Abstract
Keywords - Health Monitoring, Diagnosis, Prognosis, Non linear statistics, Formula One
Abstract - MIRIAD Technologies setup a software component allowing to diagnose on the fly mechanical hazards on a critical part of the engine of a RENAULT F1 Team formula one car. The technical point is that during a race, this kind of mechanical event is only recorded by under-sampled sensors. At such a low rate of acquisition, the symptoms of a soon failure of the piece are not known to be recognized. Moreover the failure has to be detected before it really occurs and damages the engine.
This paper exposes MIRIAD Technologies approach. It is based on :
computation of a statistical estimate of characteristics of the original high rate signals ;
contextual computation of a large amount of statistically meaningful signals (more that 300) that describe the mechanical situation of the engine part;
automatic selection of a subset of those descriptors containing valuable information and computation of a «normality score» that quantify the health of the mechanical parts
monitoring of this health score to emit a warning if it reaches critical zones or if a disturbing trend appears.
The first point strongly relies on theoretical results that can be used to estimate the underlying value of under sampled mechanical signals over specific conditions (1), (2) and (3). All the monitoring system is based on the accuracy of this estimate. The second point is a multi scale approach that belongs to the wavelet analysis family (5), (6). Its contextual part comes from the use of quantities like engine speed and load to automatically build and select homogeneous states. The third point uses information theory (Shannon (7) and Kolmogorov (8)) to keep only the part of signals containing a priori information. All this information is mixed into a normality score that has a probabilistic meaning : the more this score is close to zero, the more the current situation is likely to be a normal one ; and the more it is close to one, the more the situation is a normal one. The last point uses trend detection on the real time variations of the health score.
All those technologies have been actually validated off line on recorded datasets during races ; MIRIAD Technologies is considering a real time implementation. It is to be noticed that the first point has been validated comparing the estimate (coming from the under sampled signal) and the underlying one, really recorded on test beds. Those results have a great potential for online health monitoring of mechanical car parts. This way, such sensors with a poor acquisition rate can be used to monitor pieces subject to high level of vibrations.