Fault diagnosis employing probabilistic models and statistical learning

Abstract

A computer implemented fault diagnosis method employing both probabilistic models and statistical learning that diagnoses faults using probabilities and time windows learned during the actual operation of a system being monitored. In a preferred embodiment, the method maintains for each possible root cause fault an a-priori probability that the fault will appear in a time window of specified length as well as maintaining—for each possible resulting symptom(s)—probabilities that the symptom(s) will appear in a time window containing the fault and probabilities that the alarm will not appear in a time window containing the fault. Consequently, the method according to the present invention may advantageously determine—at any time—the probability that a fault has occurred, and report faults which are sufficiently likely to have occurred. These probabilities are updated based upon past time windows in which we have determined fault(s) and their cause(s). Advantageously, each root cause fault may be assigned its own time window length. By maintaining these probability parameters for several different window lengths, a window length that is particularly well-suited to a particular set of conditions may be chosen.