1. Introduction
The design of input signals is undoubtedly one of the key elements for the success of system identification [
1]. Well-designed signals can stimulate and highlight the main dynamics of the system, allowing these behaviors to be accurately estimated and represented by parametric models [
2]. Consequently, they constitute a fundamental element for obtaining accurate models and, in turn, for tuning model-based controllers [
3], designing state observers [
4], and developing simulators [
5], among other applications [
6].
However, the task of Optimal Signal Design (OSD) is not a trivial problem. As presented in the literature, especially in [
7,
8,
9], this task is defined as a mixed-integer nonlinear optimization problem. It requires competent approaches to simultaneously address signal design and parametric estimation problems.
Regardless of the methodology adopted, there is a consensus that the approach must generate signals that provide effective stimuli to the main dynamics of the system [
10]. These signals must exhibit appropriate excitation persistence, covering a wide range of amplitudes and frequencies. Furthermore, it is essential that the signals be feasible for the actuator system and adhere to safety conditions during the identification experiment development [
11].
Meeting all these requirements is challenging, especially in nonlinear systems, where signals with high persistence and detailed encoding are necessary. In such cases, there can be a significant increase in computational complexity, leading to time-consuming computational solutions [
12].
To address the Optimal Signal Design (OSD) problem, it is common to formulate the optimization problem using the inverse of the Fisher Information Matrix (FIM). This approach allows for measuring signal excitation persistence, irrespective of the type of estimation algorithm implemented.
Regarding the solution to the optimization problem, various deterministic and probabilistic optimization algorithms are employed [
13]. These include Steepest-descent, Conjugate-gradient, Quasi-linearization, and Newton–Raphson, among others. Additionally, bio-inspired meta-heuristics are classic examples of probabilistic methods [
14].
A comprehensive survey of the literature reveals several algorithms in this context. Concerning solutions based on deterministic methods, works such as [
15,
16] serve as classic examples of applications in the field of OSD. In these approaches, signals are optimized using deterministic algorithms that target metrics based on the Fisher Information Matrix (FIM), including criteria like A-Optimality, E-Optimality, and D-Optimality
Another approach found in the literature combines deterministic methods with optimal control theory. Examples of these combined approaches can be found in [
17,
18,
19,
20].
Specifically, in the work by [
17], an identification methodology is applied to aerial vehicles. The optimal signals are generated through a combination of dynamic programming and optimal control, which enhances traditional identification maneuvers. It is worth noting the significant reliance on the optimal control input signal. While these approaches yield satisfactory results, it is evident that further improvements to the entry maneuvers could lead to a substantial increase in the problem’s spatial complexity, rendering it intractable for the techniques presented.
In terms of methodologies, it is worth highlighting the works of [
7,
12,
21], which focus on designing multilevel pseudo-random sequence signals. The work in [
7] employed the PSO algorithm, ref. [
12] utilized the Ant Colony Optimization (ACO) algorithm, and [
21] opted for the Genetic Algorithm (GA) to address the OSD problem. What sets these works apart is their ability to design signals with advanced coding, reasonable processing times, and high-quality solutions. In the case of [
7], the signal design also optimized the experiment time, a feature not found in other works. Additionally, the authors developed a dual-layer optimization methodology, using their own metrics that cater to specific needs.
Considering all the mentioned works, it becomes apparent that Optimal Signal Design can be highly complex, particularly when dealing with signals featuring multiple levels of amplitude or intricate coding. In such cases, the number of potential combinations in the solution space can become prohibitively large, potentially rendering the project infeasible or leading to extremely high computational processing times, as exemplified in [
12].
In this context, the application of meta-heuristics has shown promise, as demonstrated in the works of [
7,
12,
21]. However, a more comprehensive investigation, involving a comparison of various techniques, is still needed. It is worth noting that this area of research is not yet extensive in the literature, but it holds promise for OSD and deserves further attention.
In light of this context, this paper presents a comparative study of meta-heuristics applied to the OSD problem for estimating parameters of Nonlinear Dynamic Systems (NDSs). Several meta-heuristics will be investigated using the methodology proposed by [
22], referred to as rSOESGOPE. While the PSO algorithm [
23] served as the optimizer, this paper will also explore other meta-heuristics to enhance the performance of rSOESGOPE. The traditional algorithms such as GA [
24], Bat Algorithm (BA) [
25], Whale Optimization Algorithm (WOA) [
26], Grey Wolf Optimizer (GWO) [
27], Salp Swarm Algorithm (SSA) [
28], Arithmetic Optimization Algorithm (AOA) [
29], and Multi-trial vector-based Differential Evolution (MTDE) [
30] will be analyzed.
This comparative study aims to explore and assess the potential of these approaches for solving this class of optimization problems. To this end, real case studies involving an ASV with three DoFs and an aerial holonomic propulsion system are presented and discussed.
After this summary, the contributions of this paper can be highlighted as follows:
a performance comparison of eight meta-heuristics applied in the rSOESGOPE concept for parametric estimation, not yet explored in the literature;
a description of the rSOESGOPE concept that allows obtaining robust models from the use of multiple projection excitation signals;
a real case application involving a 3 DoF ASV, which presents a problem with a complex solution space for signal generation and a challenging parametric estimate due to the hydrodynamic phenomena involved;
an in-depth analysis of the ASV’s identification process, providing valuable insights into the modeling and generation of identification signals.
This paper is organized as follows:
Section 2 presents the OSD problem statement, demonstrating its characteristics and complexities;
Section 3 introduces the rSOESGOPE method;
Section 4 elaborates on the fundamental aspects for the application of the algorithms in the rSOESGOPE concept;
Section 5 presents the results of applying the meta-heuristics to a real case study problem;
Section 7 presents some concluding remarks.
2. Problem Formulation
Consider a nonlinear dynamic system
which can be satisfactorily approximated by the nonlinear model
:
where
and
are the system-dependent nonlinear functions on the state vector
, the output vector
, the input vector
, and the set of parameters of the model
. Additionally, it is admitted that the
has some restrictions that need to be respected during its operation, represented by
where
and
represent, respectively, the upper and lower vector states’ operating limits,
and
express the lower and upper vessels’ output vector operating limits, respectively, while
and
represent the upper and lower input signal operating limits, respectively.
It is also assumed that represents the unknown optimal set of the system’s parameters, and best represents for any given signal , where represents the entire possible input signal domain.
In this context, designing an optimal identification signal involves finding the best signal
that, when applied to the real system
, enables the subsequent parametric estimation
, satisfying the following relationship:
. Therefore, designing an optimal identification signal requires discovering
through an optimization search defined by
where
is the initial parameter estimation
and
is an arbitrary mixed-integer optimization algorithm that searches for the best
. Finally, the optimal parameter set is defined by
where
is an arbitrary optimization procedure that uses the mathematical model
and the signal
to estimate the parameter set
in which
best represents
. Additionally, in this notation, it is possible to replace
by
or
without loss of generalization.
3. Robust SOESGOPE
The Robust SOESGOPE (rSOESGOPE) is a methodology developed for designing identification signals and promoting robust parametric estimation.
This approach is a derivation of the original SOESGOPE method developed by [
7] for designing single optimal signals for optimal parametric estimation of NDS.
However, as demonstrated by [
8], the SOESGOPE method does not yield satisfactory results in situations where the initial estimate has a high level of uncertainty regarding the real system.
To address these uncertainties, the rSOESGOPE method proposes the use of multiple identification signals designed from a set of well-spatially distributed benchmark parameters around the initial estimation . With this change, the aim is to minimize uncertainties about and enhance excitation persistence through multiple signals projected by different characteristics derived from the reference parameters. Consequently, this provides a more robust model for parametric uncertainties.
Mathematically, this new concept can be succinctly described by the following hypothesis:
“If the set is a rough approximation of , where represents a set of well-spatially distributed benchmark parameters around the initial estimation , and leads to a set of signals capable of exciting and estimating . Then, the same will also be able to excite and consequently obtain that respects .”
These concepts can be better understood in
Figure 1, which illustrates the trust region surrounding
, the designed signals, and the stages of parametric estimation.
Thinking in stages, the rSOESGOPE method proposes a two-step approach:
Generate the benchmark parameter sets and use it as a parameter to find each signal of .
Apply to the real system to find .
Thus, according to
Section 2, the following steps can be defined:
To tackle the two-step approach, a two-layer optimization strategy was developed to solve
. Two well-known optimization algorithms were employed: the PSO algorithm [
31] and the Interior-Points Algorithm (IPA) [
32].
The PSO algorithm was dedicated to the external optimization layer, specifically for the design of and, consequently, the parameterization of , which represents the swarm’s individuals.
In the proposed method, the projected signals are AutoRegressive exogenous inputs with an Amplitude-Modulated Pseudo-Random Binary Signal (APRBS). The internal layer is assigned to the IPA, which is responsible for the parametric estimation required for evaluating signal quality. This same IPA is used to solve and find .
To assess the signal quality generated by , an objective function is proposed, consisting of a weighted sum of three metrics.
To ensure the analysis of excitation persistence, the following procedure was employed:
To ensure compliance with the constraints,
was used to penalize a given signal
if it drives the system out of the desired operational restrictions. Mathematically, this can be expressed as follows:
where
and
represent the lower and upper bounds of the
i-th state, respectively, and
and
represent the lower and upper limits of the
i-th output, respectively.
Using the metrics described above, ref. [
22] proposed the following objective function to be minimized. This objective function is composed of the weighted sum of the three metrics presented:
where
are the parameters that encode the signal, and
,
, and
are constant weightings related to metrics and established according to the priority.
6. Qualitative Analysis of Algorithm Performance
Taking a closer look at the algorithmic results, it is worthwhile to compare the techniques that achieved success, referred to as Group 1 (GWO, WOA, PSO, SSA, BA, and MTDE) and Group 2 (AOA and GA), compared to techniques that did not perform well in the studied problem.
In the case of Group 2, the GA exhibited low performance. Despite its wide applicability in complex problems, it is known that the GA can face challenges related to parameterization and slow convergence, especially in complex and multidimensional problems. These challenges became evident in the final model of the algorithm, which was developed using only one optimal excitation signal. Given the complexity of the problem, this resulted in a model with limited overall learning and reduced performance.
Similarly, within Group 2, the AOA also failed to achieve satisfactory results. Premature convergence and difficulties in exploring the search space were observed. These issues can be attributed to the use of only one signal in constructing the model. Similar to the GA, this led to a model with limited overall learning. Contributing factors include the dependence on the exploration–exploitation trade-off of two parameters with fixed values, the absence of information sharing and guidance towards the best global solution, and the modification of only the best solution based on arithmetic operators during the exploration and exploitation stages of the algorithm.
Turning to the results from Group 1, it is evident that all meta-heuristics generated models using three or more sub-optimal signals. This indicates that these algorithms did not encounter issues with premature convergence or difficulties in exploring the solution space. Furthermore, they allowed for the development of models with increased learning through the addition of new signals.
Among the six techniques in Group 1, three stood out as highly efficient: GWO, and the meta-heuristics MTDE and PSO. These techniques yielded very similar results and demonstrated efficiency in exploring the search space and addressing the complexities of OSD problems. GWO achieved the best performance, which can be attributed to (i) its lower sensitivity to parameters, making it easier to configure and fine-tune compared to other algorithms, and (ii) its ability to explore a wide range of solutions within its search space before converging towards solutions closer to the global optimum.
It is important to emphasize that each algorithm possesses individual characteristics and may excel in different scenarios. In this study, GWO demonstrated superiority, even outperforming the originally used PSO in the rSOESGOPE method. Finally, for a better understanding,
Table 13 below provides a comprehensive comparison of the entire study conducted in this article.
7. Conclusions
This paper presented a comparison of metaheuristic performance when applied to Optimal Signal Design for estimating parameters of a nonlinear system. In particular, various metaheuristics were investigated and applied in the context of rSOESGOPE, which originally utilized the PSO algorithm as the primary optimizer. To assess the performance of these metaheuristics, a modeling problem of a catamaran-type ASV was employed. From the results, several key points can be emphasized:
ASV was chosen because it presents several difficulties in the scope of modeling and identification, which must deal with geometric and mechanical asymmetries, coupling between the DoFs, a high level of uncertainty about the initial estimates, and other aspects such as environmental disturbances.
Furthermore, these challenges also include the design of signals in large search spaces and the parametric estimation of many parameters.
The techniques were analyzed at each stage of the rSOESGOPE methodology, from configuration, signal design, and parameter estimation to the model validation obtained in different identification scenarios.
Considering the Optimal Signal Design phase, it has been proposed and demonstrated that conducting the study to define the characteristics of the APRBS signal plays a fundamental role in the projected signal quality. This study shows the importance of defining the trade-off between the persistence of the excitation and the problem solvability. Although long multi-stage APRBS signals provide richer excitation characteristics, the nonlinear solution space grows exponentially, which can make it difficult, costly, or even intractable.
In general, the performance of the estimated model improved with the insertion of signals, minimizing the uncertainties of the initial estimates, as initially proposed. Exceptions occurred only for the GA and AOA models, reaffirming the difficulties in the design stage regarding compliance with the established constraint restrictions and improvement of the persistence of excitation. There is a high similarity between the estimated models and the real system. In these scenarios, the GWO method presents the best RMSE among all the methods used, followed by the MTDE and PSO methods, respectively. It can also be observed that the AOA method could not find a solution that satisfies the system, as its results were about 5 times larger than the other ones. Furthermore, regarding the RMSE for each DoF, the GWO obtained the best results considering surge and sway motions and is surpassed in the yaw behavior only by the SSA method, which is still ranked only in fifth place among the metaheuristics’ results.
Finally, it is possible to appreciate the validity of the rSOESGOPE concept, which has proven its effectiveness with various algorithms. Regardless of the adopted metaheuristic, it was possible to mitigate the impact of initial uncertainties on the system and signal design, leading to the estimation of robust models. This achievement opens the door to the development of robust control strategies and fault-tolerant systems for the use of ASV in more challenging and hostile environments.
As part of future work and research, the aim is to conduct a more comprehensive assessment of the algorithms’ performance under various environmental conditions and in harsher scenarios (e.g., fast-flowing water, waves, and/or winds) to gain a more comprehensive understanding of their applicability and limitations.
Furthermore, a future project involves the creation of a hybrid metaheuristic model that combines the strengths of the top-performing algorithms. This could lead to a more robust and efficient solution, further enhancing the capabilities of our models and their adaptability to a broader range of challenges.