1. Introduction
Inertial measurement units provide continuous and accurate estimates of motion states in between sensor measurements. Future unmanned naval vessels depicted in
Figure 1a require very accurate motion measurement units including active sensor systems and inertial algorithms when active sensor data is unavailable. State observers are duals of state controllers used for establishing decision criteria to declare accurate positions and rates and several instantiations are studied here when fused with noisy sensors, where theoretical analysis of the variance resulting from noise power is presented and validated in over ten-thousand Monte Carlo simulations.
The combination of physical sensors and computational models to provide additional information about system states, inputs, and/or parameters, is known as virtual sensoring. Virtual sensoring is becoming more and more popular in many sectors, such as the automotive, aeronautics, aerospatial, railway, machinery, robotics, and human biomechanics sectors. Challenges include the selection of the fusion algorithm and its parameters, the coupling or independence between the fusion algorithm and the multibody formulation, magnitudes to be estimated, the stability and accuracy of the adopted solution, optimization of the computational cost, real-time issues, and implementation on embedded hardware [
1].
The proposed methods stem from Pontryagin’s treatment of Hamiltonian systems, rather than utilization of classical or modern optimal estimation and control concepts applied to future naval vessels as depicted in (
Figure 1) [
2,
3,
4].
Typical motion reference units conveniently have accuracies on the order of 0.05 (in meters and degree for translation and rotation, respectively, as depicted in
Figure 1b for representative naval vessels as depicted in
Figure 1a). These figures of merit are aspirational for the virtual sensor that must provide accurate estimates whether active measurements are available to augment the algorithm. Lacking active measurements, the algorithm is merely an inertial navigation unit, while with active measurements, the algorithm becomes an augmented virtual sensor. This manuscript investigates virtual sensoring by evaluating several options for algorithms, resulting estimated magnitudes, accuracy of each solution, optimization of resulting costs of motion, and sensitivity to variations like noise and parameter uncertainty of the translational and rotational motion models investigated (both simplified and high-fidelity). Algorithms are compared using various decision criteria to compare approaches for consideration of usage as motion reference units potentially aided by global navigation systems.
Noting the small size of motion measurement units, simple algorithms are preferred to minimize computational burdens that can increase unit size. Motion estimation and control algorithms to be augmented by sensor measurements are based on well-known mathematical models of translation and rotation from physics, both presented in equations. In 1834, the Royal Society of London published two celebrated papers by William R. Hamilton on Dynamics in the Philosophical Transactions. Ref. [
5] The notions were slowly adopted, and not presented relative to other thoughts of the age for nearly seventy years [
6], but quickly afterwards, the now-accepted axioms of translational and rotational motion were self-evidently accepted by the turn of the twentieth century [
7,
8,
9,
10] as ubiquitous concepts. Half a century later [
11,
12], standard university textbooks elaborate on the notions to the broad scientific community. Unfortunately, the notions arose in an environment already replete with notions of motion estimation and control based on classical proportional, rate, and integral feedback, so the fuller utilization of the first principals languished until exploitation by Russian mathematician Pontryagin [
13]. Pontryagin proposed to utilize the first principles as the basis for treating motion estimation and control as the classical mathematical feedback notions were solidifying in the scientific community. Decades later, the first-principal utilization proposed by Pontryagin are currently rising in prominence as an improvement to classical methods [
14]. After establishing performance benchmarks [
15] for motion estimation and control of unmanned underwater vehicles, the burgeoning field of deterministic artificial intelligence [
16,
17] articulates the assertion of the first-principles as “self-awareness statements” with adaption [
18,
19] or optimal learning [
20] used to achieved motion estimation and control commands. The key difference between the usage of first principals presented here follows. Classical methods impose the form of the estimation and control (typically negative feedback with gains) and they have very recently been applied to railway vehicles [
21], biomechanical applications [
22], and remotely operated undersea vehicles [
23], electrical vehicles [
24], and even residential heating energy consumption [
25] and multiple access channel usage by wireless sensor networks [
26]. Deterministic artificial intelligence uses first principals and optimization for all quantities but asserts a desired trajectory. Meanwhile the proposed methods in this manuscript leave the trajectory “free” and calculate an optimal state and rate trajectory for fusion with sensor data and calculates optimal decision criteria for estimation and controls in the same formulation.
This manuscript seeks to use the same notion, assertion of the first principals (via Pontryagin’s formulation of Hamiltonian systems) in the context of inertial motion estimation fused with sensor measurements (that are presumed to be noisy). Noise in sensors is a serious issue elaborated by Oliveiera et al. [
27] for background noise of acoustic sensors and by Zhang et al. [
28] for accuracy of pulse ranging measurement in underwater multi-path environments. Barker et al. [
29] evaluated impacts on doppler radar measurements beneath moving ice. Thomas et al. [
30] proposes a unified guidance and control framework for Autonomous Underwater Vehicles (AUVs) based on the task priority control approach, incorporating various behaviors such as path following, terrain following, obstacle avoidance, as well as homing and docking to stationary and moving stations. Zhao et al. [
31] very recently pursued the presently ubiquitous pursuit of optimality via stochastic artificial intelligence using particle swarm optimization genetic algorithm, while Anderlini et al. [
32] used real-time reinforcement learning. Sensing the ocean environment parallels the current emphasis in motion sensing, e.g., Davidson et al.’s [
33] parametric resonance technique for wave sensing and Sirigu et al.’s [
34] wave optimization via the stochastic genetic algorithm. Motion control similarly mimics the efforts of motion sensing and ocean environment sensing, e.g., Veremey’s [
35] marine vessel tracking control, Volkova et al.’s [
36] trajectory prediction using neural networks, and the new guidance algorithm for surface ship path following proposed by Zhang et al. [
37]. Virtual sensory will be utilized in this manuscript where noisy state and rate sensors are combined to provide smooth, non-noisy, accurate estimates of state, rate, and acceleration, while no acceleration sensors were utilized. A quadratic cost was formulated for acceleration, since accelerations are directly tied to forces and torques and therefore fuels.
“…condition of the physical world can either be ‘‘directly’’ observed (by a physical sensor) or indirectly derived by fusing data from one or more physical sensors, i.e., applying virtual sensors”.
Thus, the broad context of the field is deeply immersed in a provenance of classical feedback driving a current emphasis on optimization by stochastic methods. Meanwhile this study will iterate options utilizing analytic optimization including evaluation of the impacts of variations and random noise in establishing the efficacy of each proposed approach. Analytical predictions are made of the impacts of applied noise power, and Monte Carlo analysis agrees with the analytical predictions. Developments presented in this manuscript follow the comparative prescription presented in [
39], comparing many (eleven) optional approaches permitting the reader to discern their own preferred approach to fusion of sensor data with inertial motion estimation:
Validation of simple yet optimal inertial motion algorithms for both translation and rotation derived from Pontryagin’s treatment of Hamiltonian systems when fused with sensor data that is assumed to be noisy.
Validation of high-fidelity optimal (nonlinear, coupled) internal motion algorithms for rotation with translation asserted by logical extension derived from Pontryagin’s treatment of Hamiltonian systems when fused with sensor data that is assumed to be noisy;
Validation of three approaches for sensor data fused with the proposed motion estimation algorithm (not using classical feedback in a typical control topology): pinv, backslash, and LU inverses derived from Pontryagin’s treatment of Hamiltonian systems when fused with sensor data that is assumed to be noisy;
Comparison of each proposed fused implementation algorithm to three varieties of classical feedback motion architectures including linear-quadratic optimal tracking regulators, classical proportional plus velocity feedback tuned for performance specification and manually tuned proportional plus integral plus derivative feedback topologies, where these classical methods are utilized as benchmarks for performance comparisons when fused with sensor data that is assumed to be noisy.
Comparisons are made based on motion state and velocity errors, algorithm parameter estimation errors, and quadratic cost functions, which map to fuel used to create translational and rotational motion.
Vulnerability to variation is evaluated using ten-thousand Monte Carlo simulations varying state and rate sensor noise power and algorithm plant model variations, where noise power is tailored to the simulation discretization, permitting analytic prediction of the impacts of variations to be compared to the simulations provided.
Sinusoidal wave action is programmed in the same simulation code to permit future research, and inclusion of such is indicated throughout the manuscript.
Appendix A,
Table A1 contains a consolidated list of variables and acronyms in the manuscript.
2. Materials and Methods
Inertial navigation algorithms use physics-based mathematics to make predictions of motion states (position, rate, acceleration, and sometimes jerk). The approach taken here is to utilize the mathematical relationships from physics in a feedforward sense to produce optimal, nonlinear estimates of states that when compared to noisy sensor measurements yield corrected real-time optimal, smooth, and accurate estimates of state, rate, and acceleration. Sensors are modeled as ideal with added Gaussian noise and the smooth estimates will be seen to exhibit none of the noise. The optimization of the estimates will be derived using Pontryagin’s optimization.
Motion control algorithms to be augmented by sensor measurements are based on well-known mathematical models of translation and rotation from physics, both presented in Equation (1), where both high-fidelity motion models are often simplified to identical double-integrator models where nonlinear coupling cross-products of motion are simplified, linearized, or omitted by assumption. The topologies are provided in
Figure 2. Centrifugal acceleration is represented in Equation (1) by
$-m\omega \times \left(\omega \times {r}^{\prime}\right)$. Coriolis acceleration is represented in Equation (1) by
$-2m\omega \times {v}^{\prime}$. Euler acceleration is represented in Equation (1) by
$m\dot{\omega}\times {r}^{\prime}$. In this section, double-integrator models are optimized by Pontryagin’s treatment of Hamiltonian systems, where the complete (not simplified, linearized, or omitted) nonlinear cross-products of motion are accounted for using feedback decoupling. Efficacy of feedback decoupling of the full equations of motion is validated by disengaging this feature in a single simulation run to reveal the deleterious effects of the coupled motion when not counteracted by the decoupling approach.
where
$F$, $\tau $external force and torque, respectively
$m$, $I$mass and mass moment of inertia, respectively
$\omega $,$\dot{\omega}$angular velocity and acceleration, respectively
${r}^{\prime}$,${v}^{\prime},\mathrm{a}\prime $position and velocity, and acceleration relative to rotating reference
$\tau =I\dot{\omega}$and $F=m\mathrm{a}\prime $ are double integrator plants
$\omega \times I\omega $cross-product rotational motion due to rotating reference frame
$m\dot{\omega}\times {r}^{\prime}$cross-product translation motion due to rotating reference frame
$-2m\omega \times {v}^{\prime}$cross-product translation motion due to rotating reference frame
$-m\omega \times \left(\omega \times {r}^{\prime}\right)$cross-product translation motion due to rotating reference frame.
2.1. Problem Scaling and Balancing
Consider problems whose solution must simultaneously perform mathematical operations on very large numbers and very small numbers. Such problems are referred to as poorly conditioned. Scaling and balancing problems are one potential mitigation where equations may be transformed to operate with similarly ordered numbers by scaling the variables to nominally reside between zero and unity. Scaling problems by common, well-known values permits single developments to be broadly applied to a wide range of state spaces not initially intended. Consider problems simultaneously involving very large and very small values of time ($\overline{t}$), mass ($\overline{m}$)/mass moments of inertia ($\overline{I}$), and/or length ($\overline{r}$). Normalizing by a known value permit variable transformation such that newly defined variables are of similar order, e.g., $t\equiv \frac{\overline{t}}{{t}_{f}},I\equiv \frac{\overline{I}}{{I}_{system}}=J\equiv \frac{\overline{J}}{{J}_{system}},m\equiv \frac{\overline{m}}{{m}_{system}},r\equiv \frac{\overline{r}}{{r}^{\prime}}$ where $r$ indicates generic displacement units like $x,y,orangle$. Such scaling permits problem solution with a transformed variable mass and inertia of unity value, initial time of zero and final time of unity, and state and rate variables that range from zero to unity making the developments here broadly applicable to any system of particular parameterization.
2.2. Scaled Problem Formulation
The problem is formulated in terms of standard form described in Equations (2)–(8), where
$x(\xb7),v(\xb7)$ are the decision variables. The endpoint cost
$E\left(x\left({t}_{f}\right)\right)$ is also referred to as the Mayer cost. The running cost
$F\left(x\left(t\right),u\left(t\right)\right)$ is also referred to as the Lagrange cost (usually with the integral). The standard cost function
$J\left[x(\xb7),u(\xb7)\right]$ is also referred to as the Bolza cost as the sum of the Mayer cost and Lagrange cost. Endpoint constraints
$e\left(x\left({t}_{f}\right)\right)$ are equations that are selected to be zero when the endpoint is unity.
where
$J\left[x(\xb7),u(\xb7)\right]$cost function
${x}^{T}=\left[x,v\right]$state vector of motion state $x$ and rate $v$ with initial condition
${x}^{T}=\left[x,v\right]$$\left({x}_{0},{v}_{0}\right)$ and final conditions $\left({x}_{f-1},{v}_{f},{t}_{f-1}\right)=\left(0,0,0\right)$
$u=\left[u\right]$decision vector
$H$Hamiltonian operator corresponding to system total energy
${\lambda}^{T}$adjoint operators, also called co-states (corresponding to each state)
${\upsilon}^{T}$endpoint costates
$e\left(x\left({t}_{f}\right)\right)$endpoint constraints.
2.3. Hamiltonian System: Minimization
The Hamiltonian in Equation (8) is a function of the state, co-state, and decision criteria (or control) and allows linkage of the running costs
$F\left(x,u\right)$ with a linear measure of the behavior of the system dynamics
$f\left(x,u\right)$. Equation (9) articulates the Hamiltonian of the problem formulation described in Equations (2)–(5). Minimizing the Hamiltonian with respect to the decision criteria vector per Equation (10) leads to conditions that must be true if the cost function is minimized while simultaneously satisfying the constraining dynamics. Equation (11) reveals the optimal decision
$u$ will be known if the rate adjoint can be discerned.
2.4. Hamiltonian System: Adjoint Gradient Equations
The change of the Hamiltonian with respect to the adjoint
$\lambda $ maps to the time-evolution of the corresponding state in accordance with Equations (12) and (13).
The rate adjoint was discovered to reveal the optimal decision criteria, and the adjoint equations reveal the rate adjoint is time-parameterized with two unknown constants still to be sought. Together, Equations (11)–(13) form a system of differential equations to be solved with boundary conditions (often referred to as a two-point boundary value problem in mathematics).
2.5. Terminal Transversality of the Enpoint Lagrangian
The endpoint Lagrangian
$\overline{E}$ in Equation (14) adjoins the endpoint function endpoint cost
$E\left(x\left({t}_{f}\right)\right)$ and the endpoint constraints functions
$e\left(x\left({t}_{f}\right)\right)$ in Equation (8) and provides a linear measure for endpoint conditions in Equation (7). The endpoint Lagrangian
$\overline{E}$ exists at the terminal (final) time alone. The transversality condition in Equation (15) specifies the adjoint at the final time is perpendicular to the cost at the end point. In this problem, the endpoint cost
$E\left(x\left({t}_{f}\right)\right)=0$. These Equations (16) and (17) are often useful when seeking a sufficient number of equations to solve the system.
2.6. New Two-Point Boundary Value Problem
For the two-state system, four equations are required with four known conditions to evaluate the equations. In this instance, two Equations (3)–(10) have been formulated for state dynamics, two more Equations (18) and (19) have been formulated for the adjoints, and two more Equations (20) and (21) have been formulated for the adjoint endpoint conditions. Four known conditions, Equations (22)–(25) have also been formulated. Combining Equations (11) and (13) produce Equation (26).
Evaluating Equation (27) with Equation (23) produces the value
$c=0$. Evaluating Equation (28) with Equation (22) produces the value
$d=0$. Evaluating Equation (27) with Equation (25) produces Equation (29), while evaluating Equation (28) with Equation (24) produces Equation (30).
Solving the system of two Equations (29) and (30) produces
$a=-12$ and
$b=6$. Substituting Equation (26) into Equation (11) with
$a$ and
$b$ produces Equation (31), and substitution of
$a$ and
$b$ into Equations (27) and (28), respectively, produce Equations (32) and (33) the solution of the trajectory optimization problem.
Equations (31)–(33) constitute the optimal solution for quiescent initial conditions and the state final conditions (zero velocity and unity scaled position). To implement a form of feedback (not classical feedback), consider leaving the initial conditions non-specific in variable-form as described next.
2.7. Real-Time Feedback Update of Boundary Value Problem Optimum Solutions
Classical methods utilize feedback of asserted form
$u=-Kx$ for state variable
$x$, where the decision criteria (for control or state estimation/observer) and gains
$K$ are solved to achieve some stated performance criteria. Such methods are used in
Section 3 and their results are established as benchmarks for comparison. So-called modern methods utilize optimization problem formulation to eliminate classical gain tuning substituting optimal gain selection but retaining the asserted form of the decision criteria. Such methods are often referred to as “linear-quadratic optimal” estimators or controllers. These estimators are also presented as benchmarks for comparison, where the optimization problem equally weights state errors and estimation accuracy.
Alternative use of feedback is proposed here (whose simulation is depicted in
Figure 3b). Rather than classical feedback topologies asserting
$u=-Kx$ utilization of state feedback in formulating the estimator or control’s decision criteria, this section proposes re-labeling the current state feedback as the new initial conditions of the two-point boundary-value problems used to solve for optimal state estimates or control decision criteria in Equations (22) and (23). The solution of (26)–(28) using the initial values of (22) and (23) manifest in values of the integration constants:
$a=-12$ and
$b=6$. As done in real-time optimal control, the values of the integration constants are left “free” in variable form, and their values are newly established for each discrete instance of state feedback (re-labeled as new initial conditions). This notion is proposed in Proposition 1, whose proof expresses the form of the online calculated integration constants that solve the new optimization problem. The two constants
$\widehat{a}$ and
$\widehat{b}$ are utilized in the same decision Equation (31) where the estimates replace the formerly solved values of the boundary value problem resulting in Equation (40).
Proposition 1. Feedback may be utilized not in closed form to solve the constrained optimization problem in real time.
Proof of Proposition 1. Implementing Equations (34)–(37) in matrix form as revealed in Equation (38) permits solution for the unknown constants as a function of time as displayed in Equation (39), and subsequent use of the unknown constants form the new optimal solution from the current position and velocity per Equation (40).
In
Section 3, estimation of
$\widehat{a}$ and
$\widehat{b}$ becomes singular due to the inversion in Equation (39) as approaching the terminal endpoint, where switching to Equations (31)–(33) is implemented as depicted in
Figure 4d to avoid the deleterious effects of singularity when applying Proposition 1. The cases with switching at singular conditions are suffixes with “
with switching” in the respective label.
2.8. Feedback Decoupling of Nonlinear, Coupled Motion Due to Cross Products
The real-time feedback update of boundary value problem optimum solutions is often used in the field of real-time optimal control, but a key unaddressed complication remains the nonlinear, coupling cross-products of motion due to rotating reference frames. Here, a feedback decoupling scheme is introduced, allowing the full nonlinear problem to be addressed by the identical scaled problem solution presented, and such is done without simplification, linearization, or reduction by assumption. In proposition 2, feedback decoupling is proposed to augment the optimal solution already derived. The resulting modified decision criteria in Equation (42) is utilized in simulations presented in
Section 3 of this manuscript, but a single case omitting Proposition 2 is presented to highlight the efficacy of the approach.
Proposition 2. The real-time optimal guidance estimation and/or control solution may be extended from the double-integrator to the nonlinear, coupled kinetics by feedback decoupling as implemented in Equation (41).
Proof of Proposition 2. For nonlinear dynamics of translation or rotation as defined in Equation (1), where the double-integrator is augmented by cross-coupled motion due to rotating reference frames, the same augmentation may be added to the decision criteria in Equation (40) using feedback of the current motion states in accordance with Equation (42). The claim is numerically validated with simulations of “cross-product decoupling” that are nearly indistinguishable from open loop optimal solution, and a single case “without cross-product decoupling” is provided for comparison.
2.9. Analytical Prediction of Impacts of Variations
Assuming Euler discretization (used in the validating simulations) for output
y, index
i and integration solver timestep
h Equation (43) would seem to indicate a linear noise output relationship. Equation (44) indicates the relationship for quiescent initial conditions indicating the results of a style draw. In a Monte Carlo sense (to be simulated) of a very large number
n, Equation (45) indicates expectations from theory Equation (46) in simulation for scaled noise entry to the simulation to correctly reflect the noise power of the noisy sensors in the discretized computer simulation. Equation (46) was used to properly enter the sensor noise in the simulation (
Figure 2a and
Figure 3a).
Assuming this implementation of noise power for a given Euler (ode1) discretization in SIMULINK, 1 −
$\sigma $ error ellipse may be calculated as Equation (47) for the system in canonical form in accordance with [
40] and was implemented in
Figure 3a and depicted on “scatter plots” in
Section 3’s presentation of results of over ten-thousand Monte Carlo simulations.
2.10. Numerical Simulation in MATLAB/SIMULINK
Validating simulations were performed in MATLAB/SIMULINK Release R2021a with Euler integration solver (ode1) and a fixed time step of 0.01 s, whose results are presented in
Section 3, while this subsection displays the SIMULINK models permitting the reader to duplicate the results presented here. Sensor noise was added per
Section 2.8. The classical feedback subsystem is displayed in
Figure 4a. The optimal open loop subsystem implements Equation (31), and is elaborated in
Figure 4b,c. The real time optimal subsystem implements Equations (42) and (31) augmented by feedback decoupling as in Equation (42). The “switch to open loop” subsystem switches when the matrix inverted in Equation (39) is singular indicated by a zero valued determinant and is elaborated in
Figure 4d. The quadratic cost calculation computes Equation (3) and is elaborated in
Figure 4b, while the cross-product motion feedback implements the cross product of Equation (42). The P + V subsystem and PD/PI/PID subsystems depicted in
Figure 4a implement classical methods not re-derived here, but whose computer code is presented in
Appendix B, Algorithms A1 and A2.
Figure 5 displays the SIMULINK subsystems used to implement the three instantiations of real-time optimization (labeled RTOC from provenance in optimal control) where the switching displayed in
Figure 3b permits identical simulation experiments to be performed with all conditions fixed, varying only the proposed implementation. The subsystems execute Equation (39) with three variations of matrix inversion: (1) MATLAB’s backslash “\”, (2) Moore-Penrose pseudoinverse (
pinv), (3)
LU-inverse.
Section 2.10 presented SIMULINK subsystems used to implement the equations derived in the section.
Table 1 displays the software configuration used to simulate the equations leading to the results presented immediately afterwards in
Section 3.