Availability and usability of data coming from a process/environment, e.g., those generated by a sensor network, introduce serious issues about their quality. In fact, not rarely acquired measurements are affected by sensor aging and faults which might introduce errors impacting on the correctness of the subsequent decision making process. The ability to detect faults is a mandatory step, which cannot be underestimated or neglected in real deployments.
In this direction, Fault Diagnosis Systems (FDS) are tools designed to supervise a process operation in order to detect, isolate and identify potential faults and, possibly, design accommodation actions.
However, most FDS assume that some of -not necessarily amenable- hypothesis are satisfied, e.g., a description for the process is available; the system model is linear; a fault dictionary containing the fault signatures is provided; the nature of the fault profile and its development are known.
Current research in machine learning aims at removing/weakening the above assumptions so that FDS can be designed directly from available data, possibly within a cognitive framework.
The talk will focus on aspects related to the design of cognitive FDSs for sensor networks able to discriminate between faults, changes in the environment and model bias within an evolving framework.
