· · · Institutional Repository

The future demand for larger and lighter civil transport aircraft leads to more flexible aircraft, which bring their own controlling and handling problems. A review of established handling qualities methods showed that they were either unsuitable for aeroelastic aircraft, or had significant disadvantages. After consideration of the basic principles behind a number of handling qualities methods, a new handling qualities method was developed, the Experimental Behavior Measurement Method (EBMM). This new method is based on the principle that a satisfactory match between the aircraft characteristics and the human operator's behavioral characteristics is required to acquire proper vehicle handling qualities. The EBMM requires pilots to perform a number of pitch tracking tasks with a pursuit display, in the aircraft or a moving-base simulator. During these tasks the tracking signal bandwidth is increased, while the pilot's control behavior is determined using system identification techniques. A sudden decrease of the crossover frequency in the pilot's control behavior can be observed when the tracking signal bandwidth exceeds the pilot-vehicle capabilities. This phenomenon is called crossover regression, and the bandwidth at which crossover regression occurs is defined as the crossover-regression frequency. Since the crossover-regression frequency is dependent on the pilot-vehicle capabilities, it can be considered to be a measure of the handling qualities. The validity and applicability of the EBMM were investigated by conducting an experiment in the TU Delft SIMONA flight-simulator. Three aircraft models with varying levels of aeroelasticity were evaluated, using both the new EBMM as well as the well-established Cooper-Harper rating method, which assesses flying qualities as a surrogate for handling qualities. When the effects of the amount of aeroelasticity on the flying and handling qualities were compared, both a quantitative and qualitative correlation was seen between the results of the Cooper-Harper rating method and the new EBMM. In conclusion, the results suggest that the new EBMM method can contribute to the development of improved handling qualities of large and flexible future aircraft aircraft.

Optimization is an important element in aerospace related research. It is encountered for example in trajectory optimization problems, such as: satellite formation flying, spacecraft re-entry optimization and airport approach and departure optimization; in control optimization, for example in adaptive control algorithms; and in system identification problems, such as online aircraft model identification or human perception modeling. The main goal of this thesis is to investigate how Interval Analysis (IA) can be used as a tool for aerospace related optimization problems; to examine its theoretical and practical limitations, and to explore the ways in which optimization algorithms can benefit from interval analysis. A subset of goals is to improve the solutions for a number of aerospace related optimization problems. The scientific contribution of this thesis consists of the design and implementation of interval optimization algorithms for four important aerospace problems. The first contribution concerns finding the trim points for a nonlinear aircraft model. Trim points, defined as the combination of control settings for which all linear and rotational accelerations on the aircraft are zero, are important for flight control system design, since they provide information about the flight envelope and stability properties of the aircraft. Unlike other trim algorithms, the interval based method can guarantee that all trim points are found. In the second application, an interval optimization algorithm is developed for fitting pilot input/output data from an experiment in the SIMONA Research Simulator to a multi-modal human perception model. Perception models improve the understanding of how humans perceive motion and are an essential tool in the design of flight simulators. Results show that the minimum of the cost function found by the interval method is lower than the one previously found, resulting in an improved human perception model. This second application particularly demonstrates the capabilities of IA optimization as a parameter identification tool. The third contribution is an interval based algorithm for solving the integer ambiguity problem related to Global Navigation Satellite Systems (GNSS). Phase measurements of the carrier wave of a GNSS signal are used to estimate the length and orientation of baselines between two or more antennas. This estimation procedure contains an optimization problem in which the integer number of carrier wavelengths between antennas has to be determined. The new interval method provides guarantees that correct solutions are found when the measurement noise is encapsulated by an interval number. The final contribution is an interval optimization algorithm that minimizes fuel consumption during rendezvous and docking procedures of satellites in circular orbits. To avoid integration of interval functions, an analytical solution to the system of differential equations that describes the relative motion of the satellites is used to generate trajectories resulting from a set of thruster pulses of varying amplitudes. Introduction of obstacles, in the form of forbidden areas in the path between the two satellites, makes the problem nonlinear, such that gradient-based optimization algorithms can fail to obtain the globally optimal solution. The interval algorithm always converges to the trajectory that avoids all obstacles and results in minimum fuel consumption. It can be concluded that IA is an excellent tool for solving nonlinear optimization algorithms, providing guarantees on obtaining the global minimum of the cost function.

The field of global optimization has been an active one for many years. By far the most applied methods are gradient and evolutionary based algorithms. The most appearing drawback of those types of methods is that one cannot guarantee that the global solution is found within finite time. Moreover, if the global solution is found (by chance), the methods cannot provide a guaranteed feedback to the user stating that the provided solution is the global one. Therefore, no natural stopping conditions are available for most of the existing optimization algorithms. There are, however, other tools available, which do provide the guarantee that the global solution is found and that have natural stopping conditions. Interval analysis in combination with interval arithmetic is such a tool. Interval arithmetic was initially developed to cope with rounding errors in digital computers. Using interval arithmetic, one can perform reliable computing such that catastrophic numeric errors can be prevented (the explosion of the Ariane 5 rocket on June 4, 1996 was caused by a simple numeric overflow). It was soon found, that interval arithmetic could be used to form guaranteed bounds on any type of function or numeric algorithm for any domain. These bounds provide the crucial information needed to perform global optimization. Interval analysis is the group name of all methods that use the information obtained from guaranteed bounds to solve global optimization problems. Developed in the 1960’s, interval analysis gained popularity during the 90’s when digital computers became increasingly powerful. Nowadays, interval analysis has been widely applied in the field of static optimization, i.e. optimization that does not involve differential algebraic equations, and verified integration. However, interval analysis has not been applied often in the field of dynamic optimization. The goal of the research is to investigate whether interval analysis, in combination with interval arithmetic, can be used to solve non-linear, constrained, dynamic optimization problems. Moreover, the possibility of extending existing theory in the field of static optimization is investigated. The focus of the research lies on trajectory optimization (a specific case of dynamic optimization). The most important condition of the designed solvers is that the dynamic constraints, formed by the equations of motion, must be satisfied for all time instances. To reach the research objectives, the theory and application of both interval arithmetic and interval analysis have been thoroughly investigated. The work is divided into two parts. The first part is on static optimization, which includes the discussion on interval arithmetic and describes the basics regarding interval analysis. The existing theory of inclusion functions, formed via interval arithmetic, has been evaluated and extended upon. The development of the Polynomial Inclusion Function, a new type of inclusion function, shows that significant improvements are possible in this field. During the review of interval analysis, its main virtues and limitations were demonstrated. The most important advantages are the guarantee that all optimal solutions are found to any degree of accuracy and that the user knows when the solution set has been found. The main limitation is the curse of dimensionality: the computational load grows, for most problems, exponentially with al linear increase in problem dimension. The author believes that this curse is mainly caused by two aspects of the current implementation of interval analysis. The first aspect is the widening of the inclusion function due to the dependency effects. The dependency effects can be partially prevented by efficient implementation of function evaluations and through application of advanced inclusion functions. However, a generic efficient method for preventing dependency effects is still not available. The other aspect causing the curse of dimensionality is the current inefficient handling of available information. The optimization algorithms within interval analysis are commonly based on branch and bound algorithms. Through a process of elimination, one is left with a list of domains in which the optimal solution set must lie. Current methods for eliminating (part of) the domain, such as the Newton step, do not use the gathered/available information efficiently. This is mainly due to the definition of the domain and the storage of the information, i.e. keeping track of infeasible regions. It is the author’s opinion that this is the reason that the application of interval analysis is limited to solving lower dimensional problems. Despite the curse of dimensionality, interval analysis based solvers can solve complicated, non-linear, constrained problems. This has been shown in multiple chapters in the first part. Complicated problems, such as neural network output optimization and the problem integer ambiguity resolution in the field of Global Navigation Satellite Systems, are solved rigorously by interval analysis based solvers. The applications show that equality and inequality constraints are efficiently handled using interval analysis. Moreover, they show that interval analysis can be used to solve real-life problems and demonstrate that interval analysis is a strong global optimization tool. The second part of the research is on dynamic optimization, thereby focusing on trajectory optimization. The trajectory optimization problem is infinite dimensional with begin and end-point constraints, dynamic constraints (the equations of motion), and possibly additional equality and inequality constraints. The problem is infinite dimensional since the states and controls need to be specified for each time instance. In the field of trajectory optimization one can identify two classes of methods: indirect methods and direct methods. Disregarding the optimization problems for which an analytic solution is present, both classes require a transformation to make the problem solvable. Three transformation methods have been considered: control parameterization, state parameterization, and control and state parameterization. With control parameterization, the control is defined for each time step using a polynomial and the states are computed using explicit integration. For state parameterization, the states are defined and the controls are deduced via the equations of motion (implicit integration). The last method applies parameterization of both the states and controls with respect to time. Trajectories are sought that satisfy the dynamic constraints at given time instances. The nature of the transformation methods implies that the first two methods can be used to find trajectories that satisfy the dynamic constraints at all time instances, while the latter cannot be used for this purpose. Therefore, only the first two methods have been thoroughly investigated. The last method was only briefly reviewed. The main conclusion regarding the control parameterization approach is that it suffers greatly from the required explicit integration. Although verified integration is possible and sharp bounds on the trajectories can be provided, the problem is to prove the existence of a solution within a given domain of the search space. Without the ability to update the estimate of the minimal cost function value early in the optimization process, the computational load becomes very high. Despite the drawback of control parameterization, it has been demonstrate that this approach can be used to find the global solution, although, currently, only very low dimensional problems can be solved. Higher dimensional problems can be solved using the state parameterization approach. By using simplex splines, the begin- and end-point constraints can be implicitly satisfied, which significantly reduces the problem complexity. The limitation is that the approach is only suitable for fully controllable systems. For systems that are not fully controllable one needs to apply explicit integration for all dependent states. This will increase the computational load significantly and would eliminate most of the benefits of the state parameterization approach. An interval analysis based solver has been applied to solve the problem of satellite trajectory planning for formation flying. Although still suffering from the curse of dimensionality, the results demonstrate that interval analysis can be used to solve the problem rigorously. Moreover, it has been shown that the performance of the solver is superior to gradient based solvers when constraints are imposed. The main conclusion of the research is that it is possible to apply interval analysis to dynamic optimization. The current status of the solvers (in this thesis and in literature) allows one to solve only ‘lower’ dimensional problems. Radical changes in the approach of handling information and keeping track of infeasible regions must be made to make interval analysis applicable to higher dimensional problems. Despite the limitations of interval analysis, the presented results clearly demonstrate the virtues of interval analysis based solvers in the field of global optimization. Several new exciting research opportunities have been identified, such as nonlinear stability analysis using interval analysis, the combination of interval analysis and evolutionary algorithms, and a new way of forming inclusion functions to boost the efficiency of interval analysis based solvers. Overall, the potential of interval analysis is very large and the author believes that interval analysis will become one of the most important tools in the field of global optimization in the near future.

At present, model based control systems play an essential role in many aspects of modern society. Application areas of model based control systems range from food processing to medical imaging, and from process control in oil refineries to the flight control systems of modern aircraft. Central to a model based control system is a mathematical model of the physical system or process that is being controlled. The field of science concerned with the identification of models of physical systems is called system identification. In this thesis, a new methodology is proposed for the identification of models of nonlinear systems using multivariate simplex splines. This new methodology has the potential to increase the performance of any model based control system by improving the quality of system models. Multivariate simplex splines consist of polynomial basis functions, called B-form polynomials, which are defined on geometric structures called simplices. Every simplex supports a single B-form polynomial which itself consists of a linear combination of Bernstein basis polynomials. Each individual Bernstein basis polynomial is scaled by a single coefficient called a B-coefficient. The B-coefficients have a special property in the sense that they have a unique spatial location inside their supporting simplex. This spatial structure, also known as the B-net, provides a number of unique capabilities that add to the desirability of the simplex splines as a tool for data approximation. For example, the B-net simplifies local model modification by directly relating specific model regions to subsets of B-coefficients involved in shaping the model in those regions. This particular capability has the potential to play an important role in future adaptive model based control systems. In such a control system, an on-board simplex spline model can be locally adapted in real time to reflect changes in system dynamics. The approximation power of the multivariate simplex splines can be increased by joining any number of simplices together into a geometric structure called a triangulation. Triangulations come in many shapes and sizes, ranging from configurations consisting of just two simplices to configurations containing millions of simplices. Triangulations can be optimized by locally increasing or decreasing the density of simplices to reflect local system complexity. The new methodology was applied in the identification of a complete set of aerodynamic models for the Cessna Citation II laboratory using flight data obtained during seven test flights. In total, 247 flight test maneuvers were flown which together provided a significant coverage of the flight envelope of the Citation II. The complete identification dataset consisted of millions of measurements on more than sixty flight parameters. More than 2000 prototype spline based aerodynamic models were identified using a newly developed, highly optimized software implementation of the simplex spline identification algorithm. Using the developed methods for simplex spline model validation it was proved that the models are both accurate and of guaranteed numerical stability inside the spline domain. The identification and validation results of the simplex spline models were compared with those of ordinary polynomial models identified using standard identification methods. These results showed that the multivariate simplex spline based aerodynamic models were of significantly higher quality than the aerodynamic models based on ordinary polynomials.

Personal air transportation utilizing small aircraft is a market that is expected to grow significantly in the near future. However, seventy times more accidents occur in this segment as compared with the commercial aviation sector. The majority of these accidents is related to handling and control problems. In commercial aviation, Fly-By-Wire (FBW) technology is used to prevent these types of accidents. Instead of downscaling advanced and high-cost FBW platforms, a low-cost solution should be considered for the general aviation market. In the European project “Small Aircraft Future Avionics Architecture”, a FBW platform is developed specifically for small aircraft. In this environment, Flight Control Law (FCL) designs are needed that have robustness against model uncertainties, sensor bias, sensor noise and time delays, while being fast and accurate enough to accommodate the relatively agile dynamics of a small aircraft. FCL designs that meet these requirements are called practical FCL designs in this thesis. Based on a dynamic model of a Diamond DA 42 and a description of the dynamic properties of the FBW platform, two different FCL designs are synthesized and analyzed in this thesis. The first design uses classical control theory and the second design uses a newly developed nonlinear design method, based on backstepping, singular perturbation theory and approximate dynamic inversion. This latter method, called Sensor-Based Backstepping (SBB), uses no dynamic model information and relies solely on measurements. Both FCL designs are compared on sensitivity to parametric uncertainty, sensor noise, disturbances, time delays, handling qualities, design effort, certifiably and the option to add flight envelope protection. In the scope of this thesis, SBB is selected as the preferred FCL design. This method produces good aircraft responses without knowing the exact dynamic behavior of the aircraft during FCL synthesis, as long as the system is minimum phase, controllable and sufficiently time-scale separated.

In high–risk domains like aviation, medicine and nuclear power plant control, automation has enabled new capabilities, increased the economy of operation and has greatly contributed to safety. However, automation increases the number of couplings in a system, which can inadvertently lead to more complexity from the perspective of the operator. The automation of a system transforms the work domain of the human operator, and his role changes from controlling the core processes to managing the automated processes. The complexity of the automation and the lack of proper support can make the control task’s overall difficulty larger than it needs to be, restricting safety, productivity, and efficiency. To address and limit the automation introduced complexity in the operator’s work domain, and to find representations to support him, the ecological approach to automation design was taken. The ecological approach focuses on the relationship between the human operator and his work domain including the system he is controlling. The main research goals were to find how the ecological approach could be used to help limit the automation introduced complexity, and how the ecological approach could be used to support the human operator in controlling automated processes. The formulation of Ecological Automation Design (EAD) was based on the Ecological Interface Design (EID) paradigm. One of the main underlying questions asked about the interface between the work domain and the human operator is: “how to represent work domain complexity?". The inter face design paradigm was transformed into an automation design paradigm by first separating the automation component from the work domain and asking the same underlying question about the interfaces between the work domain, the human operator, and the automation. Then, the conceptual shared domain representation was defined to visualize that the apparent complexity of the system could be reduced when both the human operator and the automation view the same representation of constraints that the work domain imposes on control. As part of the ecological approach, Work Domain Analysis (WDA) was used to analyze and represent the constraints in a work domain. However, WDA is not yet fully developed and suffers from some methodological and conceptual issues. The research therefore, focused on the further development and extension of WDA to include the representation of automated processes. Four case studies were conducted, and each case study generated new insights into the application of and extension of WDA. In the first case study, EID was applied to the design of the Energy Augmented Tunnel In the Sky display. This display was designed to aid a pilot to fly the approach to landing by presenting energy management information. The WDA revealed the significance of the energy coupling between vertical flight path and speed control as an intermediate control goal. Based on the analysis, a creative design process resulted in a novel display that has the energy representations fully, and graphically integrated in the tunnel in the sky display. A preliminary evaluation indicated that the additional energy management information shown in relation to the control actions and control goals helped pilots to fly the approaches. The display is not expected to give a performance increase but to change the way in which pilots control the throttle and elevator to fly approaches. The second case study was the analysis of the already existing Total Energy Control System (TECS). TECS is an unconventional automated flight control system that was based on the same energy management constraints as that were represented in the energy augmented display of the first case study. The design of TECS was mapped onto the abstraction hierarchy to represent the energy management principles as part of the whole automated system. The analysis and useful representation of TECS using the abstraction hierarchy was not straightforward. It involved a search for the interpretation of the levels of the abstraction hierarchy and the use of the means–ends relationship in conjunction with the aggregation relationship. The resulting WDA showed that the abstraction hierarchy could be used to map out the reasons for TECS’s design features. Many constraints were represented in the same space, which cluttered the energy management principles. The focus was put on the energy management principles through selective aggregation of the represented functions, but other design principles were omitted. To provide a complete representation of the system but without the clutter, the levels of control sophistication were introduced to represented nested control problems separately. At each level of control sophistication the abstraction hierarchy was applied, resulting in the Abstraction–Sophistication Analysis (ASA). In the third case study, the ASA framework was used to guide the design of SmartUAV. SmartUAV is a newly designed mini–UAV system that is capable of controlling multiple small UAVs from a laptop computer. By designing and developing SmartUAV we gained hands–on experience with how WDA, and especially the ASA, helped to keep track of and deal with the automation introduced constraints in the design phase. The levels of control sophistication were used from the beginning to separate the different control problems in the domain. They ranged from flying the platform to the achievement of missions. Starting at the lowest level of control sophistication, each higher level allowed the designer to include a larger part of the complete work domain incrementally, and to focus on more sophisticated control of the UAV. Furthermore, the ASA supported the visualization of how automation transformed the work domain, thus how automated functionalities that were created at lower levels of control sophistication affected the (automated) functions at higher levels of control sophistication. This study showed that the ASA could span a much larger problem space than the original WDA through the nesting of abstraction hierarchies. The ASA provided a systematic way to address the abstraction of the control problems (levels of control sophistication) and the abstraction of functions per control problem (abstraction hierarchy). The fourth case study dealt with the analysis of a subset of a well structured domain that lacks automation; sailboat racing. This study generated a clearer view on the nested structure that is inherent in a work domain, as apposed to the nested structure of the automation as found in TECS and SmartUAV. The nested structure inherent to this work domain was found to be the result of how sailboat racing has evolved over time, based on the capabilities of equipment, human performance and the racing rules. Due to the lack of automation, it became clear that human performance is in fact part of the work domain, in contrast to the original formulations of WDA. The crew’s performance formed the basis for achieving the more sophisticated control of boat speed, tactics and strategy, thus was essential in the analysis. It was shown that the performance of the human crew could be represented in the ASA at a level of control sophistication, while this could not be supported in a non–nested WDA based on a single abstraction hierarchy. The four case studies exemplified WDA and led to its extension with a structure to explicitly nest abstraction hierarchies that map out different control problems: the ASA. Through generating the analyses, extensive modeling experience with the abstraction hierarchy was obtained, reducing its ambiguity and potential methodological and conceptual problems. We found that the abstraction hierarchy could be used to model the structure of the knowledge about a work domain but could not model the knowledge itself. Therefore, the abstraction hierarchy is a framework for structuring knowledge, linking different representations of a control problem, and explaining the reasons for design features of a system. The abstraction hierarchy addressed the abstraction of elements belonging to a control problem, and the levels of control sophistication addressed the abstraction of the control problem itself. Representations in the ASA framework ranged from physical at the lower levels of control sophistication to non–physical at the higher levels of control sophistication. It allowed the structuring of, for example: the sailboat racing rules at the higher levels, and the law of conservation of energy at the lower levels. Although the application of the ASA did not inherently reduce the complexity of the design of SmartUAV, it enabled us to better understand the elements of the work domain that contribute to complexity of the system prior to and during its design. The extension of work domain analysis with the levels of control sophistication has led to a richer representation of the studied work domains than a single abstraction hierarchy or the abstraction–decomposition space.

Flight simulators provide a flexible, efficient, and safe environment for research and training at much lower costs than real flight. The ultimate validity of any simulation would be achieved when – for a particular task – human cognitive and psychomotor behavior in the simulator corresponds precisely to the behavior in the aircraft being simulated. However, it has been shown that for skill-based aircraft control tasks, pilot performance and control behavior are significantly affected by simulator motion cueing settings. Current technology centered fidelity metrics do not reflect to what extent a simulator is able to induce real flight pilot behavior, as they do not incorporate knowledge about human perception and control processes. This warrants the development of a new fidelity metric that determines the simulator’s ability to induce real-flight pilot control behavior. At the Faculty of Aerospace Engineering of Delft University of Technology a research project is dedicated to develop such a behavioral fidelity metric using a cybernetic approach. A prerequisite for developing this fidelity metric is to know exactly how pilot control behavior is affected by the limited physical motion stimuli that are typically provided in a simulator. Furthermore, the knowledge on pilot control behavior in a real aircraft – the baseline – needs to be greatly expanded. The project consists of five steps. In the first step of the project, new identification techniques and methodologies are developed to accurately identify multimodal control behavior. In the second step, pilot control behavior is determined in real flight. The third step constitutes the identification of control behavior in the simulator under an array of different motion cueing conditions. This allows for a systematic comparison of control behavior between real and simulated flight. In the fourth step of the project, this knowledge is used to trace behavioral discrepancies back to the way motion stimuli are presented in the simulator, and improve motion cueing algorithms to increase simulator behavioral fidelity. In the fifth and final step, standards and metrics for behavioral fidelity are developed. This thesis work covers the first three steps of the project. The final goal of this thesis is then to determine how pilot control behavior is affected by the limited motion cues provided in a simulator by comparing control behavior in the simulator under different motion cueing conditions to control behavior in the aircraft. The research will be limited to the effect of different motion cueing settings in a pitch control task. Steps four and five of the project will be covered in another thesis by ir. D.M. Pool. Using a cybernetic approach, pilot control behavior can be characterized by estimating the parameters of quasi-linear pilot models. In previous studies, this approach was used to compare pilot control behavior between real and simulated flight. However, in all these studies only a single, lumped, pilot response function was identified, without distinguishing between the contributions of different perceptual modalities, for example, visual and vestibular. In a multi-sensory environment, such as a motion-base flight simulator, this may obscure the pilot’s ability to adopt a different control strategy by a different use of perceptual modalities. Therefore, to compare control behavior between real and simulated flight adequately, multimodal pilot models need to be identified that are able to model the pilot’s use of modalities separately. The identification of these models requires a combined target-following disturbance-rejection task, as multiple forcing functions need to be inserted at different locations in the control loop, to allow for accurate estimation of the model parameters. At the start of this project, multimodal pilot control behavior had never been identified in real flight. The requirement for a combined target-following disturbance-rejection task complicates the setup of the in-flight experiments significantly. In the aircraft, the introduction of a single target tracking signal is relatively straightforward, as it can be visualized on a display in the cockpit. In-flight disturbance-rejection tasks are much more difficult to perform however, as introducing a deterministic disturbance on the stick-free aircraft motion requires an additional input other than the pilot’s control actions to be sent through the flight control system of the aircraft. To facilitate the experiments for the identification of in-flight multimodal pilot control behavior as part of this thesis work, a novel fly-by-wire system was developed for the Cessna Citation II laboratory aircraft of the Delft University of Technology. This fly-by-wire system allows for the disturbance of the aircraft motion by adding a disturbance signal to the pilot control signal. The system is novel in its design due to the use of the existing electric automatic control system in the aircraft, limiting the modifications to the aircraft. As the variations in pilot model parameters between different experiment conditions are often subtle and the aircraft measurements are relatively noisy, the model parameters should be estimated with the highest accuracy. Traditional two-step parameter estimation methods – used in many previous experiments – often produce less accurate results, as the inaccuracies from the frequency response identification step propagate to the parameter estimation step. To increase the accuracy of the multimodal pilot model parameter estimates, a new parameter estimation technique for multimodal pilot models was developed and addressed in this thesis in Chapter 2. The technique is based on the well-known concept of maximum likelihood estimation. Due to the relatively high levels of remnant noise in experiment data and the presence of redundant parameters in the pilot model, the maximum likelihood parameter optimization problem is very complex with many local minima. To increase the likelihood that the global optimum solution of the parameter vector is found, a genetic algorithm is combined with a more common gradient-based Gauss-Newton algorithm. The advantage of the new identification technique is that it operates solely in the time domain, increasing the accuracy of the parameter estimates compared to the two-step methods traditionally used in this type of research. Using the new identification technique, the parameters of a multimodal pilot model are estimated most accurately when as many forcing functions are inserted into the closed-loop control task at different locations, as the number of modalities to be identified. However, the power requirements of the different forcing functions were not known. In addition, little was known about how pilot control behavior is affected by using multiple forcing functions in a control task as opposed to a single forcing function. The first experiment in this thesis, discussed in Chapter 3, was therefore performed to investigate these two unknowns. The results of the experiment showed that multimodal pilot control behavior is significantly affected by the relative power settings of the target and disturbance forcing functions. When the power of the target forcing function is increased – simultaneously reducing the power of the disturbance forcing function – the cue conflict between the visual and physical motion cues increases, as the target forcing function – as opposed to the disturbance forcing function – is only presented visually. This causes pilots to control with a lower visual gain, while the visual perception time delay becomes higher. In addition, pilots reduce their visual lead and increase their vestibular gain when the power of both forcing functions becomes similar. The result of this change in control strategy is a reduction in tracking performance and control activity. It was found that multimodal pilot control behavior can be evaluated by using a combined target-following disturbance-rejection task with an additional signal with relatively small magnitude. In contrast to frequency-domain identification methods, the maximum likelihood based parameter estimation method has no strict requirement for the use of multi-sine signals as forcing functions in a closed-loop control task. This allows for an exploration into new types of signals to be used in manual control experiments. New types of target signals allow for manual control tasks that are more comparable to real piloting tasks. For example, roll ramp or step target signals introduce a task that is similar to flying a turn maneuver. The second experiment in this thesis (Chapter 4) was developed to investigate the identifiability of multimodal pilot control behavior using ramp and step target signals. In addition, the effect of these signals on pilot performance and control behavior itself was investigated. The experiment revealed that, in terms of performance, a task with ramp target inputs is comparable to a task without a target input. The step target inputs introduce a large increase in error variance, due to the transient behavior at the location of the steps. The step target inputs also result in significantly different response functions of the modalities of the pilot compared to the multi-sine and ramp target inputs, which induce comparable response functions. Based on the findings of the experiment, ramp signals as target forcing function are the best alternative to multi-sine target signals in keeping the ability to separate the pilot response functions for different modalities, while creating a task that is more equivalent to an actual piloting task. The results of the first two experiments were used to optimize the experiments in the remainder of the thesis. Before the actual in-flight and simulator comparison experiments were performed, several preliminary studies were undertaken to get insight into how pilot control behavior is affected by the different motion components that make up the total aircraft motion. In a pitch control task, the total aircraft motion at the pilot station can be decomposed into pitch rotational motion, pitch heave motion, and center of gravity heave motion. Pitch heave is the linear acceleration induced by the pitch rotation of the aircraft and the relative position of the pilot station in front of the center of gravity. Center of gravity heave results from relatively slow changes in aerodynamic lift due to the change in aircraft angle of attack while pitching. In conventional hexapod simulators, the center of gravity heave component is the most problematic to simulate accurately, as its low-frequency high-amplitude characteristics drive the simulator motion system to its limits. Therefore, in most simulator applications, the linear accelerations are heavily attenuated by a motion filter. By increasing the knowledge on how the different motion components are used by the pilot to form a control action, the individual components could be filtered more efficiently to increase behavioral fidelity in future simulator applications. The third experiment, discussed in Chapter 5 of this thesis, was set up to increase this knowledge on how pitch rotational motion, pitch heave motion, and center of gravity heave motion are used by a pilot performing a pitch control task. The results of the experiment indicated that – in a pitch target-following disturbance-rejection task – pitch motion significantly improved tracking performance, with an increased cross-over frequency for the disturbance open-loop. The increase in performance is a result of an increased visual gain and a reduction in visual lead, resulting in a lower effective time delay for the disturbance open-loop. For the Cessna Citation dynamics used, pitch heave motion showed effects similar to pitch rotational motion, but less strongly in part due to the relative short distance of the pilot station to the center of gravity and the motion filter that was used in the experiment. The presence of the center of gravity heave motion cue was found to have no significant effect on performance, however, visual lead significantly increased. This indicates that pilots reduce the use of motion cues in exchange for visual cues in the presence of center of gravity heave motion. A follow-up study focused on the effects of the filtering of pitch heave motion on pilot control behavior. The fourth experiment described in this thesis (Chapter 7) was the first experiment to identify multimodal pilot control behavior – separating the pilot’s visual and vestibular responses – in real flight. The experiment was performed with the new fly-by-wire system in the Cessna Citation II laboratory aircraft and was designed to gain more insight into the control system limitations and the optimal use of the system in future in-flight experiments on the identification of multimodal pilot control behavior. The required deterministic disturbance of the aircraft motion was facilitated by adding a disturbance forcing function to the fly-by-wire control signal. Accurate pilot model parameter estimation results could be obtained using multi-sine and ramp target signals in a pitch and roll task, despite some limiting of the fly-by-wire control signals. The limiting of control signals introduces nonlinearities in the closed control loop. However, the achieved accuracy and the multimodal pilot model parameter values were comparable to estimates from previous flight simulator experiments. These results allowed for the optimization of the final experiment discussed in his thesis in which multimodal pilot control behavior between real and simulated flight was compared. In the final experiment, discussed in Chapter 8, pilots performed a pitch target-following disturbance-rejection task in a simulator under different motion cueing settings, in addition to performing the task in an aircraft in flight, the baseline condition. Except for the applied variation in motion fidelity, differences in experimental setup between the aircraft and the simulator were kept as small as possible. Pilot performance and control behavior were slightly affected by differences in the display and sidestick setup. However, the effects introduced by motion fidelity were far more apparent. For the pitch target-following disturbance-rejection task performed, pilot performance and multimodal pilot control behavior were significantly affected by simulator motion fidelity. For improved motion fidelity towards the full-motion condition in the aircraft, pilot disturbance rejection improved. For higher levels of motion fidelity, the visual lead and lag time constants decreased, while visual and vestibular time delays increased. The lead and lag time constants approximate the characteristic time constants of the controlled aircraft dynamics much better when physical motion is present in the simulator and in flight, revealing the importance of simulator motion for these skill-based tracking tasks. From the limited number of motion conditions tested in this thesis, multimodal pilot control behavior in the simulator motion condition with full pitch motion and filtered pitch and c.g. heave motion best approximates in-flight pilot control behavior. The cybernetic approach proved to be a valuable concept in assessing simulator motion fidelity. Distinct variations in multimodal pilot model parameters were found between conditions with different motion fidelity showing the pilots’ adaption to the supplied motion cues, while the pilots rated these conditions the same on a motion fidelity rating scale. As a recommendation for future work, to investigate the effect of reduced motion fidelity more accurately, experiments should be performed on a single apparatus capable of large motion displacement that allows for simulation of full aircraft motion. This would eliminate the effects of differences in experimental setup experienced in this thesis work. Future research should also be more geared towards simulator fidelity as a whole, as the reduced fidelity of simulator systems other then the motion system (for example, out of the window visual systems) have also shown to influence pilot control behavior. Finally, more research should be devoted to identification techniques that are capable of separating more perceptual modalities and techniques for modeling pilot control behavior in tasks that are more comparable to real piloting tasks.

The thesis considers the trajectories that are used to guide aircraft from en-route flight to an airport. Although there is currently some room to adapt the trajectories to a specific situation, they are usually largely based on published standardised trajectories. In this study the potential of optimising the arrival trajectories using data about the current situation is investigated. A trajectory scheduling algorithm based on a multi-objective genetic optimisation algorithm is developed. Fast-time simulations are carried out for traffic arriving at Amsterdam Airport Schiphol. It is shown that optimising the arrival trajectories can increase throughput and has the potential of reducing impact on the environment significantly. Optimising the arrival trajectories may have implications for the pilots and air traffic controllers tasks though. It is studied by means of experiments in a flight simulator how the shape of the trajectory influences the task demand load imposed on the pilot. A number of metrics are proposed to describe the task demand load. Off-line analyses using these metrics indicate that task demand load may increase when optimised arrival trajectories are used instead of standard trajectories. It is shown though that recently proposed displays for the flight deck that give 4-D guidance information may help to reduce the task demand load. In addition, the task demand load imposed on air traffic controllers is assessed with metrics describing airspace complexity. Off-line analyses indicate that task demand load for air traffic controllers may increase in case conventional air traffic control is maintained.

The objective of the research was to identify and evaluate promising mathematical techniques for re-entry flight clearance. To fulfil this objective, two mathematical methods were investigated and developed: μ analysis for linear models and interval analysis for both linear and non-linear models. The stability of re-entry vehicles in the presence of model uncertainties was chosen as the clearance criterion, which is represented by two mathematical criteria: worst-case eigenvalues (linear) and the Lyapunov stability (non-linear). Two vehicle models including flight control systems were developed and used as case studies for the evaluation of the clearance techniques. These models are based on the DART (Delft Aerospace Re-entry Test Demonstrator) and SPHYNX (Subscale Precursor Hypersonic X) re-entry vehicle models. The suitability of the two mathematical techniques for re-entry flight clearance was evaluated based on the results of the clearance application on these models. Non-linear simulations were also performed to verify the clearance results generated by the two techniques. Non-linear interval analysis has been found to be the most reliable method of all other methods investigated in this research, because it could perform the clearance for the non-linear dynamic models of the re-entry vehicles with uncertainties, and the results were confirmed by the non-linear simulations.

Like most other types of instrumentation systems, flight test instrumentation is not produced in series; its development is a one-time achievement by a test department. With the introduction of powerful digital computers, instrumentation systems have included data analysis tasks that were previously limited to post-experiment processing. However, the resulting integrated systems are hard to maintain in the traditional environment of instrumentation development. Software-Enabled Modular Instrumentation Systems describes the theory and praxis of a new methodology to analyze, design, implement, and validate a digital signal processing system for test and evaluation applications in the information age. Based on life cycle concepts from software engineering, this dissertation presents an object-oriented approach that allows to combine proprietary and off-the-shelf components in a way that reusability of the elements and extensibility of the application are ensured. The methodology covers all phases of test and evaluation: desktop simulation, hardware- and pilot-in-the-loop simulation, flight test, and post-experiment data analysis. Moreover, optimum reusability of the components is ensured not only through the phases of the project, but also from one test program to the next. The methodology thus results in reduced system development time and cost, and improved system reliability.

The response of aircraft to stochastic atmospheric turbulence plays an important role in aircraft-design (load calculations), Flight Control System (FCS) design and flight-simulation (handling qualities research and pilot training). In order to simulate these aircraft responses, an accurate mathematical model is required. Two classical models will be discussed in this thesis, that is the Delft University of Technology (DUT) model and the Four Point Aircraft (FPA) model. Although they are well estabilished, their fidelity remains obscure. The cause lies in one of the requirements for system identification; it has always been necessary to relate inputs to outputs to determine, or identify, system dynamic characteristics. From experiments, using both the measured input and the measured output, a mathematical model of any system can be obtained. When considering an input-output system such as an aircraft subjected to stochastic atmospheric turbulence, a major problem emerges. During flighttests, no practical difficulty arises measuring the aircraft motion (the output), such as the angle-of-attack, the pitch-angle, the roll-angle, etc.. However, a huge problem arises when the input to the aircraft-system is considered; this input is stochastic atmospheric turbulence in this thesis. Currently, during flighttests it still remains extremely difficult to identify the entire flowfield around an aircraft geometry subjected to a turbulent field of flow; an infinite amount of sensors would be required to identify the atmospheric turbulence velocity component's distribution (the input) over the vehicle geometry. In an attempt to shed some more light on solving the problem of the response of aircraft to atmospheric turbulence, the subject of this thesis, it depends on the formulation of two distinct models: one of the atmospheric turbulence itself (the atmospheric turbulence model), and the other of the aircraft response to it (the mathematical aircraft model). As concerns atmospheric turbulence, stochastic, stationary, homogeneous, isotropic atmospheric turbulence is considered in this thesis as input to the aircraft model. Models of atmospheric turbulence are well established. As for mathematical aircraft models, many of them have been proposed before. However, verifying these models has always been extremely difficult due to the identification problem indicated above. As part of the mathematical aircraft model, (parametric) aerodynamic models often make use of (quasi-) steady aerodynamic results, that is all steady aerodynamic parameters are estimated using either results obtained from windtunnel experiments, handbook methods, Computational Aerodynamics (CA) which comprises Linearized Potential Flow (LPF) methods, or Computational Fluid Dynamics (CFD) which comprises Full-Potential, Euler and Navier-Stokes methods. In this thesis the simplest form of fluid-flow modeling is used to calculate the time-dependent aerodynamic forces and moments acting on a vehicle: that is unsteady Linearized Potential Flow (LPF). The fluid-flow model will result in a so called "unsteady panel-method" which will be used as a virtual windtunnel (or virtual flighttest facility) for the example discretized aircraft geometry, also referred to as the "aircraft grid". The application of the method ultimately results in the vehicle's steady and unsteady stability derivatives using harmonic analysis. Similarly, both the steady and unsteady gust derivatives for isolated atmospheric turbulence fields will be calculated. The gust fields will be limited to one-dimensional (1D) longitudinal, lateral and vertical gust fields, as well as two-dimensional (2D) longitudinal and vertical gust fields. The harmonic analysis results in frequency-dependent stability- and gust derivatives which will later be used to obtain an aerodynamic model in terms of constant stability- and gust derivatives. This newly introduced model, the Parametric Computational Aerodynamics (PCA) model, will be compared to the two classical models mentioned earlier, that is the Delft University of Technology (DUT) model and the Four-Point-Aircraft (FPA) model. These three parametric aircraft models are used to calculate both the time- and frequency-domain aerodynamic model and aircraft motion responses to the atmospheric turbulence fields indicated earlier. Also, using the unsteady panel-method the aircraft grid will be flown through spatial-domain 2D stochastic gust fields, resulting in Linearized Potential Flow solutions. Results will be compared to the ones obtained for the parametric models, i.e. the PCA-, DUT- and FPA-model. From the results presented, it is concluded that the introduced PCA-model is the most accurate for all considered gust fields. Compared to the Linearized Potential Flow solution (which is assumed to be the benchmark, or the model that approximates reality closest) the new parametric model shows increased accuracy over the classical parametric models (the DUT- and FPA-model), especially for the aircraft responses to 2D gust fields. Furthermore, it shows more accuracy in the aircraft responses to 1D longitudinal gust fields. Although results will be presented for a Cessna Ce550 Citation II aircraft only, the theory and methods are applicable to a wide variety of fixed-wing aircraft, that is from the smallest UAV to the largest aircraft (such as the Boeing B747 and the Airbus A380).

Simulation fidelity is an intrinsic element of any simulation system, one that all its developers and users have to deal with one way or the other. It is commonly recognized by the modeling and simulation community that simulation fidelity is an essential vehicle in properly assessing the validity and credibility of simulation results. Furthermore, fidelity is one of the main cost-drives of any model or simulation development. Rigorous assessment of fidelity is, however, one of the most difficult and hard to grasp issues within the model and simulation community. Substantial and exhaustive research endeavors in this area are very limited. Due to this, simulation fidelity still remains a hardly touched upon and rather uncultivated area. This thesis tries to fill this void by the analysis, extension and integration of existing simulation fidelity approaches into a single unified fidelity theory and practice. All this is done from a general simulation system life cycle perspective, not limited by any specific application or problem domain aspects. The foundation for this developed unified fidelity framework comprises a precise mathematical formulation for fidelity and the fundamental concepts underlying its characterization and measurement. The unified fidelity framework is completed with a fidelity management process model outlining a series of generic stages, activities and tasks, which together provide a structured but generic approach to properly integrate and apply all other unified fidelity framework elements in the simulation system development and validation process.

The upgrade of the flight deck instruments from electro-mechanical dials and gauges towards computer-driven systems and interfaces was a necessary step to accommodate the increasing demands in flight technical performance and safety. The upgrade was a relatively slow process, however, where new systems were developed and installed as soon as the technology was available. As a result, many systems are not always well integrated in terms of presenting information. Together with the increasing amount of automation, the flight deck has become prone to issues such as information ambiguities and misunderstandings between the pilot and the (automated) avionic systems. This phenomenon is commonly labeled as a lack of "situation awareness" (SA) and has become a new cause for accidents. That is, pilots are unaware sometimes of the current flight situation, a situation that in itself may be caused by the automation. A recent example of this phenomenon is the Turkish Airlines accident near Schiphol on February 25 2009. The focus of this thesis is on aircraft terrain avionics, such as the Terrain Awareness Warning System (TAWS) and the Synthetic Vision System (SVS), that form a typical illustration of the evolution process and its issue regarding SA. The work in this thesis aimed to identify and address the missing information that would span the information gaps between the SVS and the TAWS to benefit pilot SA. The Ecological Interface Design (EID) framework was explored to accomplish this goal. EID was originally developed for the process industry (like nuclear power plants) and is therefore a rather novel approach in the field of flight deck design. The results of experimental evaluations indicated that pilots managed to benefit from the ecological interface enhancements to successfully and safely deal with hazardous terrain conflicts, even when encountering unanticipated events. It was found that the ecological interfaces made pilots more aware of the aircraft capabilities and much more actively involved in the decision-making loop to prevent fatal mishaps.

The main goal of this thesis is to investigate the potential of the nonlinear adaptive backstepping control technique in combination with online model identification for the design of a reconfigurable flight control system for a modern fighter aircraft. Adaptive backstepping is a recursive, Lyapunov-based, nonlinear design method, that makes use of dynamic parameter update laws to deal with parametric uncertainties. The idea of backstepping is to design a controller recursively by considering some of the state variables as ‘virtual controls’ and designing intermediate control laws for these. Backstepping achieves the goals of global asymptotic stabilization of the closed-loop states and tracking. The proof of these properties is a direct consequence of the recursive procedure, since a Lyapunov function is constructed for the entire system including the parameter estimates. The tracking errors drive the adaptation process of the procedure. Furthermore, it is possible to take magnitude and rate constraints on the control inputs and system states into account in such a way that the identification process is not corrupted during periods of control effector saturation. A disadvantage of the integrated adaptive backstepping method is that it only yields pseudo-estimates of the uncertain system parameters. There is no guarantee that the real values of the parameters are found, since the adaptation only tries to satisfy a total system stability criterion, i.e. the Lyapunov function. Increasing the adaptation gain will not necessarily improve the response of the closed-loop system, due to the strong coupling between the controller and the estimator dynamics. The immersion and invariance (I&I) approach provides an alternative way of constructing a nonlinear estimator. This approach allows for prescribed stable dynamics to be assigned to the parameter estimation error. The resulting estimator is combined with a backstepping controller to form a modular adaptive control scheme. The I&I based estimator is fast enough to capture the potential faster-than-linear growth of nonlinear systems. The resulting modular scheme is much easier to tune than the ones resulting from the standard adaptive backstepping approacheswith tracking error driven adaptation process. In fact, the closed-loop system resulting from the application of the I&I based adaptive backstepping controller can be seen as a cascaded interconnection between two stable systems with prescribed asymptotic properties. As a result, the performance of the closed-loop system with adaptive controller can be improved significantly. To make a real-time implementation of the adaptive controllers feasible the computational complexity has to be kept at a minimum. As a solution, a flight envelope partitioning method is proposed to capture the globally valid aerodynamic model into multiple locally valid aerodynamic models. The estimator only has to update a few local models at each time step, thereby decreasing the computational load of the algorithm. An additional advantage of using multiple, local models is that information of the models that are not updated at a certain time step is retained, thereby giving the approximator memory capabilities. B-spline networks are selected for their nice numerical properties to ensure smooth transitions between the different regions. The adaptive backstepping flight controllers developed in this thesis have been evaluated in numerical simulations on a high-fidelity F-16 dynamicmodel involving several control problems. The adaptive designs have been compared with the gain-scheduled baseline flight control system and a non-adaptive NDI design. The performance has been compared in simulation scenarios at several flight conditions with the aircraft model suffering from actuator failures, longitudinal center of gravity shifts and changes in aerodynamic coefficients. All numerical simulations can be easily performed in real-time on an ordinary desktop computer. Results of the simulations demonstrate that the adaptive flight controllers provide a significant performance improvement over the non-adaptive NDI design for the simulated failure cases. Of the evaluated adaptive flight controllers, the I&I based modular adaptive backstepping design has the overall best performance and is also easiest to tune, at the cost of a small increase in computational load and design complexity when compared to integrated adaptive backstepping control designs. Moreover, the flight controllers designed with the I&I based modular adaptive backstepping approach have even stronger provable stability and convergence properties than the integrated adaptive backstepping flight controllers, while at the same time achieving a modularity in the design of the controller and identifier. On the basis of the research performed in this thesis, it can be concluded that a RFC system based on the I&I based modular adaptive backstepping method shows a lot of potential, since it possesses all the features aimed at in the thesis goal.

Safety is of paramount importance in all transportation systems, but especially in civil aviation. Therefore, in civil aviation, a lot of developments focus on the improvement of safety levels and reducing the risks that critical failures occur. When one analyses recent aircraft accident statistics, it is clear that a significant portion is attributed to “loss of control in flight”. A recent worldwide civil aviation accident survey for the 1989 to 2003 period, conducted by the Civil Aviation Authority of the Netherlands (CAA-NL) and based on data from the National Aerospace Laboratory NLR, indicates that this category accounts for as much as 17% of all aircraft accident cases. This has led to a common conclusion: from a flight dynamics point of view, with the technology and computing power available on this moment, it might have been possible to recover a part of the aircraft in the accident category described above on the condition that non-conventional control strategies would have been applied. These non-conventional control strategies involve the so-called concept of fault tolerant flight control (FTFC), where the control system is capable to detect and adapt for changes in the aircraft behaviour. One FTFC strategy option is using a model based control routine. This research focuses on a physical modular approach. In this setup, not only a reconfiguring controller is needed, but also a suitable FDI/identification strategy. This research focuses on both components. In this reseach project, a real-time aerodynamic model identification procedure has been combined with a model based adaptive control method. A manual as well as an autopilot version have been developed. The autopilot version has been evaluated on desktop simulations, the manual version has been tested in the Simona Research Simulator involving professional airline pilots. Both tests have demonstrated promising results. The autopilot performance is very good, and the manual controller has demonstrated to increase handling qualities and to reduce pilot workload of the damaged aircraft. These are very promising results that motivate further research in this field.

There exist many examples of aircraft incidents in which the pilots have successfully used the remaining control authority over an aircraft to save the airframe and its passengers and cargo from apparently hopeless failure conditions. Unfortunately, the opposite is also true. Several accidents happened in which the crew was not able to save the aircraft, although post-flight analysis showed that it was possible with alternative, perhaps unconventional, control strategies. These aircraft accidents indicate that there is a potential benefit of fault tolerant flight control techniques, which are able to accommodate changes in the aircraft’s dynamics due to damage to the aircraft and failures of its systems. In this dissertation a modular adaptive flight control approach was developed based on adaptive backstepping with a recursive least squares estimator. The proposed control design was evaluated in numerical simulations on high-fidelity fighter aircraft models. The performance has been compared in simulation scenarios at several flight conditions with the aircraft model suffering from actuator failures, longitudinal center of gravity shifts and changes in aerodynamic coefficients. Results of the simulations demonstrate that the adaptive flight controller provides a significant performance improvement over classical, non-adaptive flight control designs. Although adaptive flight control techniques have shown that it may be possible to stabilize a damaged aircraft for a variety of faults and failures, it is still unclear what maneuvers are still possible and how much the performance of the aircraft has degraded due to these faults and failures. The safe flight envelope is defined as the region in the state space for which safe operation of the aircraft, and safety of its cargo and passengers can be guaranteed. In this dissertation the level set method was researched to determine the safe region of operation of the aircraft. Application of this method to an F-16 model at different flight conditions showed shrinking of the safe flight envelope and decreased maneuverability with decreasing dynamic pressure.