Nomenclature

$\textbf{A}$

state matrix

A–PC

aircraft–pilot coupling

$\textbf{B}$

input matrix

$\textbf{C}$

output matrix

$\textbf{c}$

input vector

$\textbf{D}$

feedforward matrix

${e(s)}$

error signal between the desired value and the actual aircraft

$g$

gravitational acceleration

$\mathrm{I}_{y}$

moment of inertia about the y-axis

$\mathrm{K}_{p}$

gain

LQR

linear quadratic regulator

$m$

mass of the aircraft

PIO

pilot-induced oscillation

$q$

pitch rate

ROVER

real-time oscillation verifier

$\mathrm{T}_{L}$

general lead time constant of human pilot modelling function

$\mathrm{T}_{I}$

general lag time constant of human pilot modelling function

$u(s)$

pilot control signal

$\Delta u$

longitudinal velocity variation

$w$

vertical velocity

$\textbf{x}$

state vector

$\mathrm{X}_{\cdot }$ , $\mathrm{I}_{\cdot }$ , $\mathrm{M}_{\cdot }$

stability and control longitudinal derivatives

$\textbf{y}$

linearised flight dynamics

$\mathrm{Y}_{c}$ (s)

transfer function for the aircraft behaviour

$\mathrm{Y}_{p}$ (s)

transfer function for human pilot action

Greek symbol

$\delta_{e}$

elevator angle

$\Delta\theta$

pitch angle variation

$\zeta$

damping coefficient

$\theta_{0}$

initial pitch angle

$\tau$ , $\tau_{e}$

time delay constant

$\omega_{c}$

crossover frequency

$\omega_{n}$

natural frequency

2.0 Modelling

In this manuscript, potential PIO conditions were assessed using the control arrangement shown in Fig. 1. It essentially consisted of a generic task to be performed (Syntask), a human pilot model, an aircraft model and a PIO detection algorithm (ROVER). Figure 1 also includes the elevator actuator and the elevator surface, representing the whole pitch flight control system. These are sources of delays and may affect the behaviour of the system in PIO [Reference Wang, Santone and Cao2,Reference Ashkenas3,Reference Moura, Bidinotto and Belo7–Reference Moura, Alegre, Bidinotto and Belo9], but in the current study, such delays were disregarded because they could compromise the information sought by the research.

Figure 1.

Diagram of the system used.

The study focused on pitch PIO, thus only the longitudinal motion of the aircraft was considered and the pitch angle was chosen as the parameter to be controlled. That said, the error signal e(s) shown in Fig. 1 accounted for the difference between the longitudinal reference attitude (the Syntask) and the actual pitch angle of the aircraft. Based on this signal, the pilot model actuated on the aircraft dynamics by deflecting the elevator (being the only possible input signal).

The behaviour of the human pilot is either represented by three different pilot models in the form of transfer functions proposed in the literature or by a real human, whereas the aircraft is modelled by the closed-loop configuration with a Linear Quadratic Regulator (LQR) controller. It is noteworthy that the actuators of the aircraft control surfaces have some limitations, thus their rate and position saturation were also modelled and incorporated into the vehicle dynamics model. The rate limit of the actuators was chosen to be $40^{\circ}$ /s, and the position of the control surfaces was limited to deflect $30^{\circ}$ for each side. Finally, the ROVER algorithm is utilised to identify in real time the occurrence of PIO [Reference Moura, Alegre, Bidinotto and Belo9].

The main blocks of the control arrangement of Fig. 1 are detailed in the remainder of this section.

2.1 Syntask

Syntask is a synthetic task consisting of a profile of attitudes (in our case, the pitch angle) to be carried out by either human or computational pilots. In this study, the Syntask was displayed on the artificial horizon by means of a movable red line (Fig. 2a). The line moved in accordance with the profile indicated in Fig. 2b. All simulations with the computational pilot models were performed using the same Syntask profile (Fig. 2b).

Figure 2.

Synthetic task system [Reference Moura8].

2.2 Aircraft model

A state-space approach according to Etkin and Reid (10) is selected to model the dynamics of the vehicle, in this case, a Boeing 747-100 in a cruise condition equivalent to an altitude of 40,000ft and Mach 0.8. Equation (1) describes the system via the state-space optics, in which the state vector $\textbf{x}$ represents the longitudinal velocity variation ( $\Delta$ u), vertical velocity (w), pitch rate (q) and pitch angle variation ( $\Delta\theta$ ). It can thus be seen that only the longitudinal behaviour is considered, as the research focuses on pitch-axis PIO. For the same reason, the input vector of Equation (1), that is, $\textbf{c}$ , stands for the elevator angle ( $\delta_{e}$ ), while the output vector $\textbf{y}$ in the present research coincides with the state vector. Therefore, the output matrix $\textbf{C}$ is a $4\times4$ identity matrix, whilst the feedthrough matrix $\textbf{D}$ is a $4\times1$ null matrix. The throttle is considered to be constant.

(1) \begin{eqnarray}\dot{\textbf{x}} = \textbf{Ax} + \textbf{Bc}\end{eqnarray}

\begin{eqnarray}\textbf{y} = \textbf{Cx} + \textbf{Dc},\nonumber\end{eqnarray}

which are presented as the Equations (2) and (3) when written in full mode

(2) \begin{eqnarray}\left\{\begin{array}{c} \Delta \dot{u}\\ \dot{w}\\ \dot{q} \\ \Delta \dot{\theta} \end{array}\right\} =\left(\begin{array}{cccc} \frac{X_{u}}{m} & \frac{X_{w}}{m} & 0 & -g\cos(\theta_{0}) \\ \frac{Z_{u}}{m - Z\dot{_{w}}} & \frac{Z_{w}}{m - Z\dot{_{w}}} & \frac{Z_{q} - m u_{0}}{m - Z\dot{_{w}}} & \frac{- m g \sin(\theta_{0})}{m - Z\dot{_{w}}} \\ \frac{1}{I_{y}}\left[M_{u} + \frac{M\dot{_{w}} Z_{u}}{\left(m - Z\dot{_{w}}\right)}\right] & \frac{1}{I_{y}}\left[M_{w} + \frac{M\dot{_{w}} Z_{u}}{\left(m - Z\dot{_{w}}\right)}\right] & \frac{1}{I_{y}}\left[M_{q} + \frac{M\dot{_{w}} \left(Z_{q} + m u_{0}\right)}{\left(m - Z\dot{_{w}}\right)}\right] & -\frac{M\dot{_{w}} m g \sin(\theta_{0})}{I_{y}\left(m - Z\dot{_{w}}\right)} \\ 0 & 0 & 1 & 0 \end{array}\right)\left\{\begin{array}{c} \Delta u\\ w\\ q \\ \Delta\theta \end{array}\right\}\nonumber \\+\left(\begin{array}{c} \frac{X_{\delta_{e}}}{m}\\ \frac{X_{\delta_{e}}}{m - Z\dot{_{w}}}\\ \frac{M_{\delta_{e}}}{I_{y}}+\frac{M\dot{_{w}} Z_{\delta_{e}}}{I_{y}\left(m - Z\dot{_{w}}\right)} \\ 0 \end{array}\right)\delta_{e},\end{eqnarray}

(3) \begin{equation}\textbf{y} =\left(\begin{array}{c@{\quad}c@{\quad}c@{\quad}c} 1 &0 &0 &0 \\ \\[-7pt] 0 &1 &0 &0 \\ \\[-7pt] 0 &0 &1 &0 \\ \\[-7pt] 0 &0 &0 &1 \end{array}\right)\left\{\begin{array}{c} \Delta u\\ \\[-7pt] w\\ \\[-7pt] q \\ \\[-7pt] \Delta\theta \end{array}\right\}+\left(\begin{array}{c} 0 \\ \\[-7pt] 0 \\ \\[-7pt] 0 \\ \\[-7pt] 0 \end{array}\right)\delta_{e}.\end{equation}

According to the assumption of decoupling between the longitudinal and lateral directional motions, Equations (2) and (3) apply to the motion relative to the axis in the wingspan direction (y). In these equations, m is the mass of the aircraft, g is the gravitational acceleration and $I_{y}$ is the moment of inertia about the y-axis; the other terms are the longitudinal stability derivatives of the aircraft, detailed in Table 1.

Table 1.

The stability and control derivative of the models [Reference Moura, Alegre, Bidinotto and Belo9,Reference Etkin and Reid10]

A thorough investigation is conducted in the sense that three different sets of stability derivatives are analysed, as presented in Table 1. The first, model A, represents the original parameters for the mentioned flight condition [Reference Etkin and Reid10], whilst models B and C [Reference Moura, Alegre, Bidinotto and Belo9], correspond to modified variables, with B being a model with lower propensity to PIO occurrence and C having a higher propensity. In Table 1, the derivatives omitted for models B and C mean that the original value was maintained.

2.3 ROVER algorithm

The ROVER algorithm seeks to detect PIO in real time by observing four parameters: (i) pitch rate magnitude, (ii) pitch rate frequency, (iii) pilot input amplitude and (iv) phase difference [Reference Mitchell, Arencibia and Munoz11]. When all these parameters simultaneously reach pre-established threshold values, the algorithm flags that the system is in PIO.

These parameter values were chosen according to Liu [Reference Liu12] and have been previously tested in other works from the same authors [Reference Moura, Bidinotto and Belo7–Reference Moura, Alegre, Bidinotto and Belo9]. In these trials, some human pilots were subjected to similar conditions and their subjective opinion corroborated the use of the values presented in Table 2.

Table 2.

ROVER parameters [Reference Mitchell, Arencibia and Munoz11,Reference Liu12]

2.4 Pilot models

Three mathematical pilot models were used in this study. Although differing slightly, they all seek to mathematically represent the control action of a human pilot by means of a transfer function $Y_{p}(s)$ modelling the pilot control signal u(s) for the error signal e(s). The three models together with their transfer functions are described below.

2.4.1 Tustin model

Tustin [Reference Tustin5] proposed the transfer function presented herein as Equation (4) to model the human pilot control action. In this model, the pilot’s control behaviour, level of experience and neuromuscular limitations are taken into account. These parameters are respectively represented in Equation (4) by the gain $K_p$ , the time constant of anticipation $T_{L}$ and the time delay constant $\tau$ .

(4) \begin{equation}Y_{p} = \frac{u(s)}{e(s)} = \frac{K_{p}(1+T_{L}s)e^{-\tau s}}{s}.\end{equation}

2.4.2 Crossover model

According to McRuer [Reference McRuer and Krendel6], human pilots adapt their control actions depending on how they perceive the dynamic condition of the controlled vehicle, thus that author proposed an open-loop transfer function for the pilot–vehicle system as presented in Equation (5).

(5) \begin{equation}Y_{p}Y_{c} = \frac{\omega_{c}e^{-\tau_{e} s}}{s}.\end{equation}

In this model, $Y_{c}(s)$ and $Y_{p}(s)$ represent the individual transfer functions of the vehicle and pilot, respectively, whereas $\omega_{c}$ is the so-called crossover frequency and $\tau_{e}$ is a total time delay constant modelling the time elapsed between the instant when a change in the system behaviour is identified by the human body and the moment the aircraft changes its attitude as a consequence of an action performed by the crew member.

Based on the mentioned pilot–vehicle system model, McRuer further explored the pilot behaviour and introduced a transfer function solely describing the human being as expressed in Equation (6). The presence of a time delay constant ( $T_{I}$ ) suggests the adoption of this model when the response of a controlled plant is dominated by a second-order transfer function, which serves well the aircraft modes of phugoid, short-period and Dutch roll. Additionally, $K_{p}$ stands for the gain of the pilot, and $\tau$ serves as their response time delay.

(6) \begin{equation}Y_{p} = \frac{u(s)}{e(s)} = \frac{K_{p}e^{-\tau s}}{T_{I}s+1}.\end{equation}

2.4.3 Precision model

Likewise developed by McRuer [Reference McRuer and Krendel6], the precision model expands the crossover model to take into account the human neuromuscular actuation system and an anticipation component for the pilot’s control action. The general form of its describing function is defined in Equation (7). In addition to the pilot gain $K_p$ , time delay $\tau$ and general lag time constant $T_I$ also considered in the crossover model, Equation (7) introduces the lead time constant $T_L$ towards the lead-lag pilot control actuation, as well as the term in parentheses to account for the neuromuscular system. Regarding the latter, $\omega_{n}$ and $\zeta$ respectively represent the natural frequency and damping ratio of the pilot’s neuromuscular system.

(7) \begin{equation}Y_{p} = K_{p}e^{-\tau s}\frac{T_{L}s + 1}{T_{I}s + 1}\left(\frac{1}{\frac{s^2}{\omega_{n}^2}\frac{2\zeta}{\omega_{n}}s+1}\right).\end{equation}

To utilise the models presented above, the parameters of each transfer function were obtained from experimental data gathered from human pilot trials. Thus, a model estimation methodology was applied, as detailed in the next section.

3.0 Model estimation

Human pilot trials are conducted to estimate the variables required by the pilot models. For this purpose, three volunteer pilots, identified as pilot 1, pilot 2 and pilot 3, were asked to fly the Boeing 747 model in the above-detailed cruise situation (Section 2.2). In these tests, the system shown in Fig. 1 still applies, except for the pilot model which is substituted by the human pilot, and the ROVER system, which in this case is not utilised.

The trials consist of the generation of a synthetic task in the form of a step from 0 $^\circ$ to 10 $^\circ$ (or 10 $^\circ$ to 0 $^\circ$ ).Footnote 1 with respect to the artificial horizon (Fig. 2a). The pilots are then required to capture the angle using a commercial joystick with a 40Hz acquisition rate.

In addition, for the characterisation of a PIO condition, the pilot must apply the command in an aggressive manner [Reference Mitchell and Klyde13], simulating an emergency situation. This way, both conditions are simulated: slow application (called normal gain condition) and fast application (called high gain condition). Hereinafter, these modes are also named susceptibility conditions.

Figures 3 and 4 present the results obtained for the three test pilots performing the task just described. In both cases, the Pitch Task curve (red) represents the synthetic task that the pilot should follow, while the Pilot Command contour (black) shows the angle applied on the joystick by the test pilot, and finally, the curve labelled Model Angle (blue) is the attitude of the aircraft model following the command application. Figure 3 applies to the normal gain condition, whilst Fig. 4 refers to the high gain situation.

Figure 3.

Pilot model estimation: Syntask, aircraft pitch angle and pilot input for the normal gain condition.

Figure 4.

Pilot model estimation: Syntask, aircraft pitch angle and pilot input for the high gain condition.

From the data presented in Figs. 3 and 4, the parameters of interest for the three pilot models were extracted, for both the normal and high gain conditions, as presented in Table 3 and 4, respectively. To accomplish this task, the MATLAB^® system identification toolbox was used, although due to limitations of the mentioned toolbox in identifying the real value of the pilot time delay, the pole of Tustin’s model, originally at the origin, was dislocated, resulting in the transfer function described in Equation (8)

(8) \begin{equation}Y_{p} = \frac{K_{p}(T_{L}s + 1)}{(T_{I}s + 1)}e^{-\tau s}.\end{equation}

Table 3.

Pilot model estimation: parameters for the normal gain condition

Table 4.

Pilot model estimation: parameters for the high gain condition

Finally, Table 5 provides the transfer functions for the three pilot models according to the gain condition. Entries in Table 5 are in the Z-domain (discrete-time), i.e.,

(9) \begin{equation}Y_{p} = \frac{u(z)}{e(z)}\end{equation}

Table 5.

Pilot model estimation: Z-domain transfer functions