«a paper presented at SPIE’s 1996 Symposium on Smart Structures and Integrated Systems Jennifer L. Pinkerton Anna-Maria R. McGowan Robert W. Moses ...»
The PARTI model was first tested in March 1994 to obtain basic flutter characteristics and transfer functions relating the model response to piezoelectric actuator input for various tunnel conditions. A closed-loop wind-tunnel test of the PARTI model was completed in November 1994. Preliminary results have been previously published regarding data collected during these tests. Analysis and design of several control laws along with preliminary test data have also been published. 24 Further discussion of wind-tunnel testing results and a discussion of experimental methodologies can be found in reference 25. The current paper will present a brief description of the PARTI model and wind-tunnel testing.
3.1.1 Wind-tunnel model
The model is a five-foot long, high aspect ratio semi-span wing designed to flutter at low speeds to simplify aerodynamic analyses and wind-tunnel testing. The model consists of two primary structures: an exterior fiberglass shell used to obtain aerodynamic lift and an interior composite plate that contains the piezoelectric actuators and acts as the main load carrying member. The fully assembled model mounted in the TDT is shown in Figure 3. Figure 4 shows the interior construction of the model. The interior plate is composed of an aluminum honeycomb core sandwiched by graphite epoxy face sheets.
/The face sheets are of [-2Oo2/O0]laminate, referenced to the wing quarter-chord which is swept aft 30". The unsymmetric, unbalanced composite lay-up provides a static bendhwist coupling. The model has two additional components: a trailingedge aerodynamic control surface and a wing-tip flutter-stopper device. The flutter-stopper tip-mass assembly was constructed as a safety device for wind-tunnel testing. Piezoelectric actuators cover the inboard 60% of the span of the internal composite plate and account for 7.3% of the total wing weight. Fifteen groups of piezoelectric actuator patches are adhered to the top and bottom of the interior plate. The actuators are configured to impart differential bending moments to the plate; however, the ply orientation of the graphite epoxy and the wing sweep angle make it possible for piezoelectric actuation to affect both the bending and torsion natural modes of the model. The piezoelectric patches were only used for actuation; ten strain gages and four accelerometers were used as sensors. The complete design of the PARTI model is documented by Reich and CrawIey. 26
3.1.2 Wind-tunnel testing
The PARTI model was ground tested and wind-tunnel tested during two entries in the TDT. Ground vibration tests were conducted to determine the dynamic characteristics of the model at zero-airspeed and to validate analytical models.
Experimental flutter characteristics and open-loop time-history data at subcritical conditions were obtained during the first wind-tunnel test in March 1994. Time history data was acquired at several dynamic pressures by actuating each of the 15 groups of piezoelectric actuators individually as well as in five sets of several actuator groups. The open-loop data were used to construct state-space models for control law design, to examine the linearity of piezoelectric actuation, and to verify analytical models and techniques. A time domain system identification method was used to generate the state-space models from the time histories obtained during wind-tunnel testing. Using system identification, all of the PARTI model dynamics were fully captured using a math model with 120 states; however, state-space models with as little as 40 states were used as well. The open-loop data were also used to examine the linearity of piezoelectric actuation. Despite the presence of nonlinearities, the results indicate that superposition can be used to combine the responses of individual piezoelectric actuator groups. Furthermore, preliminary results show that there is a linear increase in model response with a linear increase in piezoelectric command voltage.
Twenty-eight control laws designed to increase flutter speed and reduce response at subcritical conditions were tested during the second wind-tunnel test. A variety of control law design techniques were used and both single-inputlsingle-output (SISO) and multi-input/multi-output (MIMO) control laws were designed utilizing up to five inputs and nine outputs. The PARTI model can be represented by up to 15 control outputs (1 5 groups of piezoelectric actuators) and 14 control inputs (10 strain gages and 4 accelerometers). However, several piezoelectric actuator groups receiving the identical output signal were considered as one control output. The most successful flutter suppression control law was also effective in reducing response at subcritical speeds, demonstrating a 12% increase in flutter dynamic pressure and a 75% reduction in the power spectral density of peak response at subcritical speeds as shown in Figures 5 and 6. This was a SISO control law that used all 15 piezoelectric actuator groups. Another SISO control law that used strain gage 4 for feedback and piezoelectric actuator groups 3,4, 6,7, and 10 successfully increased flutter speed 8%.
The PARTI program successfully demonstrated active flutter suppression and reduced response at subcritical speeds using piezoelectric actuation on a five-foot span wind-tunnel model. Through this program, a number of issues associated with applying piezoelectric actuation to aeroelastic problems have been addressed; however, there exist several issues that require further study, some of which are planned as follow-on studies to the present research. These include an examination of scaling laws, power consumption, increasing control effectiveness through optimal actuator and sensor selection and placement, and the analytical development of state-space models.
3.2 AdaDtive Neural Control of Aeroelastic Response (ANCAR) Dromam
An important aspect in the development of a smart structure or integrated system is the corresponding development of an adaptive controller. The Adaptive Neural Control of Aeroelastic Response (ANCAR) program is a cooperative effort between NASA LaRC and McDonnell Douglas Aerospace to develop and demonstrate an integrated flutter suppression system that uses an adaptive neural network controller. The ANCAR program is comprised of three phases. Phases I and I1 use the LaRC Benchmark Active Controls Testing (BACT) wind-tunnel model. The Phase I objectives were to develop and demonstrate a hybrid control system that incorporated conventional control algorithms and neural networks to suppress flutter. Wind-tunnel testing for this phase took place in January 1995. The Phase I1 objectives are to develop and demonstrate an adaptive neural-network-based flutter suppression system during a January 1996 wind-tunnel entry. Phase I11 objectives are to combine the neural adaptive control system with an active structure like the PARTI model, Currently, wind-tunnel testing for this phase is scheduled for 1998. This section of the paper will focus on the results of Phase I and the plans for Phase 11.
3.2.1 BACT wind-tunnel mode I
The BACT model *' was originally developed as part of the Benchmark Models Program (BMP). *' The BMP was a NASA LaRC program that included a series of models which were used to study different aeroelastic phenomena and to validate aeroelastic, aeroservoelastic, and computational fluid dynamic methods in several wind-tunnel tests. Because the BACT model dynamics are well understood and it has control surfaces and relatively benign flutter mechanisms, it is an ideal testbed for initial testing of new control schemes, including neural-based systems.
The BACT model is depicted in Figure 7. The model is a rigid, rectangular wing with a NACA 0012 airfoil section. It is equipped with a trailing-edge control surface and upper- and lower-surface spoilers, all independently controllable. The model, shown in the figure, is attached to a mount system called the pitch-and-plunge apparatus (PAPA) that allows both pitch and plunge degrees-of-freedom. The model is extensively instrumented with pressure transducers and accelerometers to measure surface pressures and model dynamic response, and the mount system is instrumented with strain gages to measure normal force and pitching moment.
3.2.2 Phase I
The schematic for the Phase I neural network control system is shown in Figure 8. This is a hybrid control system that uses a neural network to gain schedule conventional control law parameters with varying Mach number and dynamic pressure.
Figure 8 shows how the system was implemented. The neural network was trained to use Mach number and dynamic pressure as input and provide the coefficients of the control law as output. After conversion from continuous to discrete time, the control laws could be downloaded and run on a real time digital controller.
The neural network was trained by using fifty-six different state-space analytical models, each corresponding to a different combination of Mach number and dynamic pressure, which ranged from 0.3 to 0.9 and 75 psf to 250 psf, respectively. These models were used to generate fifty-six 3-zero and 4-pole, single-input/single-output control law designs, each optimized to achieve minimum RMS model response to tunnel turbulence. For each of the fifty-six control laws, there were 7 parameters that could be varied with Mach number and dynamic pressure, and a multi-layer perceptron neural network was trained with backpropagation to output these seven control law parameters as a function of Mach number and dynamic pressure. For comparison purposes, the same design strategy was used to generate a fixed-gain flutter suppression control law to provide the best possible performance over the whole flight envelope.
In the TDT, testing is accomplished easily and conveniently by frst choosing a tunnel stagnation pressure and then varying drive motor RPM, which simultaneously changes Mach number and dynamic pressure. During Phase I testing, four different stagnation pressures were chosen, resulting in wide-ranging combinations of Mach number and dynamic pressure within the training space. Data was acquired at each of the four constant stagnation pressure conditions for the open-loop system, for the closed-loop fixed-gain system, and for the closed-loop neural-network-scheduled system. The RMS of the trailing edge accelerometer response for one of the stagnation pressures is shown in Figure 9. For this pressure, open-loop flutter occurs at a dynamic pressure of 161 psf. Here, the neural-network-scheduledsystem has the lowest response.
3.2.3 Phase I1
The goal of Phase I1 is to incorporate neural networks into adaptive flutter suppression systems. Unlike the Phase I controller, the Phase I1 neural systems will not use conventional control designs. Instead, they will adapt, on-line, to changing test conditions. A variety of adaptive neural-network-based systems will be tested during the January 1996 test.
Several of the algorithms to be tested use a predictive control algorithm. Predictive control algorithms require generation of a math model for the system being controlled. Once the model is complete, the loop can be closed and a search or optimization algorithm can be used to select the control input commands that further minimize response over some time interval. In this application, a neural network is trained using experimental data acquired in the wind tunnel to create a neural model of the system that can be used to predict future responses to future control inputs. A neural network realization of predictive control is shown in Figure 10.
This experimental study is investigating the use of neural networks for adaptive control of flutter with the BACT wind-tunnel model. During Phase I testing, a hybrid control system using a neural network to gain schedule conventional control law parameters was implemented for the first time, demonstrating slightly better flutter suppression than a fixed gain control system. Demonstration of an advanced neural network that can adapt to changing conditions is the goal for Phase 11.
3.3 The Activ elv Controlled Response of Buffet Affected Tails (ACROBAT) p r m
The goal of the Actively Controlled Response of Buffet Affected Tails (ACROBAT) program was to demonstrate the feasibility of using active control for alleviating buffeting caused by leading edge extension (LEX) vortex burst. This phenomenon is shown in Figure 1 1. For this program, an existing 1/6-scale rigid full-span model of the F/A- 18 A/B aircraft was refurbished, and three new flexible and two new rigid vertical tails were fabricated. The model, shown in Figure 12, was tested during July 1995 and November 1995 with each of the new flexible tail surfaces at a Mach number of 0.09 in a test medium consisting of air at atmospheric pressure. The three flexible tails were fabricated from a l/S-inch thick aluminum plate and covered with balsa wood to produce the desired airfoil shape. The three flexible vertical tails used: (1) a rudder surface; (2) a tip vane configured with a slotted cylinder; 3) an embedded slotted cylinder; or (4) piezoelectric actuation devices. For the vertical tail with piezoelectric actuation, a hatch cover was incorporated in the design to cover the platemounted piezoelectric devices while maintaining the proper airfoil shape. The piezoelectric actuator arrangement itself, shown in Figure 13, consisted of three root actuators, two tip actuators, and two central actuators per side. Each of the tip and central actuators was comprised of two stacks of two layers of piezoelectric wafers, while each root actuator utilized two stacks offour layers for added effectiveness. The response of each flexible vertical tail was measured using a root bending strain gage and two tip accelerometers. Pressure transducers were surface mounted on both sides of all tails, excluding the tail with piezoelectric actuation.