Drift Detection and Characterization for Fault Diagnosis and Prognosis of Dynamical Systems Antoine Chammas, Moamar Sayed-Mouchaweh, Eric Duviella, and St´ ephane Lecoeuche Univ Lille Nord de France, F-59000 Lille, France EMDouai, IA, F-59500 Douai, France {antoine.chammas,moamar.sayed-mouchaweh,eric.duviella, stephane.lecoeuche}@mines-douai.fr Abstract. In this paper, we present a methodology for drift detection and char- acterization. Our methodology is based on extracting indicators that reflect the health state of a system. It is situated in an architecture of fault diagno- sis/prognosis of dynamical system that we present in this paper. A dynamical clustering algorithm is used as a major tool. The feature vectors are clustered and then the parameters of these clusters are updated as each feature vector arrives. The cluster parameters serve to compute indicators for drift detection and char- acterization. Then, a prognosis block uses these drift indicators to estimate the remaining useful life. The architecture is tested on a case study of a tank system with different scenarios of single and multiple faults, and with different dynamics of drift. Keywords: Fault Diagnosis, Prognosis, Drift, Dynamical Clustering. 1 Introduction Incipent faults are undesirable changes in a process behavior. The affected process state passes from normal to failure through an intermediate state where it is in degraded mode. This degraded mode is also called faulty mode [6]. Condition-Based Maintenance (CBM) was introduced to try to maintain the correct equipment at the right time. It enables the pre-emptive maintenance of systems that are subject to incipient faults. For this purpose, CBM uses a supervision system (diagnosis/prognosis) in order to determine the equip- ments health using real-time data [16,26], and act only when maintenance is actually nec- essary. Prognosis is the ability to predict accurately the Remaining Useful Life (RUL) of a failing component or subsystem. Prognostics and Health Management (PHM) is the tool used by a CBM system. The reason is that a PHM provides, via a long-term prediction of the fault evolution, information on the time where a subsystem or a com- ponent will no longer perform its intended function. In order to estimate the RUL, it is important to accurately determine the fault conditions: detection of fault and isolation of fault. Moreover, it is important to continuously determine the current condition state of a process when operating online. These requirements are achieved by a diagnosis module. Approaches for diagnosis have been largely studied in the literature. Globally, we can cite model-based approaches and data-driven based approaches. The model-based E. H¨ ullermeier et al. (Eds.): SUM 2012, LNAI 7520, pp. 113–126, 2012. c  Springer-Verlag Berlin Heidelberg 2012