Flight Verification and Validation of an L1 All-Adaptive Flight Control System Enric Xargay, Naira Hovakimyan Department of Aerospace Engineering University of Illinois at Urbana-Champaign e-mail: {xargay, nhovakim}@illinois.edu Safe & Secure Systems & Software Symposium - June 15, 2010
Outline Adaptive Control in Transition Robust Fast Adaptation: L1 Adaptive Control AirSTAR Project L1 All-Adaptive FCS Piloted Sim Evaluations Flight Test Evaluations (NOT flight demonstrations) Conclusions March 2010 AirSTAR Deployment June 2010 AirSTAR Deployment Closing certification gaps Real alternative to gain scheduling
Adaptive Control in Transition Fast adaptation Single design AFCS NPS FlightTest Program Sig RASCAL IRAC (NASA) GTM T2 60s 90 95 00 05 10 X-15 (NASA/USAF/ US Navy) IFCS (NASA/Boeing) F-15 ACTIVE RESTORE (AFRL-VA/Boeing) X-36 Adaptive Control for Munitions (AFRL-MN/GST/Boeing) MK-84 MK-82 L-JDAM J-UCAS (DARPA/USAF/US Navy) Boeing X-45A & X-45C in production MK-84 JDAM Gen I: flown 1999, 2003 Gen II: 2002 2006 flight test 4th Q 2005 Gen III: 2006 Slow adaptation Expensive gain-scheduled AFCS evaluated in flight sim environment in production Source: Kevin Wise, Boeing (adapted)
Robust Fast Adaptation: the key to safe flight Predictable :: Repeatable :: Testable :: Safe Control law objectives: Keep aircraft in the wind tunnel data envelope (accurate models) Is A/C controllable here? Eventually, return to normal flight envelope Control actions within 2-4 seconds of failure onset are critical: Need for transient performance guarantees Predictable response Need for fast adaptation Source: NASA Failure of conventional adaptive control (limited to slow adaptation)
Main Features of L 1 Adaptive Control Separation between adaptation and robustness Performance limitations reduced to hardware limitations Guaranteed robustness with fast adaptation Guaranteed transient response for input and output NOT achieved via high-gain feedback or persistence of excitation or gain-scheduling or control reconfiguration Guaranteed (bounded away from zero) time-delay margin Uniform scaled transient response dependent on changes in initial conditions, uncertainties, and reference inputs Verifiable software with computationally predictable numerical characteristics Systematic design guidelines suitable for flight verification Suitable for development of theoretically justified Verification & Validation tools for feedback systems
NASA Langley AirSTAR :: Generic Transport Model High-risk flight conditions, some unable to be tested in target application environment. 5.5 % geometrically and dynamically scaled model 82in wingspan, 96in length, 49.6 lbs (54 lbs full), 53 mph stall speed Model angular response is 4.26 faster than full scale Model velocity is 4.26 times slower than regular scale
AirSTAR :: Challenges Inner-loop state-feedback controller for tracking angle of attack, roll rate, and sideslip angle commands. Challenges: Single all-adaptive design for the entire flight envelope (including stall and post stall high α conditions), without gain scheduling Compensation for structural damage/actuator failures without FDI methods Compensation for unmatched uncertainties variations is α, β, V dynamics with flight condition Strict performance requirements: High precision tracking Reduced workload Predictable response!!! L 1 AFCS Hardware requirements: Euler integration at 600Hz
AirSTAR :: All-Adaptive FCS
Baseline p [deg/s], [deg] p [deg/s], [deg] p [deg/s], [deg] p [deg/s], [deg] L1 Controller p [deg/s], [deg] p [deg/s], [deg] p [deg/s], [deg] p [deg/s], [deg] AirSTAR :: Piloted Sim Evaluation (Bank capture) Bank angle capture task : φ 45 deg Acquire in 2 sec, hold for 2 sec; desired ± 5 deg, adequate ±10 deg. Nominal ΔClp = -50% ΔClp = -75% ΔClp = -100% 30 20 10 0-10 -20-30 -40 p cmd p L1 bank -50 0 5 10 15 time [s] 30 20 10 0-10 -20-30 -40 p cmd p L1 bank -50 0 5 10 15 time [s] 40 30 20 10 0-10 -20-30 -40 p cmd p L1 bank -50 5 10 15 20 time [s] 40 20 0-20 -40 p cmd p L1 bank -60 0 2 4 6 8 10 12 time [s] Level I Level I Level I Level II (CH4) 60 40 20 0-20 -40 p cmd p Base bank 0 5 10 15 time [s] 40 20 0-20 -40-60 p cmd p Base bank 8 10 12 14 16 18 20 22 time [s] 60 40 20 0-20 -40-60 p cmd p Base bank -80 2 4 6 8 10 12 14 time [s] 150 100 50 0-50 p cmd p Base bank -100 35 40 45 time [s] Level I Level I Level III (CH9) Uncontrollable (CH10)
AirSTAR :: Piloted Sim Evaluation (asymmetric engine failure) 1. Full throttle (100%) 2. Climb at 25-30 deg pitch 3. Left Throttle cut to 0% in <0.5sec L1 Adaptive Control Stick-to-Surface Stick-to-Surface L 1 Controller
GTM T2 :: Flight Test Evaluation (March 2010) L1 all-adaptive FCL: provides performance/stability for nominal and impaired aircraft Not an augmentation to a baseline controller that provides nominal aircraft performance, like other adaptive controllers implemented Flight Control Law related tasks during March 2010 deployment: Flight Control Law Block : Injected longitudinal and lateral stick doublets for each fault, continuous stick doublets on straight legs during latency fault Latency fault: starting at 20msec, continuously increase in latency (5msec every 5sec) through the turns, etc until aircraft is neutrally stable or unstable want graceful performance degradation Robust to 105msec of additional time delay Simultaneous longitudinal and lateral stability degradation (Cmα/Clp): 50%: nominal performance 75%: nominal performance 100%: small degradation of performance observed by the pilot 125%: divergent closed-loop system Left elevator inboard and outboard segments locked-in-place failure (<2deg): nonevent for the adaptive controller
GTM T2 :: Flight Test Evaluation (March 2010) FLT14: Mode 3.2 (L1 all-adaptive) FCL under moderate (+) turbulence
GTM T2 :: Flight Test Evaluation (March 2010) this is the first successful flight of an all-adaptive control law that deals with aircraft stability degradation as well as actuator failures it is the first flight of a direct all-adaptive controller with a pilot in the loop NASA RTD weekly key activities report Dr. Irene M. Gregory
GTM T2 :: Flight Test Evaluation (June 2010) L1 all-adaptive FCL: provides performance/stability for nominal and impaired aircraft Not an augmentation to a baseline controller that provides nominal aircraft performance, like other adaptive controllers implemented Flight Control Law related tasks during June 2010 deployment: Flight Control Law Block : Injected longitudinal and lateral stick doublets for each fault, continuous stick doublets on straight legs during latency fault Latency fault: starting at 20msec, continuously increase in latency (5msec every 5sec), carried through the turns, until aircraft is neutrally stable or unstable want graceful performance degradation Robust to 130msec of additional time delay Simultaneous longitudinal and lateral stability degradation (Cmα/Clp): 50%: nominal performance 75%: nominal performance 100%: small degradation of performance observed by the pilot 125%: large amplitude roll with pitch doublet Left elevator inboard and outboard segments locked-in-place failure (<2deg): nonevent for the adaptive controller Modeling Tasks: L1 used for β-sweep in flat turn maneuver
GTM T2 :: Flight Test Evaluation (June 2010) FLT23: Mode 3.6 (L1 all-adaptive) FCL under light turbulence SP Research Pilot SP SP Research Pilot SP High AOA flight Post-stall regimes ~12.5 mins of flight with L1
GTM T2 :: Flight Test Evaluation (June 2010) Post-stall, high angle of attack flight Open-loop aircraft tends to aggressively roll off between 13deg and 15deg AOA and exhibits significant degradation in pitch stability Stick to surface Aggressive departure Roll rate above 60dps Normal flight FQ Level I A/C All 3 stick-to-surface attempts in maintaining controlled flight at AOA=18deg were unsuccessful
GTM T2 :: Flight Test Evaluation (June 2010) Post-stall, high angle of attack flight L1 provides departure resilient control for aircraft in post-stall flight L1 adaptive controller achieved a very well controlled aircraft (pilot assessment) L1 AFCS Repeatable results Two AOA=18deg acquisitions with L1 AFCS Pitch break Oscillations around 15deg AOA A well controllable aircraft during stall and post-stall flight Dan Murri AirSTAR GTM T2 research pilot Roll rate below 20dps
GTM T2 :: Flight Test Evaluation (June 2010) Post-stall, high angle of attack flight L1 provides departure resilient control for aircraft in post-stall flight Stick to surface L1 AFCS
GTM T2 :: Flight Test Evaluation (June 2010) L1 All-AFCS Flight Test Summary: All-adaptive FCS that provides nominal aircraft performance and takes care of large changes in aircraft dynamics No baseline to assist in A single controller design at a nominal flight condition (4deg AOA) to provide satisfactory FQ and robustness No gain scheduling of control parameters Predictable response to the pilot under stability degradation and graceful performance degradation once nominal response was unachievable Departure resistant in post-stall flight: L1 provides a controllable aircraft to the pilot and facilitates safe return to normal flight Good time-delay margin Robust to control saturation (occurred during high AOA flight) This is the first post-stall flight of an adaptive controller
GTM T2 :: X-29 :: X-48B GTM T2 All-adaptive NASA Langley Grumman X-29 Augmentation of an LQR-PI NASA Dryden X-48B Blended Wing Body Augmentation of a dynamic inversion controller AFRL/Boeing
Closing the Certification Gaps (I) From : Closing the Certification Gaps in Adaptive Flight Control Software Jacklin, 2008 Verification, Validation, and Certification Challenges for Adaptive Flight-Critical Control Systems Software Jacklin et al., 2004 [ ] the Lyapunov analysis only guarantees the ultimate stability of the learning algorithm; the proof does not guarantee how fast the system returns to the origin. Ultimate boundedness and asymptotic convergence are weak properties for (nonlinear) adaptive closed-loop systems. L1 achieves guaranteed uniform transient and steady-state performance. This is an important point for system performance, because if learning happens too slowly, an adaptive controller may be rendered ineffective for the control task at hand. Fast adaptation -only limited by hardware- allows for compensation of undesirable effects of rapidly varying uncertainties and significant changes in system dynamics. Yet another gap is that methods to find acceptable gains for stable learning must be found other than the time intensive and error process. Adaptation rate set as fast as hardware permits. Conventional tools and methods from classical and robust control can be applied to tune the LINEAR filter of L1 architectures.
Closing the Certification Gaps (and II) From : Closing the Certification Gaps in Adaptive Flight Control Software Jacklin, 2008 Verification, Validation, and Certification Challenges for Adaptive Flight-Critical Control Systems Software Jacklin et al., 2004 A persistent excitation signal added to the control signal yields more convergent learning, but at the expense of poorer steady-state controller performance. With an adaptive system [ ] test inputs might provide the excitation needed for better learning and adaptation that might not be available to the fielded system. Guaranteed performance of L1 does not rely on persistence of excitation assumptions or on the injection of persistently exciting signals. Closed-loop response does not depend on the nature of the external inputs. Fast adaptation is the key for predictability and repeatability. On-line adaptation of network connection weights makes it impossible to predict, in a deterministic sense, the future form of the neural network mapping. [ ] the computation time was affected by the number of connections between nodes [of the NN]. [ ] With all this variability, it proved difficult challenge to validate the worstcase computation time required for this type of neural network. Fast adaptation allows for control of time-varying nonlinear systems by adapting two parameters only. No need to resort to neural networks. Deterministic control algorithm.
L 1 adaptive control architectures: Performance and robustness guarantees Systematic design guidelines Conclusions Computationally predictable characteristics Design of robust adaptive flight control systems: Single design for the entire flight envelope (including stall and post-stall conditions) without Gain-scheduling/Persistency of excitation/control reconfiguration/high-gain feedback Compensation for structural damage and actuator failures without FDI methods Consistent results from platform to platform, as predicted by theory Implementation as an all-adaptive controller or as an augmentation loop for baseline controllers 2 successful flights with NASA s GTM T2 and over 100 successful flights with NPS Suitable for development of theoretically justified Verification & Validation tools for feedback systems
Real Alternative to Gain-Scheduling Günter Stein (1980s): The main point made is that for conventional flight control problems, adaptive control is the losing alternative in a historical competition with explicit airdata-scheduling. L1 Adaptive Control represents the first step towards a real alternative to gain-scheduled controllers!
Acknowledgements This research was supported by: NASA under grants NNX08BA64A and NNX08BA65A AFOSR under Contract FA9550-09-1-0265 AFRL under Contract F33615-00-D-3052 Collaborators: Chengyu Cao (UConn) Irene M. Gregory (NASA Langley) Special thanks to the staff of the AirSTAR Flight Test Facility for their support with control law implementation.
Questions?