Adaptive control of aeroelastic wing component

The schematic picture of the aeroelastic wing model is given Fig. 6.1 while its equations of motion are defined in Eq. 6.1 citePrime:2010. The parameters of the wing can be completed with that of the actuator system that results in certain “effective” values given in the following list: m_h = 6.516[kg], m_α = 6.7[kg],m_β = 0.537[kg]masses,x_α = 0.21(non-dimensional), x_β = 0.233[m]distance values, rβ = 0[m],a= 0.673non-dimensional distance from the mid-chord to the elastic axis,b= 0.1905[m]

semi-chord of the wing,I_α= 0.126[kg×m²],I_β = 10⁵[kg×m²],c_h= 27.43[N×m×s/rad]the plunge structural damping coefficient, c_α = 0.215[N ×m×s/rad]the pitch structural damping coefficient, cβservo= 4.182×10⁴[N×m×s/rad]trailing-edge structural damping coefficient,kh= 2844[N /m]plunge structural spring constant, k_βservo = 7.6608×10⁻³[N ×m/rad] spring constant, = 1.225[kg/m³]air density,Clα = 6.757the aerofoil coefficient of lift about the elastic axis,Cmαef f = 1.17,Clβ = 3.774, C_{mβef f} = 2.1, aerofoil moment coefficients, S = 0.5945[m], andU = 14.1[m/s]free stream velocity.

Thek_α(α)function describes some nonlinearity in the elastic deformation of the wing and it is given as kα(α) = 25.55103.19α+ 543.24α²[A. 12].

Figure 1: Schematic picture of the aeroelastic wing model (taken from [1])

(1)

This means a serious challenge in control engineering the history of which till the beginning of the 21

century is well overviewed in [3], normally considering 2 or 3 DoF approximations.

In [4] a state space representation with linear feedback determined by pole the placement technique was applied in the early eighties of the past century. The application of more sophisticated linear techniques can be found from the nineties e.g. in [5], [6], and [7]. Via application of Tensor Produc t Mode ling Technique in combination with the Highe r Orde r Singular Value Decomposi tion linear matrix inequalities can be obtained for controller design that opens possibilities for making good controllers in the possession of the precise mode l of the system to be controlled (e.g. [8], [9], and [1]). This technique can be applied not only for Line ar Parameter Varying ( LPV ), but also is useful for quasi Linear Parame ter Varying ( qLPV ) systems.

Significantly different problem class is the control of physical systems lacking precise available models . Besides the usual problem of the presence of errors in parameter estimations as e.g. in tyre fric tion models [1012], for instance the dynamic model of freeway traffic also has ambiguous form depending on the use of finite element approximations of the gradients in the space coordinate (e.g.

forward, backward or central differences) [1315]. In this case the classical adaptive, parameter tuning controllers that are designed by the use of /\DSXQRY¶V GLUHFW PHWKRG [16], [17] cannot be applied. An alternative design approach that concentrates rater on the primary design intent instead of global stability was introduced in [18] that by the use of Robust Fixed Point Transformations ( RFPT ) generates an iterative learning sequence that under certain conditions converges to the solution of the control problem. It was successfully applied for the adaptive control of freeway WUDIILF HJ >@ DV ZHOO DV IRU WKH FRQWURO RI XQGHUDFWXDWHG V\VWHPV LQ ZKLFK WKH V\VWHP¶V VWDWH cannot be fully observed [20].

On the basis of these preliminaries in the present paper we investigate the applicability of the RFPT-based iterative adaptive controller for the control of the FRQWURO VXUIDFH GHÀ ection of the underactuated 3 DoF aeroelastic wing model. In the next section a brief overview of the RFPT-based iterative adaptive learning control is given then simulations are considered.

Figure 6.1: Schematic picture of the aeroelastic wing model (taken from [74])

(6.1) For the development of a realistic controller we cannot use the exact structure of Eq. 6.1 with approximate parameters, since we cannot assume that besidesd²β/dt² the other accelerations as d²h/dt²andd²α/dt²are also measurable since due to the (3,1) and (3,2) elements of the inertia matrix (the multiplier ofd²q/dt²in Eq. 6.1) their contribution appears inu. It may be a more realistic assump-tion that no any measurable informaassump-tion we have on h and α but we have informaassump-tion only onβ, dβ/dt, andd²β/dt². This situation can simply be modeled if instead of Eq. 6.1 we use the even more rough approximation of the model as given in Eq. 6.2 [A. 11].

(6.2) Formally the equations is the same, both for adaptive and MRAC cases, but the containing elements are different. In adaptive case the parametersIβApprox= 0.8×Iβ, cβservoApprox= 0.8×cβservo, kβservoApprox= k_βservowas chosen. In MRAC caseI_βRef = 2I_β, c_βservoRef = 1.1c_βservo, andk_βservoRef =k_βservowas cho-sen.

The simulation parameters in adaptive case was: P = 1000[1/s²], I = 200[1/s³]andD = 0were

80

chosen withB = 1, K = 10⁶ , andA = 1.25×10⁶. The simulation parameters in MRAC case was:

P = 1000[1/s²], I= 200[1/s³]andD= 0were chosen withB= 1, K= 10⁶, andA= 5×10⁷ 6.1.1.1 Simulation results

The RFPT case:

Figure 6.2: Trajectory tracking of a simple PI-type (proportional, integral feedback controller) without adaptation (LHS) and with RFPT-based iterative adaptive improvement (RHS): the nominal trajectories (upper charts), the simulated trajectories (middle charts), and the tracking error (lower charts) versus time in [s] [A. 11]

Figure 6.3: The second time-derivatives of a simple PI-type (proportional, integral feedback controller) without adaptation (LHS) and with RFPT-based iterative adaptive improvement (RHS): the desired val-ues (upper charts), the adaptively deformed “required” valval-ues (middle charts), and the simulated valval-ues (lower charts) versus time in [s] [A. 11]

Figure 6.4: The control action versus time in [s] for a simple PI-type (proportional feedback controller) without adaptation (LHS) and with RFPT-based iterative adaptive improvement (RHS) [A. 11]

Figure 6.5: Trajectory tracking of a simple P-type (proportional feedback controller) without adaptation (LHS) and with RFPT-based iterative adaptive improvement (RHS) when it has information only onβ : the nominal trajectories (upper charts), the simulated trajectories (middle charts), and the tracking error (lower charts) versus time in [s] [A. 11]

Figure 6.6: The control action versus time in [s] for a simple P-type (proportional feedback controller) without adaptation (LHS) and with RFPT-based iterative adaptive improvement (RHS) when it has in-formation only onβ[A. 11]

The appropriate simulation results are given in Figs. 6.2-6.4. In general it is expected that the small integrating contribution of the PI-type controllers well reduces the steady-state errors. In the non-adaptive version the tracking errors shows considerable fluctuation, however, in the adaptive ver-sion it can be observed that the swinging is better reduced and damped as in the non-adaptive verver-sion, furthermore, the center of swinging in the tracking error signal is shifted to zero while it remains in the negative region in the non-adaptive case [A. 11].

Figures 6.5 and 6.6 belong to the results of the P-type controller with the control parametersP =

1000[1/s²],I = 0andD= 0,B= 1,K = 10⁶, andA= 1.25×10⁶. It can well be seen that by dropping the integrated feedback no very significant differences occur, but in both cases the prescribed adaptive controller successfully reduced the swinging ofβ(t)aroundβ^{N om}(t)[A. 11].

The MRAC case:

Figure 6.7: Trajectory tracking of a PI-type (proportional, and integral feedback controller) without adap-tation (LHS) and with RFPT-based MRAC adaptive improvement (RHS): the nominal trajectories (upper charts), the simulated trajectories (middle charts), and the tracking error (lower charts) versus time in [s] [A. 12]

Figure 6.8: The control signals versus time in [s] of a PI-type (proportional, and integral feedback controller) without adaptation (LHS) and with RFPT-based MRAC adaptive improvement (RHS) (upper charts), and zoomed excerpts (lower charts) [A. 12]

Figure 6.9: Thed²β^{N om}(t)/dt²(black line),d²β^Des(t)/dt²(green line), andd²β(t)/dt² (simulated) (red line) signals versus time in [s] of a PI-type (proportional, and integral feedback controller) without adap-tation (LHS) and with RFPT-based MRAC adaptive improvement (RHS) [A. 12]

Figure 6.10: Trajectory tracking of a PID-type (proportional, integral, and derivative feedback controller) without adaptation (LHS) and with RFPT-based MRAC adaptive improvement (RHS): the nominal tra-jectories (upper charts), the simulated tratra-jectories (middle charts), and the tracking error (lower charts) versus time in [s] [A. 12]

Figure 6.11: The control signals versus time in [s] of a PID-type (proportional, integral, and deriva-tive feedback controller) without adaptation (LHS) and with RFPT-based MRAC adapderiva-tive improvement (RHS) (upper charts), and zoomed excerpts (lower charts) [A. 12]

Figure 6.12: Thed²β^{N om}(t)/dt²(black line),d²β^Des(t)/dt²(green line), andd²β(t)/dt²(simulated) (red line) signals versus time in [s] of a PID-type (proportional, integral, and derivative feedback controller) without adaptation (LHS) and with RFPT-based MRAC adaptive improvement (RHS) [A. 12]

Figure 6.7 well exemplifies that some swinging inβ(t)aroundβ^N(t)can be observed in both the adaptive and the non-adaptive cases, however, in the adaptive case this swinging is much better damped.

The non-adaptive case now is unstable. The excitation of the coupled dynamic degrees of freedom can well be observed in the charts, too [A. 12].

To reveal the significance of adaptivity in Fig. 6.8 the control signals. In the non-adaptive case u^Req≡u^Destherefore the black lines are exactly covered by the green ones, but in the adaptive solution the recalculated red lines are in the close vicinity of the desired black lines and significantly differ from the exerted green lines. This means that the dynamic illusion to be created by the MRAC controller works well: on the basis of purely kinematical considerations and using the reference model the external loop (before the adaptive deformation) calculates a control signal and obtains a realized response from the controlled system that really corresponds to this control action. (It isworthy of note that the jumps in the control signal are produced by the numerical integrator of SCILAB as it works with the time-grid in the integration [A. 12].

Their appearance and positions depend on the numerical integrator chosen, and they never appear if one applies “hand−made” Euler integration in a sequential program.) The precision of the MRAC illusion is very well revealed by Fig. 6.9 that compares thed²β^{N om}(t)/dt²,d²β^Des(t)/dt², andd²β(t)/dt² (simulated) values. In the non-adaptive case very drastic PI-corrections are added tod²β^{N om}(t)/dt²that is the realized motion is far less smooth than the nominal one. In the adaptive case the three different lines are in each other’s close vicinity that means that only very little PI corrections were needed in the kinematic error relaxation design [A. 12].

In the second set of simulations PID-type kinematic error relaxation was prescribed. It was observed that the appearance of the derivative term required far smaller feedback gains for a stable control than a simple PI-type relaxation. The error relaxation strategy(d/dt+Λ)³∫

[β^N(τ)−β(τ)]dτ= 0was chosen withτ= 15[1/s]andB= 1,K= 10³, andA= 2.5×10⁴adaptive control parameters. It can be seen in Fig.

6.10 that adding the derivative term considerably improves the operation of the non- adaptive controller, and the adaptive improvement even better reduces the fluctuation ofβ(t)aroundβ^N(t). According to Figs. 6.11 and 6.12 it can be stated that the MRAC illusion was almost perfect again [A. 12].