IFAC-PapersOnLine 49-15 (2016) 050–056 ScienceDirect Available online at www.sciencedirect.com 2405-8963 © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. Peer review under responsibility of International Federation of Automatic Control. 10.1016/j.ifacol.2016.07.613 © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved. 1. INTRODUCTION Fault tolerant behavior in an autonomous mobile robot is desir- able for a variety of reasons. Timely fault diagnosis increases the ability to complete tasks satisfactory and improves per- formance and safety – the robot becomes more efficient and economic. The ability of a system to recognize errors and draw conclusions about future actions enables it to avoid failures such as mission abortion, material damage and accidents with humans. After estimating the severity of a fault and determining its location, the system could be more easily repaired. Fault tolerant robots stand to be more flexible to new circumstances and environments. Various approaches to fault diagnosis can be found in the liter- ature: some are based on expert systems and analysis of large databases (Zaman et al. (2013), Nan et al. (2008)), others are based on hardware redundancy Kleiner et al. (2008), and yet another group is based on models that use some domain knowl- edge in order to facilitate system analysis Isermann (2006). In many cases the physical (mathematical) model of the system is known, we are thus able to calculate the correct outputs for the given input data. The main problems stem from the use the physical model of the robot, as it is hard to exactly determine. Since the world is dynamic, the robots environment can always change, therefore the robot model has to be customized accord- ing to an ever changing environment. If the physical model is unknown, it has to be estimated from a set of training data. To a given or estimated model we can apply model-based fault diagnosis algorithms. The general principle underlying this procedure is to compare the expected behavior of the system that is given by the model with the actual behavior, known through on-line observations. ⋆ This work was sponsored by the B-IT foundation and the Strukturfond des Landes Nordrhein-Westfalen for the female PhD students. For a better understanding of theoretical basics and practical application we intend to discuss and compare four fault diagno- sis approaches. These approaches model the knowledge about nominal and faulty states of a robot using various representa- tions. Additionally to the known fault diagnosis methods we also utilize the new technique (in fault diagnosis domain) which seems to be very promising for fault diagnosis due to the rich class of models and its support of non-stationary processes. In this work we point out the strengths of each algorithm and determine which method best suits the demands of fault diagnosis for a robot. Theoretical background From a general perspective fault di- agnosis (Nyberg (1999)) can be explained as follows: The task is to generate a diagnosis that states whether a fault arises or not. If a fault is determined, its location has to be identified. Thus, fault diagnosis algorithms provide two critical pieces of information: detection (whether a fault occurs) and identifica- tion (where the fault occurs). Most fault diagnosis methods are based on the concept of re- dundancy (extra resources) in the system, so that a variable of interest can be calculated in more than one way. There are two types of redundancy: hardware redundancy and analytical redundancy. Hardware redundancy Isermann (2006) is a classi- cal approach of fault diagnosis methods in control engeneering. Obvious disadvantages of using the hardware redundancy con- cept in robotics are higher costs, increased weight and complex- ity. Therefore the trend of current fault diagnosis techniques in robotics is based on the analytical redundancy Frank (1990) concept, where the mathematical model is used to determine a variable. The representation of a system model is different for various al- gorithms. Based on our experiments the parity space approach is applied only to state-space models, hidden Markov model and Abstract: The work presented in this paper focuses on the comparison of well-known and new techniques for designing robust fault diagnosis schemes in the robot domain. Correctly identifying and handling faults is an inherent characteristic that all autonomous mobile agents should possess, as none of the hardware and software parts used by robots are perfect; instead, they are often error-prone and able to introduce serious problems that might endanger both robots and their environment. Based on a study of literature covering model-based fault-diagnosis algorithms, we selected four of these methods based on both linear and non-linear models. We analyzed and implemented them in a mathematical model, representing a kinematics of four-wheel-OMNI mobile robot. Numerical examples were used to test the ability of three of the described algorithms to detect and identify abnormal behavior and to optimise the model parameters for the given training data. The final goal was to point out the strengths of each algorithm and to figure out which method would best suit the demands of fault diagnosis for a particular mobile robot. * Bonn-Rhein-Sieg University of Applied Sciences, Sankt Augustin, 53757, Germany (e-mail: anastassia.kuestenmacher@h-brs.de) ** (e-mail: paul.ploeger@h-brs.de) Anastassia Kuestenmacher * Paul G. Pl ¨ oger ** Model-Based Fault Diagnosis Techniques for Mobile Robots ⋆