SBA Controle & Automação Vol. 12 no. 01 / Jan., Fev., Mar, Abril de 2001 11 FAULT DETECTION AND ISOLATION IN ROBOTIC MANIPULATORS USING A MULTILAYER PERCEPTRON AND A RBF NETWORK TRAINED BY THE KOHONEN’S SELF-ORGANIZING MAP Renato Tinós and Marco Henrique Terra tinos@sel.eesc.sc.usp.br, terra@sel.eesc.sc.usp.br Departamento de Engenharia Elétrica - EESC / USP Caixa Postal 359 São Carlos, São Paulo, Brazil, 13560-970 Abstract: In this work, Artificial Neural Networks are employed in a Fault Detection and Isolation scheme for robotic manipulators. Two networks are utilized: a Multilayer Perceptron is employed to reproduce the manipulator dynamical behavior, generating a residual vector that is classified by a Radial Basis Function Network, giving the fault isolation. Two methods are utilized to choose the radial unit centers in this network. The first method, Forward Selection, employs Subset Selection to choose the radial units from the training patterns. The second employs the Kohonen’s Self- Organizing Map to fix the radial unit centers in more interesting positions. Simulations employing a two link manipulator and the Puma 560 manipulator indicate that the second method gives a smaller generalization error. 1 INTRODUCTION The search for Fault Detection and Isolation (FDI) systems for robotic systems should increase in the next years due to the moving of robots from accessible areas (like laboratories and factories) to unstructured and hazardous environments. It is already usual to talk of robots in space and undersea exploration, in medicine, in nuclear plants and manipulating explosives. Furthermore, robots should be a common household item in the next future. In these environments, and even in factories, a faulty robot can cause irreversible damages and inadmissible economic losses.. Unfortunately, faults in robots have been usual. In a research made by the Japanese Ministry of Labor, 28.7% of the industrial robots studied had a mean-time-between-failure of 100h or less; 60 % had mean-time-between failure less than 500 h (Dhillon, 1991 ad in Groom et al., 1999). Thus, there are good reasons to research FDI systems in robotic manipulators. Usually, the FDI techniques employ the mathematical model to reproduce the dynamical behavior of the fault-free system. The outputs of the mathematical model are compared with the real measurements generating a residual vector that when properly analyzed, gives the fault information. However, modeling errors can obscure the faults and can be a false alarm source (Gertler, 1997). In several cases, it is necessary the use of robust techniques (Patton et al., 1989; Mangoubi, 1998; Chen & Patton, 1999). Alternatively, a recurrent Artificial Neural Network (ANN) may be employed to reproduce the fault-free system dynamical behavior, generating the residual vector (Köppen-Seliger & P. M. Frank, 1996; Korbicz, 1997). Generally in robotic manipulators, the researchers have proposed FDI schemes utilizing the system mathematical model (Visinsky et al., 1995; Schneider & Frank, 1996; Naughton et al., 1996; Vemuri & Policarpou, 1998), employing different methods for residual analysis. In (Terra & Tinós, 1998a) a Multilayer Perceptron (MLP) trained with the Backpropagation algorithm has been utilized to reproduce the dynamical behavior of a fault-free two link manipulator to generate the residual to be analysed. The residual classification procedures have been received an important attention by several researchers, we have two central categories of analysis on the fault diagnosis: considering static and dynamic thresholds. For the static threshold we can see interesting applications in (Korbicz et al., 1999; Patan & Korbicz, 2000), and references therein (we can see, too, in these references, other techniques based on fuzzy-logic sets and genetic algorithms applied on fault-diagnosis). Obviously it is hopped that a dynamic threshold should improve the diagnosis quality, decreasing false alarms. For dynamics thresholds on the analysis of the residual classification, we have utilized Radial Basis Function Network (RBFN). This ANN has been trained using three different methods: the first, called Forward Selection (FS), employs Subset Selection to choose the radial unit centers from the training set; the second method (Global Ridge Regression - GRR) employs regularization, applying a penalty term in the large weights; the third method (Local Ridge Regression - LRR) employs regularization too, but instead of only one term, a penalty term is applied in each radial unit (Orr, 1996). Artigo Submetido em 22/12/99 1a. Revisão em 31/05/00; 2a. Revisão em 27/09/00. Aceito sob recomendação do Ed. Consultor Prof. Dr. Fernando Gomide