Analysis of fault diagnosability from SCADA alarms signatures using relevance indices

When a fault occurs on a complex process, it generally sets off many alarms from the SCADA system. The list of alarms raised by the control system from the occurrence of the fault until the flow of alarms is stabilized forms the fault signature and is the raw material used by operators to make their diagnosis. Some alarms may be raised on the occurrence on several faults while others may be specific to a given fault and thus be very useful to make the diagnostic. In order to assist the operator in his diagnostic process and help him better understand the alarm generation process, a method to evaluate the ability to discriminate a fault from a set of other faults from its alarms signature is proposed. The method is data driven and makes use of a set of alarm lists recorded during the occurrence of several faults. An index of relevance is calculated for each alarm and for each fault. An alarm is relevant to a fault compared to a set of other faults if it is very often (or never) raised when the fault occurs while it is seldom (or often) raised when the other faults occur. The index of relevance can be calculated in different contexts by changing the set of other faults the fault is compared to. Strategies to evaluate fault diagnosability using this index are proposed in this paper. They are validated on a set of alarm lists recorded during different fault situations and obtained by means of a very accurate simulator of a CERN LHC system.