Operation Condition Monitoring using Temporal Weighted Dempster-Shafer Theory

Abstract

System operation is a real time, dynamic decision process, a continuous observation should be implemented to support timely decision. Real time condition monitoring and diagnosis is featured with ongoing event sequence. The more recent observation, the much detailed, accurate information, and the more obsolete observations with much weak correlation to current faults and errors vise versa.

Dempster-Shafer evidence theory is best suitable for the problem of redundant sensors, insufficient data reasoning. However, D-S base applications largely focused on causational relationship between symptoms and effects, and the fusion process of evidences was performed regardless whatever order observed. As an improvement to the frame of discernment of the D-S theory, we purposed a time weighted evidence combination method. Observed events were extracted from multiple time points to form a temporal evidence sequence. Basic probability assignment was altered by temporal weights in accordance with the time proximity between the observed events and current time. The temporal weights value set was in accordance with its occurring time point. Evidences with same timestamps should be allocated with the same temporal weights. An example was discussed to illustrate the temporal weight, D-S rule based assessment framework. In the framework, latest observed evidences stream were combined into the framework to improving fault recognition.