UAV Research at Georgia Tech Eric N. Johnson Eric N. Johnson Lockheed Martin Assistant Professor of Lockheed Martin Assistant Professor of Avionics Integration, Avionics Integration, Georgia Tech School of Aerospace Engineering Georgia Tech School of Aerospace Engineering Presentation at TU Delft Presentation at TU Delft June 3, 2002 June 3, 2002 June 2002 ENJ - Georgia Tech 1
Outline • Previous Work Previous Work • – MIT and Draper Laboratory – MIT and Draper Laboratory – Ph.D. Thesis Work: Advanced Control for the X Ph.D. Thesis Work: Advanced Control for the X- -33 33 – • Current Research Current Research • – Adaptive Guidance and Control for Hypersonic Vehicles Adaptive Guidance and Control for Hypersonic Vehicles – – Aggressive Maneuvering for UAVs Aggressive Maneuvering for UAVs – – DARPA Software Enabled Control, and the GTMax UAV – DARPA Software Enabled Control, and the GTMax UAV – Aerial Robotics Competition – Aerial Robotics Competition June 2002 ENJ - Georgia Tech 2
Draper Small Autonomous Air Vehicle (DSAAV) in 1996 Receiver/Servo Interface 6 ft Rotor IMU Compass SD MotionPak Battery Sonar Altimeter D-GPS NovAtel RT-20 Camera/Tx RF Modem 486 Computer 32cc Engine Power Distribution Modified TSK BlackStar Total Weight 23 Pounds, 10 kg June 2002 ENJ - Georgia Tech 3
DSAAV at the 1996 Aerial Robotics Competition • Organized by the Association for Unmanned Vehicle Organized by the Association for Unmanned Vehicle • Systems, International (AUVSI) Systems, International (AUVSI) • Epcot Center, Orlando, Florida Epcot Center, Orlando, Florida • Contest Area, 60x120 ft Contest Area, 60x120 ft D- -GPS Reference GPS Reference D Vision Vision Processor Processor GCS GCS Helicopter Helicopter R l o a b i o r Start Box Start Box e t Ground Coverage Ground Coverage A i c s n o T t e Emergency s Emergency a o Safety Pilot Safety Pilot B m Termination Termination June 2002 ENJ - Georgia Tech 4
June 2002 Contest Flight #5 ENJ - Georgia Tech 5
Limited Authority Adaptive Flight Control • Research Project Sponsored by NASA MSFC Research Project Sponsored by NASA MSFC • – Thesis Advisor: Anthony J. Calise, Georgia Tech – Thesis Advisor: Anthony J. Calise, Georgia Tech • Exploring Flight Control Technologies Applicable to Exploring Flight Control Technologies Applicable to • X- -33 & Future Reusable Launch Vehicles (RLV) 33 & Future Reusable Launch Vehicles (RLV) X – Reduce Analysis Required per Mission Reduce Analysis Required per Mission – – Increase Tolerance to Failures and Environment Increase Tolerance to Failures and Environment – June 2002 ENJ - Georgia Tech 6
Neural-Network Adaptive Flight Control Pseudo- -Control Control Pseudo Plant Inputs (Actual Controls) Plant Inputs (Actual Controls) Command Command - ν δ ν Approximate Approximate Reference rm Reference + Dynamic Plant Dynamic Plant Model Model + Inversion Tracking Inversion Tracking Error Error ν − ν pd Neural ad Neural Network PD Network PD Control Control June 2002 ENJ - Georgia Tech 7
Single Hidden Layer Neural Network ( ) σ ⋅ V W • Feedforward Neural Networks Feedforward Neural Networks • with a Single Hidden Layer are with a Single Hidden Layer are ( ) σ ⋅ y 1 x 1 Universal Approximators. Universal Approximators. y 2 x 2 ( ) σ ⋅ • The Sigmoidal Activation The Sigmoidal Activation • o o N 1 N 3 Function has Internal Function has Internal o N 2 y N 3 x N 1 Activation Potential ‘a’. Activation Potential ‘a’. 1 ( ) σ = 1 ( ) z σ ⋅ − + az e ( ) ν = y = T T W σ V x In matrix form: In matrix form: ad June 2002 ENJ - Georgia Tech 8
Neural Network Adaptation − x x � � ν = ν + ν − ν = rm e Error Dynamics: Error Dynamics: crm pd ad rm − x x = ν − ∆ rm x � � ( ) = + ν − ∆ e A e b � ( A is Hurwitz) is Hurwitz) ( rm rm ad = e T ζ P b Define: Define: rm + = − > T A P PA Q , Q 0 ( ) ∂ [ ] σ z ( ) ( ) = σ ' z = − Γ − + κ T W D σ σ ' V x ζ | ζ | W Adaptation Law: Adaptation Law: ∂ z W [ ] V = − + κ Γ T V D x ζ W σ ' | ζ | V (Diagonal Matrix) (Diagonal Matrix) June 2002 ENJ - Georgia Tech 9
Issues • Capability is Limited Capability is Limited • – Saturation (Including Axis Priority), Rate Limits – Saturation (Including Axis Priority), Rate Limits • Not Feedback Linearizable Not Feedback Linearizable • • Sign of Control Effectiveness Becomes Zero Sign of Control Effectiveness Becomes Zero • – Discrete Control (e.g., RCS Thrusters) Discrete Control (e.g., RCS Thrusters) – • Need to Make a Flight Certification Case Need to Make a Flight Certification Case • – Show Adaptation Extremely Unlikely to – Show Adaptation Extremely Unlikely to Cause Cause Loss of Vehicle Loss of Vehicle • Assumptions for Stability Need to be Extremely Mild • Assumptions for Stability Need to be Extremely Mild • Require Recovery from Temporary “Faulty” Adaptation Require Recovery from Temporary “Faulty” Adaptation • June 2002 ENJ - Georgia Tech 10
NN Adaptive Control with Pseudo-Control Hedging (PCH) Estimate Estimate ν x Hedge Hedge hedge Command Command - ν rm Reference Dynamic Reference Dynamic + Actuator Plant Actuator Plant ν δ δ Model Inversion Model Inversion + cmd ν pd − ν Neural ad Neural Network PD Network PD Control Control Tracking Tracking Error Error June 2002 ENJ - Georgia Tech 11
Implications • “Shelter” Adaptive Element from the Adverse Effects “Shelter” Adaptive Element from the Adverse Effects • of Plant Input Characteristics: of Plant Input Characteristics: – Linear Dynamics, Latency, Saturation, Rate Saturation, etc. Linear Dynamics, Latency, Saturation, Rate Saturation, etc. – • Achievable Adaptation Performance is Increased Achievable Adaptation Performance is Increased • Dramatically Dramatically • Adaptation is Correct During Saturation Adaptation is Correct During Saturation • – Adaptive Element Can Recover from “Faulty” Adaptation Adaptive Element Can Recover from “Faulty” Adaptation – • Enables Correct Adaptation When Not in Control of Enables Correct Adaptation When Not in Control of • Plant Plant June 2002 ENJ - Georgia Tech 12
X-33 Flight Control Sponsored by NASA MSFC • Ascent Phase • Transition and Entry • Ascent Phase • Transition and Entry – Linear Aerospike Roll/Pitch/Yaw Linear Aerospike Roll/Pitch/Yaw – Reaction Control System (RCS) Reaction Control System (RCS) – – – Aerodynamic Controls: Aerodynamic Controls: – Aerodynamic Controls Aerodynamic Controls – – • Body Flaps Body Flaps • • Elevons Elevons • • Rudders Rudders • Aero Surfaces (8) Aero Surfaces (8) RCS (8) RCS (8) Aerospike Throttles (4) Aerospike Throttles (4) June 2002 ENJ - Georgia Tech 13
Nominal Ascent Phase Results Baseline Baseline • Preliminary Results, Ascent • Preliminary Results, Ascent roll pitch yaw 2.5 Flight Control Flight Control 2 attitude error (deg) 1.5 1 – 3 – 3- -Axis Attitude System Axis Attitude System 0.5 0 -0.5 • Performance Improved Over • Performance Improved Over -1 -1.5 Existing Design Existing Design -2 -2.5 0 50 100 150 200 – Attitude Error is Lower – Attitude Error is Lower time (sec) – Hinge Moments Look Good – Hinge Moments Look Good NN NN – Nothing – Nothing is Scheduled! is Scheduled! roll pitch yaw 2.5 2 1.5 1 0.5 0 -0.5 -1 -1.5 -2 -2.5 0 50 100 150 200 time (sec) June 2002 ENJ - Georgia Tech 14
Ascent Phase Multiple Actuator Failures Baseline Baseline • Half of Aero Surfaces Fail • Half of Aero Surfaces Fail roll pitch yaw 150 Hard- -Over at 60 sec Over at 60 sec Hard 100 50 • (All Right- -Hand Surfaces Hand Surfaces • (All Right 0 Give Uncommanded Left Give Uncommanded Left -50 -100 Turn) Turn) -150 0 50 100 150 200 Failure Failure • Occurs Near Max Q time (sec) • Occurs Near Max Q (60 Seconds) (60 Seconds) NN NN roll pitch yaw 120 attitude error (deg) 60 0 -60 -120 0 50 100 150 200 time (sec) June 2002 ENJ - Georgia Tech 15
Ascent Phase Multiple Actuator Failures NN Controller • Saturates on All Three Axes • Saturates on All Three Axes • Vehicle Rolls Three Times • Vehicle Rolls Three Times Effectors Effectors • Full Recovery Once • Full Recovery Once flapR flapL elevonInR Dynamic Pressure Dynamic Pressure elevonInL elevonOutR elevonOutL rudderR rudderL Drops Drops 35 surface deflection 30 25 20 15 (deg) 10 5 0 -5 -10 -15 0 50 100 150 200 time (sec) June 2002 ENJ - Georgia Tech 16
Ascent Phase Multiple Actuator Failures NN Controller • Adaptation is “Correct” • Adaptation is “Correct” During Saturation During Saturation Roll Axis Pseudo- -Control Signals Control Signals Roll Axis Pseudo • No Knowledge of • No Knowledge of Failure Used Failure Used del vad (Not Even in the 1.2 (Not Even in the 1 Hedge!) Hedge!) 0.8 0.6 0.4 0.2 0 -0.2 0 50 100 150 200 250 time (sec) June 2002 ENJ - Georgia Tech 17
Recommend
More recommend
