betaFeedback

WorksEngineering

Neural Network-based Control Systems with Application to Flight Control, From Classical Neural Control to Reinforcement Learning

Abstract

This book provides a comprehensive and systematic exploration of neural network-based intelligent control systems for aerospace applications, bridging foundational theory with practical implementation. It is structured into two primary parts: Part I covers model-based control design, detailing Feedback Error Learning (FEL), pseudocontrol strategies, and neural backstepping for strict-feedback systems, along with practical refinements such as learning rule modifications, composite learning, and discrete-time formulations, all grounded in rigorous Lyapunov stability analysis. Part II introduces model-free control, laying the groundwork for Reinforcement Learning and Approximate Dynamic Programming. Throughout, theoretical concepts are solidified through detailed case studies on quadrotor UAVs, fixed-wing aircraft (GTM), satellites, and hypersonic flight vehicles, complete with performance metrics, parameter sensitivity analyses, and discussions of actuator faults, input constraints, and wind disturbances, making the book an invaluable resource for students and practicing engineers seeking to design, simulate, and analyze intelligent flight control systems.

Springer Nature

Introduction

Intelligent control systems

Although the words Intelligence and Autonomy have been vastly employed interchangeably, they have an essential conceptual difference [1]. Different definitions have been given for both concepts in the literature [2, 3]. However, in a general view, intelligence may be defined as “a very general mental capability that involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience” [3]. On the other hand, “the ability to generate one’s own purposes without any instruction from outside” can be interpreted as the autonomy of a system [1].

Nevertheless, concerning these two general definitions, in some cases, distinguishing the high Level of Autonomy (LoA) from the low level of intelligence is not trivial at all. As systems become increasingly automated, they exhibit behaviors that resemble intelligence. This suggests a continuum between high levels of automation and simple, low-level intelligent systems. Within the control theory framework, generally, the final goal is to develop an autonomous system (rather than an intelligent system), which can adequately fulfill a set of predefined missions. However, like the common literature, one may interpret the high level of autonomy of some Unmanned Aerial Vehicles (UAVs) as intelligence. [4] Accordingly, despite the conceptual difference between these two words, in the current study (like existing literature,) we will use the words Intelligence and Autonomy interchangeably [4].

Various criteria have been provided in the literature to specify a system’s LoA [5], especially a UAV. Despite the lack of a unique categorization [6], a beneficial division has been given in (author?) [1, 7] for UAVs. Considering it, the highest LoA for a single UAV (level 4) indicates the self-accomplishment of an assigned tactical plan, where it is capable of on-board trajectory replanning, event-driven self-resource management, and compensating for most disturbances and faults in various flight conditions. This, in turn, requires different learning-based mechanisms in the whole control system, including the entire range from low-level control, such as attitude control, to high-level control, such as path planning. Such a scheme will be known as an Intelligent Control System (ICS) in this book.

Today’s academic landscape boasts a plethora of advanced Control Mechanisms (CMs) stemming from diverse methodologies like neural networks, fuzzy systems [8], decision trees [9, 10], reinforcement learning [11], different data-driven approaches [12, 13], evolutionary algorithms [14, 15], etc. Given such vast and varied CM explorations in scholarly literature, it is imperative to provide a comprehensive overview, tracing the journey of CMs from conceptual blueprints to their tangible implementations in today’s world. Leading the way, the authors of (author?) [16] offer insights into the multifaceted technical and operational aspects of CMs, exploring avenues like fuzzy logic, neural networks, evolutionary algorithms, collective intelligence tactics, and hybrid optimization methods.

However, an in-depth theoretical exposition was noticeably absent given the topic’s vast scope. To bridge this void, our current examination pivots to Neural Network (NN)-based control systems, a predominant theme in contemporary research. NNs have distinctly showcased their prowess in controller design processes and dynamic model identification. Their intrinsic ability (based on the universal approximation property) to estimate various non-linearities in dynamic frameworks sets them apart. Unlike traditional fuzzy control approaches, which heavily rely on expert knowledge and experiential data to construct their fuzzy rule base [17], NNs intuitively understand system or controller dynamics’ intricacies, even without prior knowledge. Their adaptability is commendable, harmoniously integrating with diverse adaptive learning methods. NNs are applicable in both direct and indirect control structures. In direct control, NNs are directly used to construct the control command, while in indirect control, they are used to estimate unknown functions within the system equations, e.g., model uncertainties, which are then used to design the control command (See § 1.2 for more details). Furthermore, thanks to their inherent parallel processing capabilities, neural architectures are especially suited for real-time implementations [18, 19].

The genesis of ICS in flying vehicles can be traced back to the 1990s when the world of flight control began embracing NNs as adaptive tools, responding dynamically to unexpected faults and unique flying conditions [20]. Though the fusion of NNs with flight systems began in the early ’90s, real-world applications employing dynamic NNs in practical flight controls only gained traction in 2001, stymied previously by technological and procedural barriers [21]. A monumental undertaking in this arena was the collaboration between NASA and Boeing, which spanned ten years, from 1999 to 2009, birthing the trailblazing Intelligent Flight Control System (IFCS) program [22, 23]. This endeavor was bifurcated into two pivotal phases. The first phase, focusing on indirect adaptive flight control, saw its inaugural flights using a modified F-15B model in 1999. These initial runs employed pre-configured NNs to estimate the aircraft’s stability and control derivatives. In a significant evolution, the mutable cell structure of NNs was incorporated into the control scheme, granting online modifications to these estimated derivatives. Flights harnessing this online identification mode took to the skies in 2003 [24]. Although closed-loop control simulations showed promise, genuine adaptive control was notably absent, as flight controls refrained from making full use of the identified models in actual flight [25]. In a shift of focus, the project’s second phase was dedicated to crafting a direct adaptive control strategy, intertwining a dynamic inversion element with an online NN [26]. This stage’s test flights began in 2006 and extended into 2008. These trials drew comparisons between performances with and without the dynamic NN augmentation, especially during structural challenges and control surface faults. Evaluation criteria ranged from pilot rating feedback to performance measurements. As underscored in (author?) [27, 28], NN-based improvements notably overall pitch performance amidst structural damages. Conversely, control surface faults produced varied results from slight enhancement in pitch rate to occasional pilot-induced roll oscillations. This latter stage also saw refinements such as alternating NN inputs to mitigate issues like excessive correlation and high gain in the base design, introducing a weight decrement factor (during NN weight updates) to avoid over-fitting, and adding singular dead zones in the adaptation laws for more fluid operations. The 2009 findings, mirroring these refinements, exhibited significant advancements over the project’s outset [28]. This ambitious journey concluded with the F-15B’s retirement in January 2009 [23].

The revelations from the IFCS initiative shed light on the potential concerns and challenges of implementing adaptive control mechanisms, particularly in real-world applications like human-crewed aircraft. Among these challenges were the complex intricacies of control design and unpredictable control behaviors in unexpected flight and fault conditions [29, 30]. In response to these findings, NASA embarked on a new journey in 2009 with the Integrated Resilient Aircraft Control (IRAC) project. A primary objective was to design adaptive control strategies that were both intuitive and effective while ensuring that these adaptive flight controls underwent thorough validation and verification for critical safety checks. A comprehensive discussion of this project is beyond the scope of this book. Nevertheless, inspired by the context at hand, our study seeks to deeply explore both direct and indirect NN-driven adaptive control methods, charting their evolution towards increased efficiency and reliability. Intriguingly, while direct and indirect adaptive control methodologies stem from different vantage points, they converge in their mathematical structures and dynamic NN update principles, echoing the spirit of NASA’s IFCS initiative. Hence, this book organizes ICSs into two primary categories: model-based and model-free NN control strategies. Moreover, our investigation extends to various model-free control mechanisms grounded in alternative machine learning paradigms, including Reinforcement Learning (RL). In particular, we will investigate using the introduced structures in designing IFCS for applications such as satellites, quadrotors, and fixed-wing aircraft.

Direct versus indirect adaptive control

The scholarly landscape unveils many control approaches wherein NNs are utilized to address online optimization hurdles in the control domain [31, 32] or to imitate conventional controller actions [33]. Our discussion will center on control methodologies where NNs are directly integrated into the design phase, serving as intelligent mediators that embed an added layer of intelligence into the feedback mechanism. This integration can manifest in several ways: Some research initiatives employ NNs to estimate model uncertainties, subsequently shaping the control command. This approach is termed indirect neural control. Training NNs can be executed in online mode using established learning techniques such as Feedback Error Learning (FEL) or be preset offline to provide a pre-trained dynamic system model. Within the realm of direct neural control, NNs are responsible for directly shaping the control command.

For instance, considering model-based control strategies where system dynamics are expressed as x˙=f(x)+g(x)u\dot{x}=f(x)+g(x)u, where f(x)f(x) and g(x)g(x) signify unknown nonlinear functions of the system states and control input, the control command uu can be designed in two primary manners. The first approach employs NNs to estimate the unknown functions ff and gg, which are subsequently integrated into the designed control command. Alternatively, the second approach may directly determine the control command by evaluating g−1(x˙d−f)g^{-1}(\dot{x}_d-f) via an NN (where xdx_d denotes the desired trajectory). In scholarly discussions, the first approach is typically labeled as indirect adaptive control, while the second is recognized as a direct adaptive control technique [34, 35]. Nevertheless, as will be elaborated in Chapter 2, both these model-based control methods bear remarkable similarities in mathematical formulation and in the updating mechanisms of NNs.

On the other hand, delineating model-independent adaptive control strategies brings about its own complexities, given the diverse direct and indirect adaptive control approaches detailed in prior research. A common model-free indirect ICS approach entails employing NNs to identify the complete system dynamics, creating an adaptive control strategy grounded in this newly elucidated model, which was identified online. Though derived from various design philosophies, direct model-free control predominantly revolves around two main paradigms in this book: adaptive optimal and RL-based control methodologies.

While the direct adaptive control is widely recognized for its simplicity and computational efficiency [36], its widespread application across various domains has also raised eyebrows, especially when applied to critical missions. In particular, aggressive learning, a common tactic in direct adaptive control, accelerates the reduction of dynamic inversion error by increasing the learning rate [37]. However, high-gain control due to aggressive learning in direct adaptive control is a problematic issue which can cause actuator saturation, excite unmodeled dynamics, and other well-known problems of high learning rates [38]. Moreover, system dynamics can witness significant alterations, especially in the case of failure. The absence of reliable and solid insights into the current system dynamics can result in incompetent control input, particularly when the control system merges a nominal controller with an adaptive NN-driven control command [39].

Model-based versus Model-free control

Now, we present a more detailed categorization of NN-based control systems, laying the groundwork for subsequent chapters.

Over three decades, NNs have increasingly influenced the control engineering community, particularly the aviation sector in its flight control mechanisms [40]. Broadly, these control systems can be bifurcated into two main types: those rooted in a predefined model (which is called model-based) and those that function independently of any pre-established model (which is called model-free).

Model-based approach

Over recent decades, the model-based neural control approach has blossomed, drawing upon a nominal model of the system during the control design phase. A standout learning approach in this domain is FEL, which has been integrated into numerous advanced intelligent control frameworks. This approach is an unsupervised learning approach (distinguished by the training of the NN to recognize specific patterns without relying on explicit output examples [41]) that uses various system data such as the tracking error, the prediction error of the model, the output of a baseline controller, or a combination of them to develop a Lyapunov-based control command in such a way that guarantee the tracking error and the estimation error of the NNs to remain bounded. Various interpretations of FEL-driven ICSs have surfaced from academic and research quarters. Within this schema, the NN might be utilized to estimate model uncertainty and external disturbance or direct its energies toward generating the control command itself. The former typically aligns with an indirect control strategy, whereas the latter is more indicative of a direct adaptive control approach. Various network structures, such as Radial Basis Function networks, multi-layer perceptrons, High-Order NNs (HONNs) [42], Extreme Learning Machine (ELM) [43], and Elman NNs, have been integrated within FEL-based methodologies. Moreover, when the system dynamic is non-affine, or the dimension of system states and control inputs is not equalizing, it needs to use a particular design method like a pseudo-control strategy or backstepping scheme. A comprehensive dissection of these topics with their multifaceted interpretations in the flight control system and a brief discussion on different aspects of the stability analysis in such a control framework is earmarked for Chapter 2.

However, it is worth noting that FEL-based control systems are not free from obstacles. One intrinsic drawback is the NN’s duty to estimate all model uncertainty under a single term, which might reduce training performance and impede optimal training, especially when signals Persistent Exciting (PE) isn’t sufficient. Furthermore, many FEL-base controls employ a foundational controller derived from the system’s nominal model with the NN acting as a complementary enhancer. Such an arrangement can sometimes yield amplified control commands, particularly in situations with intense structural failure, actuator faults, and dynamic changes. Therefore, we need additional supplement structures. For example, when the control direction of the system is unknown, Nussbaum functions can be used. It is also helpful to use common ideas in Fault-Tolerant Control (FTC) structures to face possible system failure. Moreover, in many cases, it is crucial to consider constraints on control commands, or system states/outputs. Therefore, using methods such as the modified tracking error or Lyapunov Barrier functions can be a way to consider system constraints. On the other hand, one of the primary assumptions in Chapter 2 was the measurability of all the system states, but in many applications, this assumption is not valid. So, in such a situation, a wise idea is to design the state estimators such as High Gain Observer (HGO) for use in output feedback control structures. Of course, in these mentioned schemes, at least the unobservable dynamics of the system must be stable. The next limiting assumption was that all the terms that the NN should estimate only depend on the states of the system, but in many cases, including disturbances imposed on the system that can depend on time, this assumption is not correct. Therefore, by using the robust adaptive approach through the combination of NN and disturbance observers, more general cases can also be considered. Another limitation is processing power. This is especially important in real-time implementation. The idea of using the minimal learning parameter method, which estimates an upper limit for the norm of uncertainties, will reduce the cost of calculations and actually create a trade-off between adaptive and robust aspects. These considerations and other design subtleties contribute to the refinements introduced to the foundational FEL structure, which will be elaborated upon in Chapter 3.

Model-free approach

Model-based methods, despite their advantages, also have important challenges. The first significant problem is that the unmodeled dynamics of the system are not necessarily time-invariant (for example, actuator dynamics). Furthermore, the variation in g(x)g(x) (x˙=f(x)+g(x)u\dot{x}=f(x)+g(x)u) due to different causes, such as exhaustion, is also the second significant challenge. These problems can overshadow the optimal response and even cause the system to diverge in some cases. In contrast, model-free strategies function without the benefit of initial insights into system dynamics during the control development phase. As a basic concept, a single NN would often design the entire control structure in this realm [44, 45], generally reliant on error back-propagation for training the NN. While such methods occasionally displayed satisfactory responses, despite significant external disturbances [46], a solid mathematical basis for closed-loop system stability often remained out of reach [47]. Moreover, the inherent tendencies of these training techniques to settle into local optima raised concerns about their suitability for critical missions.

Neural network-based system identification

Academic literature abounds with examples of indirect, model-free ICS designs. These typically feature a distinct identification stage during control formulation. Various types of neural structures, from nonlinear autoregressive with exogenous inputs network (NARX) [44], and recurrent Elman networks to Convolutional NNs, Wavelet NNs, ELMs [48], and even Fuzzy NNs [49] are employed in this phase to identify different unmodeled dynamics. Online modifications can be made to these identified models, adapting to changes in system dynamics. The subsequent control design is then informed by the insights drawn from the NN. Determining closed-loop stability in this tiered control design might pose challenges. However, the results often outperform FEL-based strategies, particularly in the wake of drastic dynamic changes. This variant of indirect IFCSs receives comprehensive coverage in Chapter 4.

Neuroadaptive optimal control and Direct adaptive control using Reinforcement learning

The contemporary landscape has seen a surge in model-free ICSs, gleaning wisdom from Approximate Dynamic Programming (ADP) and RL concept. Although these concepts hail from diverse academic origins, ranging from control theory to machine learning and data theory, their foundational operational philosophies are remarkably congruent. A predominant motif within many of these techniques is the actor-critic framework, wherein NNs frequently serve dual roles in both actor and critic function estimation. Here, the critic embodies the anticipated cost function (or value function or cost to go), while the actor determines the system’s control command. In other words, since Hamiltonian-Jacobi-Bellman’s (HJB) optimality equations (similar to the Hamilton-Jacobi-Isaacs equations) do not have an analytical solution, it is necessary to approximate a solution for these equations through ADP methods such as policy iteration and policy improvement. In an RL context, the control design transforms into a Markov Decision Process (MDP), with the value function representing a cumulative, discounted system (agent) reward based on current policies. The primary objective centers around policy optimization to maximize this value function, which may be done by different ADP-based (such as Q-learning, SARSA, and Monte Carlo) or policy gradient methods. Advancements in computational prowess have accelerated the rise of RL-based ICSs. New algorithms were introduced that are very suitable for control applications. These algorithms, which have continuous action and observation space, use deep neural networks in their structures. In the RL literature, these algorithms are called Deep RL. The first algorithms in this field are Deep Deterministic Policy Gradient (DDPG) and Trust Region Policy Optimization (TRPO), which were introduced around 2014. Although these algorithms have similarities, they have significant differences, especially in the way of updating the policy. The Soft Actor-Critic (SAC) algorithm is an improved version of the DDPG algorithm, which generally has a higher learning efficiency, a lower estimation bias, and a more suitable convergence rate. Also, the Proximal Policy Optimization (PPO) algorithm, which has a more convenient and straightforward policy update process, is considered as an improved version of the TRPO algorithm. Concluding this discussion, it’s crucial to note that RL agents are typically trained in a simulation environment and seamlessly transferred to real-world applications. A thorough examination of this design paradigm awaits in Chapters 5 and 6.

To round off, Figure 1.1 gives a brief glimpse at the most important topics discussed in this book, and Chapter 7 presents a synthesis of the discussions, paired with foresight for upcoming NN-based ICS innovations. It underscores the ongoing pursuit of an ICS that harmonizes reliability, effectiveness, and computational frugality, which can be utilized in serious tasks.

Roadmap and structure of the book, showing model-based and model-free neural-network control topics across Chapters 2–7

Figure 1.1. Roadmap and structure of the book.

Part I Model-based Control Design

Foundations of Model-based Intelligent Control

In this chapter, we present the foundation of model-based neural network-driven control systems. In this procedure, our main focus is on the use of neural networks as a compensator for uncertainties and disturbances, and this goal is achieved through proper updating of neural network weights. To this end, in the following subsection, we will first give a brief introduction to neural networks, their structures, and training methods.

On the other hand, the development of a theoretical framework to guarantee the stability of the closed-loop system including dynamic neural networks is crucial to render a reliable control method. In a general view, the control structure can include a conventional controller in the inner loop, which will be responsible for stabilizing the nominal system dynamics. In addition, a neural network can help the main controller as a compensator for the effects of unmodeled dynamics and disturbances. In the process of designing such a structure, we employ the idea of Lyapunov stability. For this purpose, in the proposed Lyapunov functions, there will be terms for the tracking error and the estimation error of NN weights so that at least the Bounded-Input Bounded-Output stability (BIBO) can be guaranteed in the presence of model uncertainties and external disturbances.

In this regard, in the first step, we introduce the fundamental theoretical basis for the most commonly used approach to incorporate NNs within the adaptive control design process, i.e. the Feedback Error Learning method, while the application of such an algorithm in flight control systems would be addressed in various example. According to the type of system dynamics, the design process may undergo changes and different methods may be used. In section 2.2, we deal with affine systems that the control input dimension and the state dimension are equal to each other. To this end, first the basic process of model-based neural network-driven controller design will be described in detail, and then some existing challenges (such as the drift phenomenon) and how to deal with them will be expressed. Composite and concurrent learning methods will be introduced as effective methods to improve the learning performance of NNs in the lack of persistent exciting signals. Besides, the introduced design will be extended to higher-order dynamic systems by incorporating filtered tracking errors into the design.

In section 2.3, we discuss about non-affine systems. Most of the intelligent controller design approaches for this class of systems are direct methods. The pseudocontrol strategy that uses the approximate inverse model of the system (it can be a very simple model of the system that has only modeled the input and output connection with the correct sign, which is called the control assignment structure-property) is one of the widely used methods in this field.

Section 2.4 talks about another group of affine systems, which are called strict feedback forms. In this category of systems, each state variable is considered as the input of the previous state. Hypersonic flight vehicles are generally modeled like this. Neural backstepping is one of the most used control strategies for this category of systems. The main challenge of the basic neural backstepping method is the derivation of the state variables at each stage of the control input calculation, which is called the explosion of terms. However, this problem can be solved with simple modifications. The first idea that will be introduced is the dynamic surface control method. In this method, derivation will be done through the first-order filter. Appropriate selection of the filter time constant is very important in obtaining the desired result. Difficulty in stability analysis is the main challenge of this method, so to solve this problem, the idea of command filter backstepping can be used. This idea is based on the introduction of a new error, which is called a compensated error. In addition to this, there are many other methods for approximating the derivative, including the First-Order Sliding Mode Differentiator (FOSMD) method. In this method, the main idea is to be able to write state equations whose time response is the derivative of the desired signal (desired state variable). The main advantage of this approach compared to the previous ones is that if we consider the derivative filter as a type of linear derivative, this type of derivative method can be considered as a non-linear derivative that if the signal is Lipschitz, the error of the derivative tends to zero. In the final part of this section, we will provide direct neural backstepping methods. Direct methods free us from a series of complications, but generally, these types of solutions do not easily pass the safety standards. In addition, the learning rate is more important than before and its inappropriate choice can lead to state divergence.

In section 2.5, we will delve into the discussion on the definitions and different aspects of stability. Stability can be categorized in terms of extent and intensity. In the concept of extend, the stability of the system (equilibrium point) is investigated from local to global aspect. Nevertheless, in the concept of intensity, error limitations and how it behaves towards the zero value are evaluated.

Finally, in the last section, which is called concluding remarks, we will discuss and review the most important content and points presented during this chapter.

Neural networks overview

Artificial NNs have become a powerful tool for modeling complex nonlinear systems and have found widespread use in control theory and aerospace applications.

As a class of machine learning algorithms, neural networks do not need explicit programming but rather rely on training from data, unlike traditional computing methods.

Understanding the basic terminology used in neural networks is essential for grasping their functioning and applications. Here, we introduce some fundamental terms:

  • Layers: NNs are organized into distinct layers of neurons (See Figure 2.1). Layers are grouped into input layers that receive data, hidden layers that extract features, and output layers that generate predictions.

  • Neurons: The building blocks of a network. Each neuron sums its input signals weighted by the connection strengths and applies an activation function to generate an output signal.

  • Activation functions: Nonlinear functions such as sigmoid, tanh, or ReLU (Rectified Linear Unit) are applied to transform a neuron’s summed input into an output signal. They introduce nonlinearity enabling complex modeling.

  • Weights: The strength of connections between neurons. Weights are adjustable parameters that represent the knowledge learned by the network.

Typical structure of an RBF neural network

Figure 2.1. Typical structure of an RBF neural network.

These core elements enable neural networks to learn complex relationships between high-dimensional inputs and outputs for problems in prediction, classification, and control.

NNs can be categorized based on their structure into single-layer vs. multi-layer, and feedforward vs. feedback networks. Each architecture has its modeling capabilities and is suited to different control tasks.

This section provides a comprehensive overview of neural network fundamentals, architectures, and training methods in each category, with a focus on their application in control theory.

Neural network taxonomy

Neural networks can be broadly classified into two categories: feedforward and recurrent. Feedforward networks allow signals to travel in only one direction from input to output. A recurrent networks contains feedback connections, allowing signals to propagate in both directions. This section delves into the mathematical models and common architectures employed for both classes.

Feedforward NNs

One of the simplest and widely used variants of NNs. They consist of multiple layers, with information flowing in a unidirectional manner from the input layer through one or more hidden layers to the output layer. The network may or may not have hidden node layers, making their architecture shallow or deep.

Here, we discuss three popular types of feedforward NNs applicable in closed-loop control.

Radial Basis Function Neural Networks (RBFNN)

RBFNNs are networks often used in control applications to estimate nonlinearities such as model uncertainties and external disturbances. As depicted in Figure 2.1, the estimation Δ(x)\Delta(x) can be described as follows

Δ^(x)=W^Tμ(x),(1)\widehat{\Delta}(x) = \widehat{W}^{T}\mu(x), \tag*{(1)}

where x∈Rmx \in\mathbb{R}^{m} is the input vector, W^∈Rn×k\widehat{W} \in\mathbb{R}^{n \times k} is the adjustable weight matrix, μ(X):Rm→Rn\mu(X): \mathbb{R}^{m} \to\mathbb{R}^{n} is the vector of basis functions, nn is the number of neurons in the network, and kk is the number of outputs, which is the length of the vector Δ(x)\Delta(x).

RBFNNs utilize radial basis functions as activation functions. The activation function is typically a Gaussian function, given by

μi(x)=exp⁡(−(x−x0i)T(x−x0i)2σi2),(2)\mu_i(x) = \exp\left(-\frac{(x-x_{0i})^{T}(x-x_{0i})}{2\sigma_i^{2}}\right), \tag*{(2)}

where x0i∈Rmx_{0i} \in\mathbb{R}^{m} is the vector of center of the iith basis function, and σ\sigma is the width of the Gaussian function.

Multilayer Perceptron (MLP)

MLP is a versatile feedforward NN architecture consisting of multiple layers of neurons with nonlinear activation functions, allowing the network to learn complex relationships. The universal approximation theorem states that MLPs with just a single hidden layer can approximate any continuous function, given enough hidden units. However, they require proper tuning of hyperparameters like depth, width, and learning rates. In the context of machine learning, a hyperparameter is distinct from a parameter in that they are set before the training process begins, whereas parameters are learned by the network during training. Hyperparameters govern the overall architecture and learning behavior of the network, while parameters represent the specific weights and biases of the connections between neurons. Carefully selecting hyperparameters is crucial for achieving optimal performance and generalization to new data.

Some commonly used activation functions are:

  • Sigmoid: f(x)=11+e−xf(x) = \frac{1}{1+e^{-x}},

  • Hyperbolic tangent (tanh),

  • Rectified linear unit (ReLU): f(x)=max⁡(0,x)f(x) = \max(0,x).

The sigmoid and tanh functions are nonlinear activation functions that allow the network to learn complex patterns in data. However, they can lead to a problem called vanishing gradients, where the error signal becomes very small as it propagates through the network, making it difficult for the weights to update effectively. ReLUs avoid this problem and enable faster training but can get stuck at zero gradients, meaning that the weights are unable to learn further. Variations like leaky ReLUs (f(x)=max⁡(0.01x,x)f(x) = \max(0.01x,x)) help alleviate the latter issue.

Backpropagation is an algorithm that allows the network to learn the optimal weights and biases by minimizing a loss function. The loss function represents the error between the network outputs and true targets that we aim to minimize. The goal of backpropagation is to propagate the error signal backward from the output layer through the network to update the weights in a way that reduces the overall error. To this end, gradient descent is used to update the weights along the negative gradient direction to reduce the loss.

During training, an MLP learns the optimal weights and biases using backpropagation to minimize a loss function. The key steps include:

  1. Forward pass: Compute the output by passing input features through the network layer-by-layer.

  2. Backward pass: The error gradient is calculated at the output layer and propagated backward through the network using the chain rule of derivatives. This process involves calculating the error gradient for each intermediate layer and using it to update the weights of the corresponding connections.

  3. Update weights: The weights are updated using gradient descent, which involves moving the weights in a direction that reduces the error. The magnitude of the

update is determined by the learning rate, which is a hyperparameter that controls the speed of learning.

This process is repeated multiple times over different training data examples, known as epochs, until the loss function converges and the network learns to produce accurate outputs for a wide range of inputs.

Wavelet Neural Network (WNN)

WNNs, introduced in the 1990s by (author?), merge wavelet theory with NNs, allowing them to efficiently deal with time-frequency analysis. WNNs are regarded as a generalization of RBFNNs [51], as they combine the universal approximation capabilities of neural networks with the properties of wavelet decomposition [52].

The basic structure of a WNN consists of an input layer, a single hidden layer employing wavelets as activation functions, and an output layer. The hidden layers contain wavelet nodes or wavelons, performing wavelet transforms on the inputs.

Wavelet transform is a magnifying glass for signals. It can zoom in to reveal high-frequency components or zoom out to reveal low-frequency components. Unlike the Fourier transform, which uses a fixed set of sinusoidal basis functions, the wavelet transform uses a set of wavelet basis functions that are localized in both time and frequency. This localization allows wavelet transform to extract local spectral and temporal information simultaneously, making it suitable for analyzing non-stationary and transient signals. The transform operates by convolving the original signal to the scaled and translated versions of the wavelet basis functions, whereby a large transform value is observed if the scale and position of the wavelet match the signal.

Common choices for the wavelet activation function include Gaussian wavelet, Mexican hat and Morlet wavelets, as depicted in Figure 2.2. The Mexican Hat wavelet is the second derivative of a Gaussian function and is defined by the following equation

Figure 2.2. Graphical representation of common mother wavelet functions

ϕ(xi)=(1−xi2)exp⁡(−xi22).(3)\phi(x_i) = (1-x_i^2)\exp\left(-\frac{x_i^2}{2}\right). \tag*{(3)}

Each wavelon is obtained by translating and dilating the mother wavelet

Φij(xi)=ϕij(xi−cijbij),(4)\Phi_{ij}(x_i)=\phi_{ij}\left(\frac{x_i-c_{ij}}{b_{ij}}\right), \tag*{(4)}

where bijb_{ij} and cijc_{ij} are the dilation and translation factors respectively.

The output of the hidden layer is obtained by taking the product of wavelets from each input dimension

μj(x)=∏i=1mΦij(xi).(5)\mu_j(x)=\prod_{i=1}^{m}\Phi_{ij}(x_i). \tag*{(5)}

Finally, the output layer combines the hidden layer outputs linearly to produce the network output

y(x)=∑i=1mωixi+∑j=1nwjμj(x)+wn+1,(6)y(x)=\sum_{i=1}^{m}\omega_i x_i+\sum_{j=1}^{n}w_j\mu_j(x)+w_{n+1}, \tag*{(6)}

where another set of weights ωi\omega_i is introduced to directly connect the input layer to the output layer. Accordingly, the weighting vector is as follows

W^=[ωi,wj,wn+1].(7)\hat{W}=[\omega_i,w_j,w_{n+1}]. \tag*{(7)}
Recurrent NNs

Recurrent Neural Networks (RNNs) are a class of artificial neural networks that incorporate feedback connections, introducing dynamics and memory to the networks. Unlike feedforward neural networks, RNNs can process sequences of inputs by maintaining an internal state that takes into account previous inputs and computations. This makes RNNs well-suited for processing time series data and sequence modeling tasks.

Elman Neural Networks (ENN)

The Elman network, introduced by Jeffrey (author?) in 1990, has a simple architecture that includes context units to store the previous hidden state of the network. As shown in Figure 2.3, the context layer holds a copy of the previous hidden layer output. At each time step, the input is fed into the hidden layer, which computes the activation values that get passed to the output layer to produce the network output. The hidden layer output is also copied to the context layer via a unit delay connection. In the next time step, the contents of the context layer are fed back into the network as additional inputs along with external inputs. The context units provide the network with dynamic temporal behavior and the ability to retain information from the past. The previous hidden states stored in the context layer allow the network to maintain useful temporal information for generating outputs that depend on prior context. The Elman network is trained via BackPropagation Through Time (BPTT), which unfolds the network through time and runs backpropagation on the unfolded network [54]. BPTT computes the error gradient across multiple time steps to update the weights appropriately. The Elman network’s simple architecture makes it easy to implement and train while providing useful sequence learning capabilities. However, it lacks more sophisticated gating mechanisms for controlling information propagation through time.

Elman network architecture

Figure 2.3. Elman network architecture.

NARX (Jordan)

Nonlinear AutoRegressive with eXogenous inputs (NARX) and (author?) networks have architectures that feed the network output back to the input. As depicted in Figure 2.4, the NARX model takes current and past input values along with past output values as input to produce the current output. The tapped delay lines on the input (not shown in the figure) and output store previous values to provide temporal context.

NARX (Jordan) network architecture

Figure 2.4. NARX (Jordan) network architecture.

These networks leverage output feedback and temporal memory to generate outputs that depend on past computations. The feedback path allows the networks to learn to make predictions based on prior outputs. NARX/Jordan networks are also trained with BPTT. An advantage over the Elman network is the ability to optimize weight connections from outputs to both hidden and input layers. However, they still lack sophisticated mechanisms for storing information over long time lags.

Long Short-Term Memory (LSTM)

LSTM networks were introduced by (author?) in 1997 to specifically address the problem of learning long-term dependencies. LSTMs contain special units called memory cells that include gates to control information flow. As illustrated in Figure 2.5, each LSTM unit consists of:

LSTM unit architecture

Figure 2.5. LSTM unit architecture.

  • A forget gate that determines what information to discard from the unit

  • An input gate that decides what new information to store in the unit

  • An output gate that controls the output activation of the unit

  • A self-connected memory cell unit to store state information

These gates adaptively reset or update the memory cell and allow LSTM networks to model long-term sequences in data. The gating mechanisms regulate information flow and mitigate the vanishing gradient problem. Numerous LSTM variants have been developed, but all use gated memory cells for learning temporal dependencies.

As depicted in Figure 2.5, the underlying mathematics behind an LSTM network is as follows

it=σ(Wixt+Uiht−1+bi),(8)i_t = \sigma(W_i x_t + U_i h_{t-1} + b_i), \tag*{(8)}
ft=σ(Wfxt+Ufht−1+bf),(9)f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f), \tag*{(9)}
ot=σ(Woxt+Uoht−1+bo),(10)o_t = \sigma(W_o x_t + U_o h_{t-1} + b_o), \tag*{(10)}
c~t=tanh⁡(Wcxt+Ucht−1+bc),(11)\tilde{c}_t = \tanh(W_c x_t + U_c h_{t-1} + b_c), \tag*{(11)}
ct=ft⊙ct−1+it⊙c~t,(12)c_t = f_t \odot c_{t-1} + i_t \odot\tilde{c}_t, \tag*{(12)}
ht=ot⊙tanh⁡ct,(13)h_t = o_t \odot\tanh c_t, \tag*{(13)}

where

  • xtx_t is the input at time step tt,

  • ht−1h_{t-1} is the hidden state output from the previous time step,

  • W,U,bW, U, b are learnable weight matrices and biases,

  • σ\sigma is the sigmoid activation function,

  • tanh is the hyperbolic tangent activation,

  • and, ⊙\odot denotes element-wise multiplication.

Single-layer vs multi-Layer NNs

The universal approximation theorem shows that a feedforward network with just a single hidden layer containing a finite number of neurons can approximate any continuous function, given appropriate weights. This property enables the modeling of complex uncertainties and dynamics with NNs. However, if the number of hidden nodes is increased excessively, it can result in the overfitting problem [57]. Overall, deep networks with multiple layers provide greater representation power over shallow single-hidden-layer models. However, they can be more difficult to train due to issues like vanishing gradients.

Single-hidden-layer NNs

As we mentioned before, the simplest NN structure consists of an input layer, a single hidden layer, and an output layer.

Despite their simplicity, single hidden layer NNs with enough nodes can approximate any continuous function. They have been applied to model aircraft dynamics [58], design flight controllers [59], and perform system identification [60]. The Extreme Learning Machine (ELM) is a popular variant that randomly initializes the hidden layer and only trains the output weights [48], which simplifies the training process.

To sum up, single-layer NNs offer efficient function fitting but may require more nodes than deep networks. Their simple structure facilitates analysis for control design.

Deep NNs

Instead of increasing the number of hidden nodes in a single hidden layer, deep NNs utilize more hidden layers that can efficiently represent highly nonlinear relationships and data patterns. Common deep architectures include MLPs, Convolutional Neural Networks (CNNs), and RNNs like LSTM.

CNN

CNNs are a significant type of deep NNs that have a cascade connection structure. Each CNN cell consists of a convolution layer, a sub-sampling layer (pooling layer), and a fully connected layer as the last layer.

The convolutional layers utilize filters to extract characteristics within the input data. The pooling layers then reduce the size of the convolutional layer’s output, decreasing data complexity by reducing the dimensions of the convolved features and improving computational efficiency. This cycle can be repeated multiple times, with additional convolutional and pooling layers to learn increasingly complex representations. Finally, the fully connected layers assess the previous layer’s output to make the ultimate classification decision based on the learned features.

CNNs are a powerful tool for the identification of nonlinear systems. They have several features that make them suitable for this task.

Firstly, the convolution operation in CNNs is similar to the input-output relation of linear time-invariant systems. This allows CNNs to capture the underlying patterns and relationships in the data.

Secondly, CNNs employ sparse connectivity and shared weights. This means that each neuron in the network is only connected to a small number of neurons in the previous layer, reducing the number of parameters in the network. Additionally, the shared weights allow the network to generalize well to new data, reducing the risk of overfitting.

Another important feature of CNNs is multi-level pooling. Pooling layers down-sample the input by summarizing local features, resulting in a more robust identification scheme against measurement noises. This helps to improve the accuracy and reliability of the system identification process.

In summary, a CNN is a type of NN that is commonly used for image processing and computer vision tasks. It consists of multiple layers, able to learn hierarchical representations of the input data, from low-level features such as edges and corners to high-level features such as object parts and whole objects. While their application in flight control systems is still limited, they have been successfully used in high-level control schemes such as localization and path planning.

Training approaches

Having looked at different structures for NNs, we can discuss two distinct approaches for training them for identification and control purposes.

Consider an RBFNN that needs to be trained. The following methods can be used:

(a) Learning in the open-Loop structure

  • Hybrid learning method: Uses k-means clustering (unsupervised learning) to determine the RBF centers and widths in the hidden layer. The output layer weights are tuned through regression (supervised learning).

  • ELM: An ELM can be used to randomly initialize the RBFNN centers and widths and then train the weights of the network in one shot.

(b) Learning in the closed-Loop structure

The RBFNN can be trained online in the closed-loop using the FEL method based on the Lyapunov stability theory. This approach introduces a parallel training strategy in relation to supervised or unsupervised learning scenarios.

  • ELM: Similar to the open-loop structure, an ELM can be used in a closed-loop structure, with random initialization of hidden layer parameters. Subsequently, FEL provides with an adaptive update rule to train the network weights.

  • Fully-tuned NN: All parameters including output weights and RBF centers and widths are tuned adaptively based on feedback error signals.

Open-loop methods identify neural network parameters using only input-output data, without feedback from the system. Closed-loop techniques, on the other hand, rely on feedback from the real system to train the network. FEL-based tuning utilizes this online feedback to ensure stability during adaptation. Selecting the right training approach depends on the problem, objectives (identification/control), constraints, availability of data, and performance requirements.

Feedback Error Learning

Here, we will introduce the fundamental theoretical basis for the most commonly used approach to incorporate NNs within the adaptive control design process, i.e. the FEL method, while the application of such an algorithm in flight control systems will be addressed in the following subsections. FEL can effectively integrate the control design procedure and the online updating law for the parameters of the NN, which is utilized to compensate for model uncertainties and disturbances. Accordingly, in a general view, the control block can consist of a conventional controller in the inner loop to stabilize the system dynamics, and the neural controller acts as an aid to the controller to compensate for model uncertainties. This arrangement is depicted in Figure 2.6. As shown in the figure, the weights of the neural network are adaptively adjusted using an update rule, which will be further elaborated upon. These updated weights then modify the control law implemented by the conventional controller. Thus,

Schematic diagram of FEL control law

Figure 2.6. Schematic diagram of FEL control law.

employing a composite Lyapunov function including both the tracking error and the estimation error of NN parameters, all signals in the closed-loop system remain bounded, even in the presence of model uncertainties and external disturbances. To be more precise, consider the dynamic model of an aircraft (in the affine form) as follows

x˙=F(x)+B(x)u+Δ(x),(14)\dot{x} = F(x) + B(x)u + \Delta(x), \tag*{(14)}

where x,u∈Rnx,u \in\mathbb{R}^{n} and Δ(x)\Delta(x) represents a lumped uncertainty that encapsulates all state-dependent model uncertainties and external disturbances. It’s worth noting that the portion of the uncertainty which is a explicit function of time is not included here and will be addressed in Chapter 3. Defining the desired trajectory as xdx_d, the tracking error is obtained as e=x−xde=x-x_d. Now, the control command can be defined as

u=B−1(−F(x)−Δ+x˙d−k1e),(15)u=B^{-1}\left(-F(x)-\Delta+\dot{x}_d-k_1e\right), \tag*{(15)}

where k1k_1 is a positive-definite matrix. However, the vector Δ\Delta is unknown. Thus, it is approximated by an NN (such as RBFNN or MLP) as Δ^=W^Tμ(x)\hat{\Delta}=\hat{W}^{T}\mu(x), where μ\mu represents the vector of basis functions (corresponding to hidden layers of the NN) and WW indicates the matrix of unknown weights that should be identified. Such a formulation can be used to represent different feedforward and recurrent NNs. Here, we will use such a general formulation, and for brevity, do not address different possible network structures and their advantages and disadvantages in flight control systems (for more details, see [61]). Accordingly, due to the universal approximations property of NNs, we have

References

  1. [1]Bruce T. Clough and Wright Patterson. Metrics, Schmetrics! How The Heck Do You Determine A UAV’s Autonomy Anyway? In Proceedings of the 2002 Performance Metrics for Intelligent Systems Workshop, Gaithersburg, MD, 2002.
  2. [2]Linda S. Gottfredson. Mainstream Science on Intelligence: An Editorial With 52 Signatories, History, and Bibliography. INTELLIGENCE, 24(1):13–23, 1997.DOI
  3. [3]Lyle Long and Troy Kelley. The Requirements and Possibilities of Creating Conscious Systems. In AIAA Infotech@ Aerospace Conference, Seattle, Washington, 2009.DOI
  4. [4]Seyyed Ali Emami, Paolo Castaldi, and Afshin Banazadeh. Neural network-based flight control systems: Present and future. Annual Reviews in Control, 53:97–137, 2022.DOI
  5. [5]Thaddeus Eze, Richard Anthony, Alan Soper, and Chris Walshaw. A Generic Approach towards Measuring Level of Autonomicity in Adaptive Systems. International Journal on Advances in Intelligent Systems, 5(3-4), 2012.
  6. [6]Marialena Vagia, Aksel A. Transeth, and Sigurd A. Fjerdingen. A literature review on the levels of automation during the years. What are the different taxonomies that have been proposed? Applied ergonomics, 53 Pt A:190–202, 2016.DOI
  7. [7]Marco Protti and Riccardo Barzan. UAV Autonomy – Which level is desirable? – Which level is acceptable? Alenia Aeronautica Viewpoint. In Platform Innovations and System Integration for Unmanned Air, Land and Sea Vehicles, Neuilly-sur-Seine, France, 2007.
  8. [8]Andriy Sarabakha, Changhong Fu, Erdal Kayacan, and Tufan Kumbasar. Type-2 Fuzzy Logic Controllers Made Even Simpler: From Design to Deployment for UAVs. IEEE Transactions on Industrial Electronics, 65(6):5069–5077, 2018.DOI
  9. [9]Kirk Y. W. Scheper, Sjoerd Tijmons, Cornelis C. de Visser, and Guido C. H. E. de Croon. Behavior Trees for Evolutionary Robotics. Artificial life, 22(1):23–48, 2016.DOI
  10. [10]G. C. H. E. de Croon, M. Perçin, B. D. W. Remes, C. de Wagter, and R. Ruijsink. The DelFly: design, aerodynamics, and artificial intelligence of a flapping wing robot / G.C.H.E. de Croon, M. Perçin, B.D.W. Remes, R. Ruijsink, C. de Wagter. Springer, Berlin, 2016.DOI
  11. [11]Jemin Hwangbo, Inkyu Sa, Roland Siegwart, and Marco Hutter. Control of a Quadrotor With Reinforcement Learning. IEEE Robotics and Automation Letters, 2(4):2096–2103, 2017.DOI
  12. [12]Lu Cheng, ChangSheng Jiang, and Ming Pu. Online-SVR-compensated non-linear generalized predictive control for hypersonic vehicles. Science China Information Sciences, 54(3):551–562, 2011.DOI
  13. [13]Eunjung Ju, Jungdam Won, Jehee Lee, Byunguk Choi, Junyong Noh, and Min Gyu Choi. Data-driven control of flapping flight. ACM Transactions on Graphics, 32(5):1–12, 2013.DOI
  14. [14]Dario Floreano, Jean-Christophe Zufferey, and Jean-Daniel Nicoud. From wheels to wings with evolutionary spiking circuits. Artificial life, 11(1-2):121–138, 2005.
  15. [15]Fernando Silva, Miguel Duarte, Luís Correia, Sancho Moura Oliveira, and Anders Lyhne Christensen. Open Issues in Evolutionary Robotics. Evolutionary Computation, 24(2):205–236, 2016.DOI
  16. [16]Fendy Santoso, Matthew A. Garratt, and Sreenatha G. Anavatti. State-of-the-Art Intelligent Flight Control Systems in Unmanned Aerial Vehicles. IEEE Transactions on Automation Science and Engineering, 15(2):613–627, 2018.DOI
  17. [17]C. C. Lee. Fuzzy logic in control systems: Fuzzy logic controller. I. IEEE Transactions on Systems, Man, and Cybernetics, 20(2):404–418, 1990.DOI
  18. [18]Giampiero Campa, Mario L. Fravolini, Brad Seanor, Marcello R. Napolitano, Diego Del Gobbo, Gu Yu, and Srikanth Gururajan. On-line learning neural networks for sensor validation for the flight control system of a B777 research scale model. International Journal of Robust and Nonlinear Control, 12(11):987–1007, 2002.DOI
  19. [19]Janusz Kacprzyk, Johann Schumann, and Yan Liu, editors. Applications of Neural Networks in High Assurance Systems. Studies in Computational Intelligence. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.
  20. [20]James E. Tomayko. The Story of Self-Repairing Flight Control Systems. NASA Dryden Flight Research Center, 2003.
  21. [21]M. Steinberg. Historical Overview of Research in Reconfigurable Flight Control. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 219(4):263–275, 2005.DOI
  22. [22]Johann Schumann, Pramod Gupta, and Yan Liu. Application of Neural Networks in High Assurance Systems: A Survey. In Janusz Kacprzyk, Johann Schumann, and Yan Liu, editors, Applications of Neural Networks in High Assurance Systems, volume 268 of Studies in Computational Intelligence, pages 1–19. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.DOI
  23. [23]Markus Kaminski. A Rapid Robust Fault Detection Algorithm for Flight Control Reconfiguration. Master Thesis, Germany, 2017.
  24. [24]Peggy Williams. Selected Flight Test Results for Online Learning Neural Network-Based Flight Control System. In AIAA 1st Intelligent Systems Technical Conference, Chicago, Illinois, 2004. AIAA.
  25. [25]Jacob Hageman, Mark Smith, and Susan Statchowiak. Integration of Online Parameter Identification and Neural Network for In-Flight Adaptive Control. NASA Dryden Flight Research Center, 2003.DOI
  26. [26]Tim Smith, Jim Barhorst, and James M. Urnes. Design and Flight Test of an Intelligent Flight Control System. In Janusz Kacprzyk, Johann Schumann, and Yan Liu, editors, Applications of Neural Networks in High Assurance Systems, volume 268 of Studies in Computational Intelligence, pages 57–76. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.
  27. [27]John T. Bosworth and Peggy S. Williams-Hayes. Flight Test Results from the NF-15B Intelligent Flight Control System (IFCS) Project with Adaptation to a Simulated Stabilator Failure. In AIAA Conference and exhibit, AIAA infotech@Aerospace Conference and Exhibit, Rohnert Park, California, 2007. AIAA.DOI
  28. [28]John Burken, Curt Hanson, Jim Lee, and John Kaneshige. Flight Test Comparison of Different Adaptive Augmentations of Fault Tolerant Control Laws for a Modified F-15 Aircraft. In AIAA Infotech@Aerospace Conference, Seattle, Washington, 2009.DOI
  29. [29]Curt Hanson, Jacob Schaefer, Marcus Johnson, and Nhan Nguyen. Design of Low Complexity Model Reference Adaptive Controllers. NASA Dryden Flight Research Center, 2012.
  30. [30]Curt Hanson, Jacob Schaefer, John J. Burken, David Larson, and Marcus Johnson. Complexity and Pilot Workload Metrics for the Evaluation of Adaptive Flight Controls on a Full Scale Piloted Aircraft. NASA Dryden Flight Research Center, 2014.
  31. [31]Konstantinos Dalamagkidis, Kimon P. Valavanis, and Les A. Piegl. Nonlinear Model Predictive Control With Neural Network Optimization for Autonomous Autorotation of Small Unmanned Helicopters. IEEE Transactions on Control Systems Technology, 19(4):818–831, 2011.DOI
  32. [32]Zhijun Li, Jun Deng, Renquan Lu, Yong Xu, Jianjun Bai, and Chun-Yi Su. Trajectory-Tracking Control of Mobile Robot Systems Incorporating Neural-Dynamic Optimized Model Predictive Approach. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 46(6):740–749, 2016.DOI
  33. [33]Mehmet Önder Efe. Neural Network Assisted Computationally Simple PID Control of a Quadrotor UAV. IEEE Transactions on Industrial Informatics, 7(2):354–361, 2011.DOI
  34. [34]Bin Xu, Chenguang Yang, and Yongping Pan. Global neural dynamic surface tracking control of strict-feedback systems with application to hypersonic flight vehicle. IEEE Transactions on Neural Networks and Learning Systems, 26(10):2563–2575, 2015.DOI
  35. [35]Bin Xu, Daipeng Yang, Zhongke Shi, Yongping Pan, Badong Chen, and Fuchun Sun. Online Recorded Data-Based Composite Neural Control of Strict-Feedback Systems With Application to Hypersonic Flight Dynamics. IEEE Transactions on Neural Networks and Learning Systems, 29(8):3839–3849, 2018.DOI
  36. [36]Karl J. Åström and Björn Wittenmark. Adaptive control. Dover books on engineering. Dover Publications, Mineola N.Y., 2nd ed., dover ed. edition, 2008.
  37. [37]Weibin Gu, Kimon P. Valavanis, Matthew J. Rutherford, and Alessandro Rizzo. A Survey of Artificial Neural Networks with Model-based Control Techniques for Flight Control of Unmanned Aerial Vehicles. In 2019 International Conference on Unmanned Aircraft Systems (ICUAS), pages 362–371, Atlanta, GA, USA, 2019. IEEE.DOI
  38. [38]P. A. Ioannou, A. M. Annaswamy, K. S. Narendra, S. Jafari, L. Rudd, R. Ortega, and J. Boskovic. L1-Adaptive Control: Stability, Robustness, and Interpretations. IEEE Transactions on Automatic Control, 59(11):3075–3080, 2014.DOI
  39. [39]Nhan Nguyen, Kalmanje Krishnakumar, John Kaneshige, and Pascal Nespeca. Flight dynamics and hybrid adaptive control of damaged aircraft. Journal of Guidance, Control, and Dynamics, 31(3):751–764, 2008.DOI
  40. [40]A. J. Calise. Neural networks in nonlinear aircraft flight control. IEEE Aerospace and Electronic Systems Magazine, 11(7):5–10, 1996.
  41. [41]Martin T. Hagan, Howard B. Demuth, Mark Hudsun Beale, and Orlando de Jesús. Neural network design. 2nd ed. edition, 2016.
  42. [42]Bin Xu and Yu Zhang. Neural discrete back-stepping control of hypersonic flight vehicle with equivalent prediction model. Neurocomputing, 154:337–346, 2015.DOI
  43. [43]Bin Xu, Yongping Pan, Danwei Wang, and Fuchun Sun. Discrete-time hypersonic flight control based on extreme learning machine. Neurocomputing, 128:232–241, 2014.DOI
  44. [44]F. Rinaldi, S. Chiesa, and F. Quagliotti. Linear Quadratic Control for Quadrotors UAVs Dynamics and Formation Flight. Journal of Intelligent & Robotic Systems, 70(1-4):203–220, 2013.
  45. [45]Valeria Artale, Mario Collotta, Cristina Milazzo, Giovanni Pau, and Angela Ricciardello. An Integrated System for UAV Control Using a Neural Network Implemented in a Prototyping Board. Journal of Intelligent and Robotic Systems, 2016.DOI
  46. [46]Jih-Gau Juang and Kai-Chung Cheng. Application of Neural Networks to Disturbances Encountered Landing Control. IEEE Transactions on Intelligent Transportation Systems, 7(4):582–588, 2006.DOI
  47. [47]Sefer Kurnaz, Omer Cetin, and Okyay Kaynak. Adaptive neuro-fuzzy inference system based autonomous flight control of unmanned air vehicles. Expert Systems with Applications, 37(2):1229–1234, 2010.
  48. [48]Seyyed Ali Emami and Alireza Roudbari. Multimodel ELM-Based Identification of an Aircraft Dynamics in the Entire Flight Envelope. IEEE Transactions on Aerospace and Electronic Systems, 55(5):2181–2194, 2019.DOI
  49. [49]Erdal Kayacan. Fuzzy neural networks for real time control applications: Concepts, modeling and algorithms for fast learning. Elsevier and Butterworth-Heinemann, Amsterdam and Oxford UK, 2016.
  50. [51]Antonios K. Alexandridis and Achilleas D. Zapranis. Wavelet neural networks: A practical guide. Neural networks : the official journal of the International Neural Network Society, 42:1–27, 2013.DOI
  51. [52]Seyyed Ali Emami and Alireza Roudbari. Identification of nonlinear time-varying systems using wavelet neural networks. Advanced Control for Applications: Engineering and Industrial Systems, 2(4):e59, 2020.DOI
  52. [54]P. J. Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10):1550–1560, 1990.DOI
  53. [57]Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.
  54. [58]Seyyed Ali Emami and Kasra K. A. Ahmadi. A self-organizing multi-model ensemble for identification of nonlinear time-varying dynamics of aerial vehicles. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering, 235(7):1164–1178, 2021.DOI
  55. [59]Bin Xu, Xia Wang, and Zhongke Shi. Robust Adaptive Neural Control of Non-minimum Phase Hypersonic Vehicle Model. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 51(2):1107–1115, 2021.
  56. [60]Heidar A. Talebi, Farzaneh Abdollahi, Rajni V. Patel, and Khashayar Khorasani. Neural Network-Based State Estimation of Nonlinear Systems, volume 395. Springer New York, New York, NY, 2010.DOI
  57. [61]Weibin Gu, Kimon P. Valavanis, Matthew J. Rutherford, and Alessandro Rizzo. UAV Model-based Flight Control with Artificial Neural Networks: A Survey. Journal of Intelligent and Robotic Systems, 100(3-4):1469–1491, 2020.DOI
  58. [62]Kumpati S. Narendra and Anuradha M. Annaswamy. Persistent excitation in adaptive systems. International Journal of Control, 45(1):127–160, 1987.DOI
  59. [63]Kumpati S. Narendra and Anuradha M. Annaswamy. Stable adaptive systems. Dover Publications, Mineola N.Y., 2005.
  60. [64]Mohd Ariffanan Mohd Basri, Abdul Rashid Husain, and Kumeresan A. Danapalasingam. Intelligent adaptive backstepping control for MIMO uncertain non-linear quadrotor helicopter systems. Transactions of the Institute of Measurement and Control, 37(3):345–361, 2015.DOI
  61. [65]Anthony J. Calise, Naira Hovakimyan, and Moshe Idan. Adaptive output feedback control of nonlinear systems using neural networks. Automatica, 37(8):1201–1211, 2001.
  62. [66]Nakwan Kim. Improved methods in neural network based adaptive output feedback control, with applications to flight control. PhD Thesis, School of Aerospace Engineering, 2003.
  63. [67]Girish Chowdhary, Maximilian Mühlegg, and Eric Johnson. Exponential parameter and tracking error convergence guarantees for adaptive controllers without persistency of excitation. International Journal of Control, 87(8):1583–1603, 2014.DOI
  64. [68]Taeyoung Lee and Youdan Kim. Nonlinear Adaptive Flight Control Using Backstepping and Neural Networks Controller. Journal of Guidance, Control, and Dynamics, 24(4):675–682, 2001.DOI
  65. [69]Shunshuai Li, Yaonan Wang, Jianhao Tan, and Yan Zheng. Adaptive RBFNNs/integral sliding mode control for a quadrotor aircraft. Neurocomputing, 216:126–134, 2016.DOI
  66. [70]Samir Zeghlache, Hemza Mekki, Abderrahmen Bouguerra, and Ali Djerioaui. Actuator fault tolerant control using adaptive RBFNN fuzzy sliding mode controller for coaxial octorotor UAV. ISA Transactions, 80:267–278, 2018.DOI
  67. [71]Yongduan Song, Liu He, Dong Zhang, Jiyue Qian, and Jin Fu. Neuroadaptive fault-tolerant control of quadrotor UAVs: a more affordable solution. IEEE Transactions on Neural Networks and Learning Systems, 30(7):1975–1983, 2019.DOI
  68. [72]Cox David. The generic transport model, 2019. Accessed: 2023-08-09.
  69. [73]Richard M. Hueschen. Development of the transport class model (tcm) aircraft simulation from a sub-scale generic transport model (gtm) simulation. Technical report, NASA/TM-2011-217169, 2011.
  70. [74]Dong-Ho Shin and Youdan Kim. Nonlinear discrete-time reconfigurable flight control law using neural networks. IEEE Transactions on Control Systems Technology, 14(3):408–422, 2006.
  71. [75]T. Zhang, S. S. Ge, and C. C. Hang. Design and performance analysis of a direct adaptive controller for nonlinear systems. Automatica, 35(11):1809–1817, 1999.DOI
  72. [76]S. S. Ge and Cong Wang. Direct adaptive NN control of a class of nonlinear systems. IEEE Transactions on Neural Networks, 13(1):214–221, 2002.DOI
  73. [77]M. Vijaya Kumar, S. Suresh, S. N. Omkar, Ranjan Ganguli, and Prasad Sampath. A direct adaptive neural command controller design for an unstable helicopter. Engineering Applications of Artificial Intelligence, 22(2):181–191, 2009.
  74. [78]E. Tzirkel-Hancock and F. Fallside. Stable control of nonlinear systems using neural networks. International Journal of Robust and Nonlinear Control, 2(1):63–86, 1992.
  75. [79]S. Fabri and V. Kadirkamanathan. Dynamic structure neural networks for stable adaptive control of nonlinear systems. IEEE Transactions on Neural Networks, 7(5):1151–1167, 1996.pubmed.ncbi.nlm.nih.gov/18263511
  76. [80]Jovan D. Boskovic, Lingji Chen, and Raman K. Mehra. Adaptive Control Design for Nonaffine Models Arising in Flight Control. Journal of Guidance, Control, and Dynamics, 27(2):209–217, 2004.DOI
  77. [81]Petros A. Ioannou and Petar V. Kokotovic, editors. Adaptive Systems with Reduced Models, volume 47 of Lecture Notes in Control and Information Sciences. Springer, Berlin and Heidelberg, 1983.
  78. [82]P. A. Ioannou and Jing Sun. Robust adaptive control. Dover Publications Inc, Mineola New York, 2012.

Paper details

vlnc:W_68rsu79z
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents