model Yiτ given in Eq. 47 as an input. The desired value of IBRBNC is
correlation between the actual and predicted state features. denoted by the subsequent vector of state features given in Eq. 48 as:
A flat trend is observed in the data sets that reflect the degree of
robustness of the IBRBNC paradigm. xτ
⎢
⎡
⎢
⎢ ⎤⎥⎥
⎢
⎢ y τ ⎥ ⎥⎥⎥
⎢
⎢
m n ⎢
Hτ Hτ − Hτ ⎢
⎢ xτ′ ⎥⎥⎥⎥. (48)
⎢
⎢ ⎥
⎣ yτ′ ⎥⎥⎦
Furthermore, the total MSEs for both bidirectional locations and ⎢
velocities of the underwater object are computed by contrasting the ωτ
observed values with the predicted values. The results of this
condition suggest that the accuracy of the IBRBNC is superior The state elements used in the estimation process exploiting the
than IMM Kalman filters when considering the responses of IBRBNC methodology consist of the instantaneous direction of
position and velocity errors. This illustrates the effectiveness of the moving object (xτ, yτ), its velocity (xτ′, yτ′), and its rotation
employing neuro computing for the estimation of state features in rate ωτ. The subsequent section showcases the simulation
underwater maneuvering target scenarios. The graphs presented outcomes, encompassing state estimations, rotational
below illustrate the bidirectional position and velocity errors approximations, errors in the position of the x and y-axes,
associated with the IMMEKF, IMMUKF, and IBRBNC techniques. errors in the velocity of the x and y-axes, a histogram
illustrating divergence, and an analysis of regression for the
current condition.
4.1.2 Condition 2: The observed noise
standard deviation σ = 0.05 radians • Figure 8A displays a comparative analysis of all three
techniques for state feature estimation by following the
In condition 2, the standard deviation of observing noise is turning route of the maneuvering target. It is important to
chosen to be σ = 0.05 radians. The intent of this selection is to highlight that, in this particular case, the motion estimations

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 8
and accuracy of IBRBNC are superior to conventional of the IBRBNC method over standard filtering techniques is
nonlinear filtering approaches. achieved for effectively determining the rotation variable.
• The estimated turn rates for the present condition of measurement • The average MSE across actual and predicted x-axis positions
noise are depicted in Figure 8B. Yet again, the superior performance of the underwater navigation object is depicted in Figure 8C.

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 9
Furthermore, this outcome illustrates the degree of accuracy position and velocity errors derived from IBRBNC and Bayesian
exhibited by the IBRBNC approach in comparison to the Kalman filters.
IMMEKF and IMMUKF strategies.
• Likewise, y-axis MSE position error of all estimation
algorithms is shown in Figure 8D. In this simulation result, 4.1.3 Condition 3: The observed noise
IMMEKF and IMMUKF are facing large spikes between standard deviation σ = 0.1 radians
120 and 140 samples. These large spikes are due to their
divergence at the left turn of the trajectory. In comparison, Under this condition, the observed noise standard deviation
IBRBNC is showing the minimum position MSE for all increases to σ = 0.1 radians, suggesting that a significant amount of
200 samples of the turning trajectory. interference has been added in the whole model. Considering
• An analysis of a maneuvering vehicle’s actual and estimated standard deviation of 0.1 radians for random detected noise, the
x-axis velocity, computed by IMM filtering techniques and the variance ℵ at time step τ is computed in Eq. 49 as follows:
IBRBNC network, is presented in Figure 8E. The findings
show that, in terms of competence, the neuro computing
technology executes better than the filtering strategies. The mathematical description of the measured noise that is derived
• Figure 8F displays a parallel view of the true and estimated from the above covariance for i sensor at time interval τ is given in
velocity along the y-axis, as computed by all techniques. The Eq. 50 as follows:
results also endorse that the neural learning is estimating
better y-axis velocity. βiτ N(0, ℵτ ). (50)
• The error histogram in Figure 8G evaluates the network error The observational model is including the achieved Gaussian
between the target dataset HReal τ , and estimated dataset measured noise as:
HEstimates
τ of the target’s state features. The bar graph
yτ − Φiy
exhibits an inclined bin positioned at its midpoint, which Yiτ atan + βiτ . (51)
signifies an error of −0.00505. This segment is associated with xτ − Φix
the input samples and contains an elevation exceeding
The above Eq. 51 represents the measurement model Y for
200 events. In the given scenario, the zero error is
hydrophone i at time step τ. This model combines the passive
positioned underneath the middle vertical column.
orientations observed by acoustic hydrophones, and correlates
• Figure 8H depicts the IBRBNC computed regression assessment
them with the white Gaussian distributed measured noise. The
for neural learning. The regression approach demonstrates the
data set utilized for deep neuro computing is comprised on the
usefulness of the IBRBNC framework by illustrating the
calculations of the measurement model. The neural network’s
correlation between the actual intake and the
output consists of state features following Eq. 52, which are
estimated outcome.
presented in state vector as:
The numerical values of average MSEs are also computed to xτ

⎢
⎡
⎢
⎢ ⎤⎥⎥
⎢
⎢ y τ ⎥ ⎥⎥⎥
quantify the difference between the actual and predicted ⎢
⎢
m n ⎢
Hτ Hτ − Hτ ⎢
⎢ xτ′ ⎥⎥⎥⎥. (52)
⎢
⎢ ⎥
⎣ yτ′ ⎥⎥⎦
bidirectional positions and velocities of the dynamic object in the ⎢
presence of σ = 0.05 radians noise. The position and velocity errors ωτ
provide more evidence that potency of the IBRBNC exceeds that of
IMM Kalman filters. This demonstrates the efficacy of neural The following part provides the simulation results on this value of
networks in forecasting state characteristics. Figure 9 lists the measured noise in terms of trajectory predictions, turning rate

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 10
estimates, least bidirectional location errors, divergences in x and • The state features estimation performance for the
y-axes velocities, network failure histogram, and a synchronized spin track is plotted in Figure 10A for all
regression study. techniques. It is evident that the deep learning model based

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 11
on IBRBNC demonstrates superior accuracy when compared nearly 350 points, and the testing samples likewise fall
to the conventional IMMEKF and IMMUKF methodologies in within this limit. The investigation reveals that the zero
this particular scenario. The IBRBNC technique demonstrates error is situated inside this upward box, such as an integer
better capability for precisely determining the dynamics of of 0.02482.
maneuvering object over bends of the route, in contrast to • Figure 10H depicts the regression conditions that occur
conventional filters, which face larger hurdles in this regard. throughout the learning process of IBRBNC. The graph
This shows the effectiveness of the neural approach. illustrates a slight discrepancy between the input and outcome
• Figure 10B illustrates the turn rate predictions based on all data, which can be indicated by increase in the observed noise.
methods. Neuro computing demonstrates far better
estimation outcomes and achieve turn rates that are close The current circumstance entails calculating the MSEs for both
to the actual values, surpassing the capabilities of bidirectional orientations and velocities. The units used for these
Kalman filters. calculations are meter and meter per second, respectively. The
• The plot shown in Figure 10C depicts the real-time mean recognition of position and velocity lapses helps to verify the
square positioning inaccuracy along the x-axis for IMM and previously described results, providing evidence that the neural
IBRBNC estimation approaches. IBRBNC demonstrates high system exhibits significantly higher convergence in comparison to
precision compared to other techniques in all instances of multi model filtering methods. The graph presented in Figure 11
turning course. illustrates the bidirectional position and velocity inaccuracies obtained
• In Figure 10D, position error along the y-axis is represented through the implementation of the IMMEKF, IMMUKF, and
for all state feature estimation algorithms. IBRBNC is IBRBNC algorithms.
experiencing some large spikes near 100 data points, while
between 120 and 180 samples, the performance of IMM filters
is poor. As a whole, the average mean square error of neural 4.1.4 Condition 4: The observed noise
learning is better than conventional techniques. standard deviation σ = 0.5 radians
• Figure 10E illustrates the velocity inaccuracy along the x-axis,
calculated in meter per second, of the navigation vehicle for every The standard deviation of noise in passive observations is
single sample over all methods. All estimation mechanisms are increased to σ = 0.5 radians in this case, so adding a large level of
facing decline in accuracy at the intersections of the turning path. Gaussian disturbance into the state feature estimation mechanism.
Yet, it is obvious that IBRBNC technique surpasses the IMM Here is the formulation of variance in Eq. 53, which includes this
filters in performance across all sample points. numerical value of the standard deviation of the observed noise as:
• Mean square velocity error computed by all estimation ℵτ diag(σ 2Y ). (53)
algorithms along y-axis of the underwater dynamic object is
shown in Figure 10F. In this result, the estimation accuracy of As well as, the calculation of observed Gaussian noise is derived in
IBRBNC method is steady for all data points while IMM Eq. 54 from the covariance in the following way:
filtering methods are showing large fluctuations, especially at
βiτ N(0, ℵτ ). (54)
turns of the target trajectory.
• Figure 10G displays a histogram comparing the error of The comprehensive framework’s measurement equation
neural learning between target HReal τ , and the expected incorporates the noise recorded at time step τ for each observer
dynamic features HEstimates
τ of the turning object. The element i in the subsequent form:
histogram features a centrally positioned upward line,
yτ − Φiy
that corresponds to an error value of 0.02482. The Yiτ atan + βiτ . (55)
maximum value of the vertical bin used for training is xτ − Φix

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 12

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 13
The IBRBNC network is given the complete measurement model all algorithms are providing large peaks of error due to
formulation Yiτ in Eq. 55 as input dataset. It includes passive enhance noise level. In comparison, the y-axis position
bearings and observed noise. The desired statistics correspond to prediction performance of neural computing is better than
the actual state vector, as depicted in Eq. 56 here: other two techniques.
• Figure 12E exhibits the discrepancy in velocity inaccuracy
xτ along the x-axis, which arises from different methods, thereby
⎡⎢⎢⎢ y ⎤⎥⎥⎥
⎢⎢⎢ τ ⎥⎥⎥ confirming the efficacy of the neural network model.
Hτ Hm n ⎢⎢⎢ x ′ ⎥⎥⎥.
τ − Hτ ⎢ ⎢⎢⎢ τ ⎥⎥⎥⎥
(56)
⎣ yτ′ ⎦ • The y-axis mean square velocity errors for all samples of
ωτ turning trajectory are shown in Figure 12F. In this finding, the
estimation accuracy of IBRBNC is far better than conventional
The bidirectional positions, velocities, and rotations of underwater techniques for all data points.
moving vehicle are computed here in real-time for the precise • A divergence histogram shown in Figure 12G presents the
curving path. The motion features are obtained via IMM Kalman frequency of deviations across the target information HRealτ ,
filters and neural network leveraging IBRBNC. The actual state vector is and the estimated target’s motion characteristics HEstimates
τ .
employed to estimate the desired characteristics of the underwater The difference of −0.00828 is observed in the massive frame
object. The diagrams below illustrate simulation findings for path positioned in the middle of the chart. In the histogram, the
predictions, rotational approximations, oversights in x and y-axes zero-error line is situated beneath the vertical bar with a
locations, errors in x and y-axes velocities, a histogram of deviations, numerical value of −0.00828.
and the description of regression. • The result of the regression in the given case is depicted in
Figure 12H, which suggests a substantial gap between the
• The state feature estimation capability of Bayesian filtering input dataset and the expected outcome. The gap can be
and IBRBNC schemes is shown in Figure 12A, assuming a lot identified as an increase in the standard deviation of the
of observation distortion. It is notable that each of these observed noise.
strategies faces difficulties when attempting to precisely
identify the real route of the target. Nevertheless, despite To assess the difference between the actual and predicted
the existence of an adverse underwater environment, it can bidirectional velocity and location of the moving vehicle, the
be seen that the IBRBNC algorithm demonstrates a greater MSE is derived here. The numerical position and velocity
level of coherence with the true trajectory in comparison to the divergences strengthen the prior outcomes that IBRBNC
other methodologies under consideration. demonstrates superior efficiency in comparison to Kalman filters.
• The rotation rate estimates in this specific scenario are The position and velocity fluctuations, as computed with the
illustrated in Figure 12B. This demonstrates that neural IMMEKF, IMMUKF, and IBRBNC techniques, are depicted
computation has more effective predictive approach than in Figure 13.
IMM filtering methods.
• The schematic diagram denoted as 12c depicts the MSE
among the actual and estimated x-axis coordinates of the 4.1.5 Condition 5: The observed noise
moving target. The results show that filtration techniques have standard deviation σ = 1 radian
a notable margin of error, whereas the IBRBNC method
provides better convergence with less average positional In the concluding scenario of this research, a maximum
inaccuracy. magnitude of σ = 1 radian is chosen to depict an environment
• The underwater object’s mean square position error along the that contains an extreme amount of turbulence. Furthermore, the
y-axis is represented in Figure 12D. This outcome shows that measurement framework with a significant level of noise is

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 14
The outcomes of IMMEKF, IMMUKF, and IBRBNC for σ = 1 radian.
employed in the state feature estimation mechanism. The maximum The mathematical modeling of Gaussian distributed observed noise βiτ
amount of σ relates to the term of covariance ℵτ, given in Eq. 57 as: for sensor i takes place at time τ through the covariance ℵτ in Eq. 58 as:
ℵτ diag(σ 2Y ). (57) βiτ N(0, ℵτ ). (58)

Ali et al. 10.3389/fphy.2024.1374138
FIGURE 15
The measurement formulation in Eq. 59 reflects the presence of significantly smaller x-axis position error in comparison to the
excessive amount of distinct white Gaussian observation noise as: IMMEKF and IMMUKF techniques.
• In Figures 14A, D comprehensive comparison between
yτ − Φiy IBRBNC and IMM Kalman filters is done for the
xτ − Φix estimation of y-axis mean position error. In this
comparative analysis, neuro computing is showing far
The input set of data denoted as Yiτ for the IBRBNC network, better convergence rate and completely dominating IMM
comprises of passive measurements incorporating the highest level filters for all data points of the turning trajectory.
of noise. The output of the neuro computing is represented by the • Figure 14E illustrates the results of the x-axis velocity deviation
state vector, as indicated in Eq. 60 here: for all methods. It shows that IBRBNC occasionally has spikes,
whereas as a whole it performs better than IMM
xτ filtering methods.
⎡⎢⎢⎢ y ⎤⎥⎥⎥
⎢⎢⎢ τ ⎥⎥⎥ • Y-axis velocity error analysis is done in Figure 14F for all
Hτ Hm n
τ − Hτ ⎢
⎢⎢⎢ xτ′ ⎥⎥⎥⎥. (60)
⎢⎢⎢ ⎥⎥⎥ applied techniques. In this finding, IBRBNC is showing far
⎣ yτ′ ⎦
ωτ better performance in the start and end of the trajectory while
experiencing some difficulties in the middle phase of the
The IBRBNC system combines observations and dynamic functions target’s trajectory. In contrast, the performance of IMM
in a particular order to precisely calculate the output dataset, which filters is getting worse after 120 data points in the
includes the estimated vector of state features. The presented trajectory. Again, IBRBNC is surpassing conventional
information in the below part includes path tracking, rotation techniques in this analysis too.
predicts, bidirectional locations and velocities discrepancies, error • Figure 14G demonstrates the investigation of neural network
bar chart, and regression approach. by executing the error statistic spectrum. This investigation
involves the specific target dataset HReal τ , along with the
• Figure 14A shows the turning trajectory estimates of estimated output dataset HEstimates τ . A disparity
IMMEKF, IMMUKF, and IBRBNC algorithms in this of −0.07229 has been identified in an upward box with a
specific scenario. The existence of considerable noise in training dataset altitude near 170 steps, whereas the testing
the underwater environment is the reason for the dataset has heights within the range of 170–180 steps. In this
dispersion of route tracking findings among all specific situation, the zero-divergence score is located beneath
approaches. Accurately determining the exact path of the vertical box, whose center point is −0.07229.
object presents considerable difficulties across all • Figure 14H displays the regression graph of the IBRBNC
computational methods. Nevertheless, in this complex system, which demonstrates a statistical association between
circumstance, the intelligent technology referred to as the input dataset and the intended outcome in this condition.
IBRBNC has superior accuracy in forecasting turning The regression evaluation shows a significant difference
path in contrast to conventional methodologies. among the input values and the expected output, which is
• In this high noise level condition, Figure 14B demonstrates the due to the huge amount of uniformly dispersed noise in the
investigation of rotation rate estimations, where the rotation state feature computation architecture.
feature is more precisely approximated with IBRBNC.
• Figure 14C displays the MSE representing the average The average MSEs for the bidirectional location and velocity of
disparity among the actual and anticipated x-axis the maneuvering vehicle are also numerically calculated in meter
coordinates of the underwater moving vehicle. The findings and meter per second correspondingly, within this cluttered
conclusively indicate that the IBRBNC paradigm reveals a surrounding. These are determined through contrasting the

Ali et al. 10.3389/fphy.2024.1374138
actual values of the state features with their respective approximate features for objects that are highly maneuverable, even when there is
values. The results of the bidirectional location and velocity non-Gaussian measurement noise. Exploring this field of research
deviations confirm the prior findings that IBRBNC demonstrates has great potential in underwater atmospheres involving single or
significantly greater precision than Kalman filters, particularly in many targets.
extremely noisy underwater environments. Figure 15 displays a
collection of bidirectional position and velocity mean square
errors that have been calculated using the IMMEKF, IMMUKF, Data availability statement
and IBRBNC techniques.
The statistics compiled from various conditions indicate that The original contributions presented in the study are included in
when the standard deviation of observed noise σ is high, all state the article/supplementary material, further inquiries can be directed
feature estimation methods encounter difficulties in accurately to the corresponding author.
monitoring the true state of navigational target in marine
environment. After doing a thorough comparison of multiple
approaches, it is obvious that deep learning using IBRBNC Author contributions
surpasses other strategies in terms of performance. This
exceptional results of IBRBNC highlights its capacity to WA: Conceptualization, Data curation, Methodology, Software,
effectively forecast nonlinear real-time state features in Visualization, Writing–original draft, Writing–review and editing.
underwater problems. MB: Data curation, Formal Analysis, Methodology, Resources,
Supervision, Validation, Writing–review and editing. AA:
Funding acquisition, Project administration, Resources,
5 Conclusion Validation, Writing–review and editing. AJ: Formal Analysis,
Funding acquisition, Investigation, Project administration,
The study introduces the ability of deep learning, specifically Validation, Writing–review and editing. AM: Funding
exploiting the robust IBRBNC paradigm, to accurately estimate state acquisition, Writing–review and editing. SH: Resources, Software,
features in real-time for a Markov chain undersea object using just Supervision, Writing–review and editing.
bearing information. The investigation aims to accurately estimate
the instantaneous motion features of a kinematic turning target
within a two-dimensional x-y coordinate framework at each time Funding
instant. The analysis begins by using a mathematical BOT approach
to create a model for estimating the state space of the target in both The author(s) declare that financial support was received for the
the dynamic and measurement frameworks. Subsequently, a neural research, authorship, and/or publication of this article. This work
computing technique based on IBRBNC is presented to predict the was supported by the National Defense Science and Industry Bureau
state features of the Markov chain passive object. The performance Stability Support Fund (Grant No. JSGG20220831103800001).
analysis of the IBRBNC network involves an extensive investigation
to find the exact trajectory of target movements using rotation
estimates, minimum mean square bidirectional location error, real- Acknowledgments
time path tracking, bidirectional velocity difference, error histogram,
and linear regression. The assessment is conducted using a dataset We are grateful to Prof. Qiao Gang from the College of
consisting of 200 samples. Further review involves challenging the Underwater Acoustic Engineering, Harbin Engineering
suggested method to different numerical values of Gaussian University, Harbin for his support and assistance.
distributed observed noise, leading to a better understanding of
its resilience. The outcomes of the simulation in the final section
highlight the neural network’s higher level of accuracy in Conflict of interest
comparison to conventional nonlinear multi-model Bayesian
filtering techniques such as IMMEKF and IMMUKF. Recognizing The authors declare that the research was conducted in the
the rapid performance decline over all techniques for noisy absence of any commercial or financial relationships that could be
measurements, highlights the difficult task of obtaining accurate construed as a potential conflict of interest.
state features in cluttered oceanic situations. This research
represents a significant advancement in enhancing the capacity of
underwater systems to estimate the features of moving object. Publisher’s note
Ultimately leading to improved decision-making and safety in
aquatic missions. The establishment of the IBRBNC paradigm All claims expressed in this article are solely those of the authors
presents a great opportunity for further progress in underwater and do not necessarily represent those of their affiliated
state estimation. It enables new possibilities for applications in organizations, or those of the publisher, the editors and the
marine science, defense, and underwater exploration. reviewers. Any product that may be evaluated in this article, or
Future research endeavors could investigate the use of recurrent claim that may be made by its manufacturer, is not guaranteed or
and radial base neural techniques to improve the estimate of state endorsed by the publisher.

Ali et al. 10.3389/fphy.2024.1374138
References
1. Yuan S, Li Y, Bao F, Xu H, Yang Y, Yan Q, et al. Marine environmental monitoring 25. Ji D, Yao X, Li S, Tang Y, Tian Y. Model-free fault diagnosis for autonomous
with unmanned vehicle platforms: present applications and future prospects. Sci Total underwater vehicles using sequence convolutional neural network. Ocean Eng (2021)
Environ (2023) 858:159741. doi:10.1016/j.scitotenv.2022.159741 232:108874. doi:10.1016/j.oceaneng.2021.108874
2. Luo J, Han Y, Fan L. Underwater acoustic target tracking: a review. Sensors (2018) 26. Cheng X, Li G, Ellefsen AL, Chen S, Hildre HP, Zhang H. A novel densely
18:112. doi:10.3390/s18010112 connected convolutional neural network for sea-state estimation using ship motion
data. IEEE Trans Instrumentation Meas (2020) 69:5984–93. doi:10.1109/tim.2020.
3. Toky A, Singh RP, Das S. Localization schemes for underwater acoustic sensor
2967115
networks-a review. Comp Sci Rev (2020) 37:100241. doi:10.1016/j.cosrev.2020.100241
27. Zuberi HH, Liu S, Bilal M, Alharbi A, Jaffar A, Mohsan SAH, et al. Deep-neural-
4. Huy DQ, Sadjoli N, Azam AB, Elhadidi B, Cai Y, Seet G. Object perception in
network-based receiver design for downlink non-orthogonal multiple-access
underwater environments: a survey on sensors and sensing methodologies. Ocean Eng
underwater acoustic communication. J Mar Sci Eng (2023) 11:2184. doi:10.3390/
(2023) 267:113202. doi:10.1016/j.oceaneng.2022.113202
jmse11112184
5. Himri K, Ridao P, Gracias N. Underwater object recognition using point-features,
28. Lin C, Cheng Y, Wang X. A convolutional neural network particle filter for uuv
bayesian estimation and semantic information. Sensors (2021) 21:1807. doi:10.3390/
target state estimation. IEEE Trans Instrumentation Meas (2022) 71:1–12. doi:10.1109/
s21051807
tim.2022.3169539
6. Sánchez PJB, Papaelias M, Márquez FPG. Autonomous underwater vehicles:
29. Ali W, Khan WU, Raja MAZ, He Y, Li Y. Design of nonlinear autoregressive
instrumentation and measurements. IEEE Instrumentation Meas Mag (2020) 23:
exogenous model based intelligence computing for efficient state estimation of
105–14. doi:10.1109/mim.2020.9062680
underwater passive target. Entropy (2021) 23:550. doi:10.3390/e23050550
7. Aman W, Al-Kuwari S, Muzzammil M, Rahman MMU, Kumar A. Security of
30. Cerqueira V, Torgo L, Mozetič I. Evaluating time series forecasting models: an
underwater and air–water wireless communication: state-of-the-art, challenges
empirical study on performance estimation methods. Machine Learn (2020) 109:
and outlook. Ad Hoc Networks (2023) 142:103114. doi:10.1016/j.adhoc.2023.
1997–2028. doi:10.1007/s10994-020-05910-7
103114
31. Zhang Y, Li C, Wang H, Wang J, Yang F, Meriaudeau F. Deep learning aided ofdm
8. Yang J, Huo J, Xi M, He J, Li Z, Song HH. A time-saving path planning scheme for
receiver for underwater acoustic communications. Appl Acoust (2022) 187:108515.
autonomous underwater vehicles with complex underwater conditions. IEEE Internet
doi:10.1016/j.apacoust.2021.108515
Things J (2022) 10:1001–13. doi:10.1109/jiot.2022.3205685
32. Polson NG, Sokolov V. Bayesian regularization: from tikhonov to horseshoe.
9. Zhang B, Ji D, Liu S, Zhu X, Xu W. Autonomous underwater vehicle navigation: a
Wiley Interdiscip Rev Comput Stat (2019) 11:e1463. doi:10.1002/wics.1463
review. Ocean Eng (2023) 273:113861. doi:10.1016/j.oceaneng.2023.113861
33. Gecili E, Sivaganesan S, Asar O, Clancy JP, Ziady A, Szczesniak RD. Bayesian
10. Menaka D, Gauni S, Manimegalai CT, Kalimuthu K. Challenges and vision of
regularization for a nonstationary Gaussian linear mixed effects model. Stat Med (2022)
wireless optical and acoustic communication in underwater environment. Int
41:681–97. doi:10.1002/sim.9279
J Commun Syst (2022) 35:e5227. doi:10.1002/dac.5227
34. Sariev E, Germano G. Bayesian regularized artificial neural networks for the
11. Ali W, Li Y, Chen Z, Raja MAZ, Ahmed N, Chen X. Application of spherical-radial
estimation of the probability of default. Quantitative Finance (2020) 20:311–28. doi:10.
cubature bayesian filtering and smoothing in bearings only passive target tracking.
1080/14697688.2019.1633014
Entropy (2019) 21:1088. doi:10.3390/e21111088
35. Abdullah AA, Hassan MM, Mustafa YT. A review on bayesian deep learning in
12. Khodarahmi M, Maihami V. A review on kalman filter models. Arch Comput
healthcare: applications and challenges. IEEE Access (2022) 10:36538–62. doi:10.1109/
Methods Eng (2023) 30:727–47. doi:10.1007/s11831-022-09815-7
access.2022.3163384
13. Herrera PA, Marazuela MA, Hofmann T. Parameter estimation and uncertainty
36. Liu T, Liu Y, Liu J, Wang L, Xu L, Qiu G, et al. A bayesian learning based scheme
analysis in hydrological modeling. Wiley Interdiscip Rev Water (2022) 9:e1569. doi:10.
for online dynamic security assessment and preventive control. IEEE Trans Power Syst
1002/wat2.1569
(2020) 35:4088–99. doi:10.1109/tpwrs.2020.2983477
14. Duník J, Biswas SK, Dempster AG, Pany T, Closas P. State estimation methods in
37. Ng KH, Tatinati S, Khong AW. Grade prediction from multi-valued click-stream
navigation: overview and application. IEEE Aerospace Electron Syst Mag (2020) 35:
traces via bayesian-regularized deep neural networks. IEEE Trans Signal Process (2021)
16–31. doi:10.1109/maes.2020.3002001
69:1477–91. doi:10.1109/tsp.2021.3057691
15. Jahan K, Rao SK. Implementation of underwater target tracking techniques for
38. Zhou C, Zhang J, Liu J, Zhang C, Shi G, Hu J. Bayesian transfer learning for object
Gaussian and non-Gaussian environments. Comput Electr Eng (2020) 87:106783.
detection in optical remote sensing images. IEEE Trans Geosci Remote Sensing (2020)
doi:10.1016/j.compeleceng.2020.106783
58:7705–19. doi:10.1109/tgrs.2020.2983201
16. Tong F, Griffiths T. Hydrodynamics for subsea systems. In: Encyclopedia of ocean
39. Ganji A, Minet L, Weichenthal S, Hatzopoulou M. Predicting traffic-related air
engineering. Singapore: Springer (2022). p. 751–7.
pollution using feature extraction from built environment images. Environ Sci Tech
17. Abbas MT, Jibran MA, Afaq M, Song WC. An adaptive approach to vehicle (2020) 54:10688–99. doi:10.1021/acs.est.0c00412
trajectory prediction using multimodel kalman filter. Trans Emerging
40. Nguyen C, Li X, Blanton S, Li X. Correlated bayesian co-training for virtual
Telecommunications Tech (2020) 31:e3734. doi:10.1002/ett.3734
metrology. IEEE Trans Semiconductor Manufacturing (2022) 36:28–36. doi:10.1109/
18. Ali W, Li Y, Raja MAZ, Ahmed N. Generalized pseudo bayesian algorithms for tsm.2022.3217350
tracking of multiple model underwater maneuvering target. Appl Acoust (2020) 166:
41. Balram D, Lian KY, Sebastian N. Air quality warning system based on a localized
107345. doi:10.1016/j.apacoust.2020.107345
pm2. 5 soft sensor using a novel approach of bayesian regularized neural network via
19. Qiu J, Xing Z, Zhu C, Lu K, He J, Sun Y, et al. Centralized fusion based on forward feature selection. Ecotoxicology Environ Saf (2019) 182:109386. doi:10.1016/j.
interacting multiple model and adaptive kalman filter for target tracking in underwater ecoenv.2019.109386
acoustic sensor networks. IEEE Access (2019) 7:25948–58. doi:10.1109/access.2019.
42. Ye L, Jabbar SF, Abdul Zahra MM, Tan ML. Bayesian regularized neural network
2899012
model development for predicting daily rainfall from sea level pressure data:
20. Kumar M, Mondal S. Recent developments on target tracking problems: a review. investigation on solving complex hydrology problem. Complexity (2021) 2021:1–14.
Ocean Eng (2021) 236:109558. doi:10.1016/j.oceaneng.2021.109558 doi:10.1155/2021/6631564
21. Akhtar S, Habibi S. The interacting multiple model smooth variable structure filter 43. Price MA, McEwen JD, Pratley L, Kitching TD. Sparse bayesian mass-mapping
for trajectory prediction. IEEE Trans Intell Transportation Syst (2023) 24:9217–39. with uncertainties: full sky observations on the celestial sphere. Monthly Notices R
doi:10.1109/tits.2023.3271295 Astronomical Soc (2021) 500:5436–52. doi:10.1093/mnras/staa3563
22. Kolat M, Törő O, Bécsi T. Performance evaluation of a maneuver classification 44. Hortúa HJ, Volpi R, Marinelli D, Malagò L. Parameter estimation for the cosmic
algorithm using different motion models in a multi-model framework. Sensors (2022) microwave background with bayesian neural networks. Phys Rev D (2020) 102:103509.
22:347. doi:10.3390/s22010347 doi:10.1103/physrevd.102.103509
23. Fu Q, Lu K, Sun C. Deep learning aided state estimation for guarded semi-markov 45. Pignat E, Calinon S. Bayesian Gaussian mixture model for robotic policy
switching systems with soft constraints. IEEE Trans Signal Process (2023) 71:3100–16. imitation. IEEE Robotics Automation Lett (2019) 4:4452–8. doi:10.1109/lra.2019.
doi:10.1109/tsp.2023.3274937 2932610
24. Ali W, Li Y, Raja MAZ, Khan WU, He Y. State estimation of an underwater 46. Yuan K, Chatzinikolaidis I, Li Z. Bayesian optimization for whole-body control of
Markov chain maneuvering target using intelligent computing. Entropy (2021) 23:1124. high-degree-of-freedom robots through reduction of dimensionality. IEEE Robotics
doi:10.3390/e23091124 Automation Lett (2019) 4:2268–75. doi:10.1109/lra.2019.2901308

IBRBNC

Uploaded by

Copyright:

Available Formats

IBRBNC

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

IBRBNC

Uploaded by

Copyright:

Available Formats

TYPE Original Research

PUBLISHED 04 June 2024

Intelligent Bayesian regularization

Frontiers in Physics 01 frontiersin.org

1 Introduction step. The aforementioned capacity enables them to effectively

Frontiers in Physics 02 frontiersin.org

Frontiers in Physics 03 frontiersin.org

Frontiers in Physics 04 frontiersin.org

calculation for spread of the dynamic shifting matrix Xτ with

while the sampling interval ∂τ is given in Eq. 6 as:

To ensure precise estimation of state features through the IBRBNC

Frontiers in Physics 05 frontiersin.org

Frontiers in Physics 06 frontiersin.org

In Eq. 28, ES represents the average sum of squares of the model

Frontiers in Physics 07 frontiersin.org

Frontiers in Physics 08 frontiersin.org

Frontiers in Physics 09 frontiersin.org

4 Discussion on simulation results

Frontiers in Physics 10 frontiersin.org

TABLE 1 Establishing various parameters of state features estimation yτ − Φiy

Frontiers in Physics 11 frontiersin.org

commendable overall performance in comparison to approximated output HEstimates

Frontiers in Physics 12 frontiersin.org

Frontiers in Physics 13 frontiersin.org

Frontiers in Physics 14 frontiersin.org

The numerical values of average MSEs are also computed to xτ

Frontiers in Physics 15 frontiersin.org

Frontiers in Physics 16 frontiersin.org

Frontiers in Physics 17 frontiersin.org

Frontiers in Physics 18 frontiersin.org

Frontiers in Physics 19 frontiersin.org

Frontiers in Physics 20 frontiersin.org

Frontiers in Physics 21 frontiersin.org

Frontiers in Physics 22 frontiersin.org

Frontiers in Physics 23 frontiersin.org

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.