THIS PAGE DESCRIBES A FINISHED PROJECT AND IS NOT UPDATED ANYMORE
NECTAR
Neural and Expert Control of Aircraft
Research into the application of Neural Networks for Flight Control
Abstract
Processes with (partly) unknown or complex dynamic behaviours need complex control schemes. Neural networks offer interesting perspectives both for identification and control of these processes, because neural networks approach any (non-linear) continue function. Especially adaptive control using neural networks offers good possibilities, due to the possibility to learn on-line. The NECTAR-project (Neural and Expert ConTrol of AircRaft) aims at studying both neural network and expert system structures for highly demanding control environments. In this publication the general structure of the project will be sketched, after which the neural network part of the project will be discussed in more detail. The final architecture will be implemented on-board of a laboratory aircraft to demonstrate the possibilities.
Keywords
Flight Control, Intelligent Control, Real Time Expert Systems, Neural Networks.
Synopsis
Control has come a long way and especially in linear control, impressive breakthroughs have been realised. Linear control has been proven very usefull and viewing things in perspective, it is fair to say that control of processes is very mature at the moment. However, as most processes are nonlinear, rather than linear, and frequently not all information of a process is available, leading to partially unknown process dynamics, there is a limit to the applicability of linear control engineering. Adaptive control and robust control have proven to be usefull in those circumstances, but the advance of neural networks over the last decades has made a study to the possibilities of neural networks for control very interesting. In literature many applications have been described, but these were always limited in size and complexity. Still, the applicability of neural networks for this kind of problems has been demonstrated to a certain level. Main properties of neural networks that can be considered as favourable for application in process control are their generalising character and their ability to learn from examples, rather than to be pre-programmed with a certain process model.
The advance of digital flight control now allows to try out different control schemes, including knowledge based control and neural control. In this project, the applicability of neural networks for a very demanding environment, a civil aircraft, is being studied. Several asapects make the application in flight control a challenging one. First of all, the airworthiness authorities have to give permission to use such advanced control systems. This entails certifying the control system, during which process the safety has to be proved. Up until this moment not enough is known of neural controllers to be able to prove the working of the controller in a mathematical manner. The stability of the controller can not be guaranteed and therefore the airworthiness authorities will be reluctant to certify such controllers. This project aims at bringing more light in these matters, by identifying a stability area in which the controllers can be regarded as stable, leaving other areas to be controlled using different means. From the airlines point of view especially the acceptance by pilots may prove to be a problem. For passengers of these airlines this may prove even harder to accept that a neural system us flying the aircraft. Still, many flight modes qualify for application of newer technologies. These flight modes are characterised by highly nonlinear behaviour and often a not completely known model. Examples of these can be found in cross-wind landings, windshear situations, etc.
Neural networks have been in the spotlight of attention for a number of years. Especially for classification tasks, they offer good alternatives to other existing techniques and also the application of neural networks for control has been studied by many. Thereby the learning and generalising capacity is used intensively. For a certain class of neural networks (3-layer feed-forward networks) it can be proven that for a given non-linear continuous function can be approximated as closely as desired. This implies that any non-linear continuous transformation from an input space (state of the system) to an output space (the control signals) can be represented by a trained neural network. For this training only examples of working input-output relations have to be given to the network that then automatically captures the desired behaviour using a training algorithm. This means on one hand that no longer a mathematical model of the system to be controlled is necessary. The possibility to adapt the trained control law on-line to changing environments and to generalise from the learned training set to input the network had not seen earlier amount to the advantages of the application of neural networks for control. There are however also some drawbacks to this application. The lack of a mathematical model no longer allows control engineers to verify the working of the controller. The theory behind neural networks is not yet well-enough developed to allow implementation of neural networks for critical tasks. Their exact working is not yet well-enough understood. Also the number of different architectures described in literature is so vast that the choice for a particular architecture is not easy and the learning algorithm the be used has to be fine-tuned in detail as well. Some general rules-of-thumb can be found in literature, but usually a trial-and-error situation remains.
The NECTAR-project
At Delft University of Technology research on neural networks in general and for control applications in particular has lead to the NECTAR-project (Neural and Expert ConTrol of AircRaft). This project aims at studying both neural network and expert system structures for highly demanding control environments. The goal is to select interesting neural network architectures for control applications, verify their proper functioning and deliver usable guidelines for their use. These controllers can then be compared to other control algorithms to verify the true potential of these new techniques. The final demonstration of the proper functioning of neural networks for control applications will be done on-board of a laboratory aircraft.
Because of the high interests at stake aeronautics, a lot of research has been carried out into the many different systems that are important for the control of an aircraft. The problems that have arisen out of this research, can be divided in two categories:
- It is impossible to obtain the (mathematical) model of a certain system.
- Though it is possible to find the (mathematical) model of a system, it is impossible to identify all the parameters and the relations between them. In such a case, the model is linearised and reduced to the most important parameters.
Since the early 70's, research to flight control techniques has focussed on digital techniques. The control signals are sent to the control surfaces in electrical / digital format, where hydraulic servo-systems effectuated the control action. At Delft University of Technology, this research used to be performed using the laboratory aircraft `DeHavilland DHC-2 ``Beaver'''. This aircraft, built in the 1950's and over the years completely modernised to a completely digitally controllable aircraft, was recently replaced by a new laboratory aircraft, jointly owned by the university and the National Aerospace Laboratory NLR. This aircraft, a Cessna Citation II business jet, is being used in the national project the National Fly-by-wire Testbed (NFT). For research to flight control systems this aircraft opens up many new possibilities in that almost any control law can be tested in full flight. In this respect more conventional control laws (PID, adapative cotnrol schemes, ...) as well as newly developed control schemes (knowledge based control, fuzzy control, neural control, ...). Linked therefore to this project, is the Active Flight Controls project (AFC), aiming at studying modern control laws for flight control. The NECTAR-project is closely linked to the framework of the NFT/AFC projects and fits within the working group "Intelligent Flight Control".
Application of neural networks for flight control
Aircraft allow up to a certain extend the development of a non-linear model. From the non-linear models, linearisations can be made around certain operating points. Designing good control law based on these simplyfied models is not easy and in extreme situations the designed control laws will not fullfill. Even more, during complex manoeuvres, such as landings and take-offs, the dynamics of the aircraft are changing rapidly within a limited time-frame. For these situations more conventional controllers are not accurately enough. Future developments such as the Microwave Landing System (MLS) emphasize on these developments. The presence of model uncertainties and the influence of the atmosphere add to this situation. It are these circumstance that require alternative methods. Neural networks may offer a good alternative.
In first instance the landing phase of an aircraft is studied, during which phase an analysis takes place for selecting for proper neural network architectures. A big difference between this application and many in literature referenced applications is the complexity of the example at hand. The following main differences can be identified.
- At a height of about 10 meters the so-called flare manoeuvre is initiated. The nose of the aircraft is lifted, reducing the vertical speed of the aircraft and allowing the main gear to touch the ground first. During this limited time interval the control law has to be adjusted contiuously.
- When an aircraft is flying at a low altitude, the so-called ground-effect starts to play a major role. The aircraft dynamics are affected immensely due to this effect.
Implementation on-board of the aircraft
The final implementation of neural networks on-board of the NFT can be done in either one of two ways. The implementation can be realised using workstations, as the NFT has a set of special workstations on-board to allow quick data instpection during flight. A second option can be the use of special neural hardware. Special neural chips are available, allowing through an inherent parallel structure very fast data processing. A good example of such chips is the Intel 80170NX. Placing several of these chips in parallel creates a very powerfull computer. It is planned to try out such chips in full flight.
In the initial phase of this project, mainly simulations will be used. The networks will be trained using both flight data and simulated data, while confirmation of proper working must be done by test data. At a later stage, the flight simulator will be used for further verification. Pilots will fly with the simulator and will verify the proper working of the neural controller. After that stage, test flights will be performed. They will at that moment serve mainly as confirmation of already known behaviour of the controller.