Fault Detection and Identification for Robot Manipulators
Several factors must be considered for robotic task execution in th presence of a fault, including: detection, identification, and accomodation for the fault. In this research work, a prediction error based dead-zone residual function and a nonlinear observer are used to detect and identify a class of actuator faults. Advantages of the proposed fault detection and identification methods are that they are based on the nonlinear dynamic model of a robot manipulator (and hence, can be extended to a number of general Euler Lagrange systems), they do not require acceleration measurements, and they are independent from the controller. A Lyapunov-based analysis is providd to prove that the developed fault observer converges to the actual fault. Simulation results are provided to illustrate the performance of the detection and identification methods.
Due to the sustained needs for robotic application in remote and hazardous environments, and with emerging applications in medicine and bioengineering for the treatment of disease (often requiring patient-robot interfacing), robot reliability and fault tolerance have received significant interest. Several factors must be considered for robot operation in the presence of a fault. These factors include: detection of the fault; characterization, quantification, and identification of the fault; and then response to the fault by halting the system and/or accomodating for the fault (e.g., through a robust or adaptive controller or through system redundancy).
Fault detection is an enabling technology for fault quantification and accomodation, and hence, has recved significant interest. As described in , faults occuring in the space shuttle Remote Manipulator System (RMS) could be inferred if redundant sensors disagreed significantly with the prescribed trajectory or each other. A simple thresholding scheme was used to infer when a fault occurred; hoeever, selection of numerical values for the thresholds proved difficult, and inappropriate choices led to false alarms during missions . Clearly, fault detection will be most effective when more complete dynamic models for the manipulator are considred in the fault detection tests (residuals). In , a robust tracking controller/fault detection scheme was proposed that utilized the full dynamic model of the robot manipulator. Unfortunately, the fault detection residuals are based on conservative thresholds, which are obtained by taking the norm of user defined upper bounds for the position and velocity tracking errors. A dynamics-based approach was shown to be effective for certain types of faults in , where faults were inferred for a standard industrial robot by monitoring sudden changes in a vector of on-line parameter estimates for the robot. However, the underlying dynamic model was higky simplified (constant inertias, and coupling between joints was neglected), which implies again the need for either conservative thresholds, or again the need for either conservative thresholds, or probable false alarms. A more rigorous approach to the sysnthesis of fault detection residuals was presented in , in which the theoretical maximum number of independent residuals were derived for a manipulator with redundant sensing, based on linearized dynamics for the robot. Dynamic thresholds were developed based on full (nonlinea) manipulator dynamics. The results were promising, however the thresholds required the measurement or estimation of manipulator acceleration which is problematic, since for most practical applications the acceleration is numerically generate from position or velocity signals, and hence, the signal is inherently noisy. In , a model-based fault detection approach was successfully demonstrated expermientally. This approach was based on the generation of residuals through a filtered torque estimate which does not rely upon the measurement of acceleration quantities. Adaptive and robust detection algorithms were also developed in  to take into account possible uncertainty in the robot parameters. Other approaches to manipulator fault detection have included the development of observers , , , neural networks , , or fuzzy logic  for residual generation (see  and  for a more comprehensive overview of fault detection methods,  for an overview of fault characterization and tolerance).
Once the fault(s) has been detected, the next step in designing a fault tolerant system is to identify the fault. Based on the desire to have fault tolerant control, significant work has been focussed on this topic. The different approaches can be determined by the manner by which the output residual signals are generated and if the focus is directed at linear or nonlinear systems. In general there are two types of residual generators, structured and fixed directional . For linear systems, the residuals have been derived in several ways, including: observer-based , parity relations , , eigen structure assignment , and identification based. Similar methods have been applied for research applied to nonlinear systems. Some of these approaches have focused on nonlinear observer approaches . Other approaches apply parity relations to the nonlinear problem , . In , a fault detection and isolation architecture for nonlinear uncertain dynamics systems was presented. This approach utilizes a bank of nonlinear adaptive estimators, one for fault isolation. This approach relies on faults to be smooth functions and full state measurability is required. This approach also relies on adaptive parameter estimates for fault detection and isolation. In , the authors consider fault detection and isolation in nonlinear Euler-Lagrange mechanical systems. The approach relies on faults acting as an additive effect on the system dynamics, where exact model knowledge and full state measurability is required. A nonlinear observer with linear error dynamics is used to generate the residual for the fault detection and isolation system. Some researchers have used other tools such as fuzzy logic and artificial neural networks to approximate the models of the system and, or to identify the fault , , , ,, .
Various approaches have also been proposed for tolerating failures in robot manipulators. Most approaches have centered on the addition of some form of redundancy (e.g., in actuation ,  joints , , , , sensors ,  or software ), where the system degrades gracefully by using the redundant components. For example, if a manipulator is kinematically redundant, its end-effector task can often still be carried out by the surviving joints following a joint failure , .
The development in this paper leverages on the research presented in  nto further develop a method for robot manipulator fault detection and identification. Specifically, a fault observer and a filtered error signal are developed that do not rely upon the measurement of acceleration quantities. The fault obsserver enables the development of an estimate system which can be compared to the real system through the system states, q(t) and dq(t)/dt. The occurence of a single or a concurrent fault(s) will result in a difference between the two systems, allowing instantaneous detection of the fault. Then the fault observer asymptotically identifies the fault. A system sypervisor could use this information along with knowledge of the system to determine specifically which fault has occurred. When more than one fault occurs on one or multiple joints occurs for a robot applicaiton, the faults can be detected and identified, but the faults that are identified encompasses all faults per joint. These cases would require an advanced system knowledge to determine what specific faults were occurring in the system.
A numerical simulation was performed to demonstrate the performance of the proposed fault detection and identification system.
Block diagram for the Fault Detection and Identification
Experimental SetupFor this experiment we use the Barrett WAM. Five joints of the robot were locked at a fixed angle during the experiment and the remaining two joints of the manipulator are controlled as a planar 2 DOF robot manipulator. The control algorithm is written in C++ and are hosted on an AMD Athlon 1.2GHz PC running the QNX 6.2.1 real time OS. Data logging and online parameter tuning is performed with Qmotor 3.0 at 1KHz.
We have set up a framework where we can inject either free-swinging or ramp actuator faults on either of the joints. We can control the time when the fault will be injected and for how long the fault lasts. Also it is possible to inject different faults on the two joints.
To detect when a fault occurs we use the residual threshold based fault detection algorithm presented in . Once a fault is detected, we start the integration routine for the fault observer.
Standard P.D. control with feedforward desired acceleration is used to track the desired trajectory,
In conclusion, a fault detection and identification method is proposed for robot manipulators. The fault detection method introduced in , and the proposed fault identification method are independent from the controller. The fault identification scheme can be applied to a generic class of actuator faults that are second order differentiable. The effectiveness of the proposed fault detection/identification methods are illustrated through a numerical simulation.
For more details on this research, please refer the following conference paper: