Abstract: This paper investigates an algorithm for robust fault diagnosis in robot manipulators. The TOSM (Third Order Sliding Mode observer) provides both theoretically exact observation and unknown fault identification without filtration. The EOI (Equivalent Output Injections) of the TOSM observers can be used as residuals for the problem of fault diagnosis and to identify the unknown faults. The obtained fault information can be used for fault detection, isolation as well as fault accommodation to the self-correcting failure system. The computer simulation results for a PUMA 560 robot are shown to verify the effectiveness of the proposed strategy. Keywords: fault detection, fault diagnosis, sliding mode observer, nonlinear model I. INTRODUCTION Various approaches to fault diagnosis in nonlinear systems as well as robot manipulators have been proposed recently. The observer based on normal measurable variables have been approached [1,2]. By using neural network learning, robust fault detection scheme for nonlinear system [3], and for robot manipulators [4,5] have been developed. The basic idea of these methods is to design the robust fault diagnosis by using the model based method, and to use neural network (NN) to approximate the faults involved in the observer design. In [6], a neural-fuzzy model is used to obtain the model based of the unknown dynamic system. One of the best advantages of robust fault diagnoses is that they are not only able to detect the occurrence of a fault, but also can be provided the fault information which is useful for compensating the affect of the faults in the dynamic systems. Due to important feature of the sliding mode in the system uncertainties such as handling disturbances and modeling uncertainties through the concepts of sliding surface design and equivalent control, SM techniques have been studied for observer states by many researchers [7,8]. However, in SM applications, chattering is the major drawback in the practical realization. To avoid chattering, different approaches have been proposed [9-11]. The most widely used in practical applications to eliminate the chattering are using higher order sliding mode [12,13]. Especially, second order sliding mode [14], for instance, sub-optimal algorithm [15], super-twisting algorithm have been proposed for states observer [16,17]. However, in the second order sliding mode approach, the unknown input is constructed from the discontinuous term which provides the undesired chattering. Hence, to reduce the chattering, the filtration is required in these designs to obtain the unknown input. On the other hand, the filtration provides the delay and error that reduce the fault estimation performance. To avoid filtration which is required of second order sliding mode, the third order sliding mode observer is investigated [18,19]. In [20], the third-order sliding mode observer is designed to estimate the velocities and external perturbation. The obtained estimation of an external perturbation is used to design the controller to compensate the effect of external perturbation in the system. This paper extends earlier results of our previous work [19], the third-order sliding mode based robust fault diagnosis scheme is designed. The fault information is constructed directly from the equivalent output injection (EOI) of SM without filtration. The obtained fault estimation is used for fault detection, isolation as well as fault accommodation. To verify the effectiveness of the third order sliding mode to fault diagnosis, the simulation is performed on PUMA 560 robot. The remainder of this paper is organized as follows: in section II, the robot dynamics and faults are investigated and problems are given. In section III, the fault diagnosis scheme is designed. The simulation results for a PUMA 560 robot is described in section IV. Section V includes some conclusions. II. PROBLEM FORMULATION Let consider a robot dynamics is described by