University of Twente Student Theses


Modelling Pilot Pitch Control : analysis of human-in-the-loop dynamics from a control theory perspective

Tan, D.J. (2015) Modelling Pilot Pitch Control : analysis of human-in-the-loop dynamics from a control theory perspective.

[img] PDF
Abstract:This internship report discusses aspects of pilot modelling from a system identification perspective. The assignment is within the research project on the investigation of the safety of curved approaches at The University of Tokyo. Modelling dynamics of human pilot control could provide valuable insights on the pilot’s capacity and performance. The multidimensional human-aircraft system is complex and consists of feedback loops. It is therefore difficult to identify the entire human-in-the-loop system. Attempts have been made to capture pitch dynamics in a predictive mathematical model. In order to acquire control data, three types of experiments have been designed: 1) a theoretical setup to test identification techniques, 2) a computer simulation augmenting pitch dynamics of a B747-400, and 3) the setup in a full flight simulator. The procedure for the experiments is similar and the pilot is asked to track the motion of the Flight Director (FD), indicated on the Primary Flight Display (PFD). The artificial FD signal is designed using system identification theory and its frequency content is well within human bandwidth. The aircraft is trimmed to a stable and level flight and all other control inputs (e.g. ailerons, rudders, thrust) apart from the elevator have been left untouched. The data from the third experiment performed by an experienced pilot in a Dornier Do-228-200 full flight simulator has been used for analysis. All models have been evaluated on model fit, complexity and calculation time. It is shown that a linear parametric Auto Regressive with exogenous input (ARX) model is sufficient to model the closed-loop dynamics. For controller identification, the direct identification method is used. In this case, a linear ARX model only captures course control action, but lacks the capacity to predict corrective control. A non-linear ARX (NARX) neural network provides significant better results in terms of model fit, but requires a high number of parameters and is computational intensive. Furthermore, one should be cautious with providing (elevator) feedback channels to self-learning model sets (as the with the NARX) since it can yield erroneous results. Consequently, an Adaptive Neuro-Fuzzy Inference System (ANFIS) has been modelled relying on the FD input signal and actual aircraft pitch. This approach combines fuzzy logic with characteristics of neural networks and the control strategy can be interpreted conveniently due to its linguistic rules. Although the ANFIS doesn’t excel in model fit due to large occasional prediction outliers, it does capture human corrective control with only 4 fuzzy rules. It has been hypothesised that the level of experience can be evaluated by observing the closed-loop and human control models. Further research is needed to test the generality of the modelled human control dynamics.
Item Type:Internship Report (Master)
University of Tokyo, Japan
Faculty:ET: Engineering Technology
Subject:52 mechanical engineering
Programme:Mechanical Engineering MSc (60439)
Keywords:Pilot modelling, system identification, human-in-the-loop dynamics
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page