ABSTRACT
In many applications, detection methods are implemented to monitor systems. Detector may be built upon analytical model of the system behaviors, or based on recorded data. Lately lots of efforts have been put in developing methods that enable to transfer a detector from an assessed system to a new one when the 2 systems are similar. These so-called transfer learning approaches have shown efficiency when dealing with a fleet of systems composed of different similar systems. For the detector of each model of system a tradeoff between false alarm and non-detection must be chosen. One question that has not retained much attention up to now is the tuning of these detection systems. The common practice is to tune the detection system associated with each model individually. In this communication we tackle the problem of the global efficiency of this practice. How can we tune all those detection systems together so that the group of detectors satisfies some performance constraint? The formalization of the problem is introduced, discussed and commented based on detector ROC curves. Three toy examples are used to show the gain one could expect from join optimization of detectors and to illustrate optimization issues. Example results show that non detection probability can easily be reduced by up to 50%. In conclusion several extensions of this work are discussed.
