des FA 4.13 Steuerung und Regelung von Robotern A Physically-Based Fault Detection and Isolation Method and Its Uses in Robot Manipulators Alessandro De Luca Dipartimento di Informatica e Sistemistica Università di Roma La Sapienza currently on leave at Institute of Robotics and Mechatronics DLR Oberpfaffenhofen
Outline FDI problems Robot dynamics and physical properties Detection and isolation of actuator faults Adaptive scheme for actuator FDI Collision detection and reaction Extension to robots with joint elasticity collision detection/reaction + motor friction compensation Conclusions
FDI problems Fault Detection recognizing that a fault is affecting a dynamic system Fault Isolation discriminating the occurrence of a fault f from that of all other considered possible faults and disturbances FDI solution approach (model-based) design a residual generator system whose output is only affected by the fault f to be detected and isolated is not affected by any other fault or disturbance converges (asymptotically) to zero whenever f = 0
Robot dynamic models fully rigid case presence of transmission/joint elasticity
Relevant physical properties kinetic and potential energy relation between inertia and velocity terms
Relevant physical properties (cont d) total energy and its variation generalized momenta and their decoupled dynamics
Robot actuators FDI faulted model commanded torque fault types fault torque total failure power loss saturation bias possibly concurrent, intermittent, incipient, abrupt,
Basic assumptions full state measurements implementation with available sensors (typically, position only) robot dynamic model accurately known adaptation might be included for uncertain parameters use of detection thresholds to handle noise (false alarms) only commanded torque available (no fault model is needed) any control input law open or closed-loop, linear or nonlinear model-based feedback no need of a specified reference motion
Early solutions 1. : compare computed model-based torque (from measures) with commanded one 2. : compare simulated acceleration (inverse robot dynamics) with those from measurements
and their limitations noisy acceleration (e.g., from double numerical differentiation of position measures) inversion of inertia matrix intrinsic delay (one or more digital steps) dependence on commanded input dynamics poor or no fault isolation (only detection)
Energy-based fault detection scalar detector and its dynamics (needed only for analysis)
Momentum-based FDI vector of residuals and its decoupled dynamics (a stable first-order linear filter)
Experimental setup Quanser Pendubot 2nd link (passive) 1st link (actuated) ABB, Ladenburg, Germany video swing-up Pendubot 25 January 2006
Actuator FDI on Pendubot partially concurrent 10% power loss on actuator 1 and total failure on (missing) actuator 2 PID control on first joint to 30 commanded torques joint 1 joint 2 joint positions
Actuator FDI on Pendubot (cont d) thresholding and dynamic filtering of residuals residuals filtered residuals joint 1 joint 2
An adaptive FDI scheme include friction (difficult to estimate) in the model linear parametrization (may be extended to gravity and inertia-related terms)
Adapt and detect using an estimate of friction parameters residual dynamics
Adapt and detect (cont d) stability analysis via standard Lyapunov and LaSalle techniques (in absence of faults) parameter estimates converge to constant values (= correct ones for sufficient excitation) by overparametrization and suitable gain scaling, one may still adapt also during faults
Adaptive actuator FDI on Pendubot situation as before, with power loss increased to 50% on actuator 1 on-line adaptation of both friction and gravity parameters commanded torques joint 1 joint 2 residuals
Collision as a fault rigid robot model use only proprioceptive sensors possible contact at any point along the arm simplifying assumptions single contact robot as open kinematic chain unfaulted actuators transpose of contact point Jacobian
Analysis of collisions y 2 x 2 y 0 y 1 F K F K q 2 x 1 d 2 d 1 q 1 x 0
Collision detection as before, scalar detector only contact forces (wrenches) that perform work on contact velocity (twists) can be detected
Directional detection and isolation as before, vector of residuals ideal situation (no noise) collision point is located up to link i
Choice of residual gains evaluation by simulation on 7-dof DLR-III arm (impact on last link) joint 2 @30 /s joint 4 @200 /s 10 ms
Collision reaction strategies normal operation in zero-gravity once collision is detected ( above threshold) either stop the robot (braking) and then possibly reverse commanded motion (backtracking) or apply a reflex strategy with torque control using directional information of residual vector (move in the same direction of sensed force)
Dissipating energy when contact is lost, the residual decays until dissipate kinetic energy at highest rate (using maximum available torque) until robot stops
Operative robot states collision = 0 normal operation in zero-gravity collision = 1 residual > low reflex reaction velocity = 0 energy dissipation residual low velocity 0
Robots with elastic joints (EJ) harmonic drives introduce joint elasticity effects motor friction and possible arm collisions DLR-III arm: motor position and joint torque sensors
Multiple detection for EJ robots it is simultaneously possible to compensate friction (a fault) on motor side detect collision at link side 1. unmodeled motor friction detection and compensation (based on motor generalized momenta) decentralized linear observer (includes acceleration estimation) motor friction compensation
Multiple detection for EJ robots (cont d) 2. collision detection: several alternatives are possible for generalizing the rigid case analysis, the most simple is replace joint to motor torque robot control laws should be modified (e.g., in DLR-III arm) reflex strategies to contact detection include torque mode reaction admittance mode reaction
DLR-III robot controller motor inertia reduction based on joint torque sensing leads to with general position/torque control law (depending on reference and gain values) obtaining a full state feedback law static gravity compensation (based on motor position)
Reflex strategies strategy 2: free-floating torque mode strategy 3: torque control mode strategy 4: admittance control mode
Experiments on DLR-III arm (1) Head Injury Criterion (HIC) tests on dummy head 3D accelerometer on dummy head
Results on dummy head impact approaching at 30 /s with each joint residual gains = diag{25} joint torque joint 1 residual 0/1 detection acceleration 2 ms
Experiments on DLR-III arm (2) strategy 4 @90 /s one of 99 luftballons ABB, Ladenburg, Germany 25 January 2006
Results on balloon impact residual & velocity on joint 4 for different reaction strategies no reaction impact at 10 /s with coordinated joint motion
Results on balloon impact (cont d) residual & velocity on joint 4 for different reaction strategies impact at 100 /s with coordinated joint motion
Human-robot interaction (1) first impact @60 /sec video HRI - 1 strategy 4: admittance control based on residuals
Human-robot interaction (2) first impact @60 /sec video HRI - 2 strategy 3: torque control based on residuals
Human-robot interaction (3) first impact @90 /sec video HRI - 3 strategy 3: torque control based on residuals
Conclusions powerful FDI technique for mechanical systems based on physical quantities (energy, momenta) direct extensions to joint elasticity, actuator dynamics, friction compensation, adaptation to uncertain parameters special case of a more general geometric theory valid for sensor/actuator faults of nonlinear (affine) plants under possible concurrency, exact FDI for a maximum number of faults = N (# of generalized coordinates) principle feasible also for industrial robots, for advanced safety requirements in human-robot physical interaction
Acknowledgments scientific contributions by Raffaella Mattone (DIS, Roma) Giulio Milighetti (ex DIS, Roma; now Fraunhofer IITB, Karlsruhe) Alin Albu-Schäffer (DLR, Oberpfaffenhofen) Sami Haddadin (DLR, Oberpfaffenhofen) work supported by Humboldt-Helmholtz Association (2005 Research Award for foreign scientists)
References De Luca, Mattone: Actuator FDI using generalized momenta, ICRA 03 De Luca, Mattone: Adapt-and-detect robot actuator faults, ICRA 04 De Luca, Mattone: Identification of robot actuator faults, IROS 05 De Luca, Mattone: Sensorless robot collision detection and hybrid force/motion control, ICRA 05 Mattone, De Luca: FDI in Euler-Lagrange mechanical systems, ASME JDSMC (submitted), May 2005 Mattone, De Luca: Relaxed FDI for nonlinear systems, Automatica, 2006 Albu-Schäffer, De Luca, Haddadin, Hirzinger: Collision detection and reaction strategies with DLR-III arm, IROS 06 (to be submitted)