EXPERIENCES ON ACTUATOR FAULT DETECTION, DIAGNOSIS AND ACCOMMODATION FOR ROVS M. Caccia, R. Bono, Ga. Bruzzone, Gi. Bruzzone, E. Spirandelli, G. Veruggio Consiglio Nazionale delle Ricerche - Istituto Automazione Navale Via De Marini, 6 16149 Genova, Italy E-mail: { max, ric, gabry, gio, edo, veruggio } @ ian.ge.cnr.it Abstract This paper addresses the problem of low cost actuator fault detection, diagnosis and accommodation for Remotely Operated Vehicles. The research is based on the analysis of the telemetry of Romeo, the over- actuated ROV developed by C.N.R.-I.A.N., operating both in nominal and different failure conditions. Results demonstrate how the monitoring of the servo- amplifiers' I/O variables enables the detection and diagnosis of actuator faults for operating ROVs, supporting the pilot in making decisions on the real- time reconfiguration of the propulsion system. The integration of the servo-level FDDA module with a conventional navigation, guidance and control system is discussed. On the basis of experienced time-varying effects of operating failures, suitable models of the damaged actuators and algorithms for the generation of fault symptoms and alarms have been designed, implemented and satisfactorily tested on a large amount of recorded ROV data. The conventional fault accommodation procedure based on the reconfiguration of the vehicle's thrust control matrix has been applied. The capability of Romeo of working in a reduced actuation configuration has been operationally demonstrated executing shallow water benthic missions for scientific users. Introduction Currently, unmanned underwater vehicles (UUVs) can satisfactorily perform a large amount of scientific missions for oceanographic surveys, exploration of unknown areas and operations in the proximity of the seabed or man-made underwater structures. As discussed in [1], at this stage the main technical issue is to improve the system reliability in order to guarantee the mission is accomplished even in the presence of some variations from the nominal conditions, i.e. faults. In this sense, fault management systems for UUVs able to detect, isolate and accommodate faults, have been investigated and designed in order to improve the vehicle degree of autonomy, reliability and capability of performing missions in harsh environments. UUVs’ operations can be affected by two kinds of faults: failures in the hardware and software subsystems of the vehicle, such as thruster seizing or sensor breaking, and unforeseen environmental conditions, which could jeopardize the working mode of the sensors measuring the vehicle-environment interactions, such as multi-path phenomena affecting echo-sounders’ performances. In the past, attention focused on the vehicle faults, distinguishing between actuator and sensor faults, and trying to classify the expected system failures on the basis of their probable response as reported in [2], where zero-response breaking, permanent offset jamming, intermittent slipping, decreasing response creeping failure and random control failure damages have been listed. In the nineties, theoretical advancements in fault diagnostics obtained in various fields of applications, from aeronautical and aerospace systems to nuclear and chemical plants, have been transferred to underwater robotics. Since unmanned underwater vehicles can be modeled by a small number of state variables, model- based techniques for fault detection have been preferred to model-free ones. In particular, fault detection techniques based on parameter estimation and analytical redundancy, i.e. the same quantity is calculated by combining data from different sensors or model-based predictors, have been developed. An example of black-box identification combined with gradient filtering to generate failure events is given in [2], while a technique to manage analytical redundancy through expert systems’ rules is presented in [3]. In [1] and [4], a model-based observer is used to generate a residual between the sensor measured values and that predicted from the model in order to detect dynamic faults. A fault is detected when the residual magnitude is higher than a suitable threshold for more than a fixed length of time. The same method to avoid false alarms, after processing residuals generated through analytical redundancy, is used in [5], where an accommodation technique for thruster failures is also presented in the case of an over-actuated vehicle. Modeled actuator faults can be detected and isolated through a bank of extended Kalman filters or sliding-mode observers as discussed in [6] and [7]. Due to the difficulties of having available recorded telemetry data of UUV actual faults, the above mentioned investigations are based on simulations of plant damages of the types