The recent dependencies studies have mainly been done in context of single unit, which means the failures involving components in different units are not explicitly treated. The major scopes are also limited to the simultaneous failures as the direct results of shared causes, namely the Common Cause Failures (CCFs). The causal relations among components or units are rarely addressed. Thus it’s the prime topic of this investigation that the dependencies among multiple units co-located at a site which is called multi-unit dependencies.
This paper seeks to propose a hybrid approach by combining physics-based models and supervised learning techniques. The essential idea is to account for the multi-unit dependencies by explicitly modeling the interactions of the underlying physical failure mechanisms. Ultimately a hybrid Dynamic Bayesian Belief Network is developed to model the possible dependencies, and supervised learning techniques are adopted to quantify the likelihood of failures due to dependency effects. Furthermore, an experiment has been designed and presented involving redundant pumping system, the performance of which are monitored by an advanced sensing system. The experiment is now operating and the gathered multi-sensor data will be used to illustrate the proposed approach as the next stage of this research. These sensor data should be of good quality to allow revealing the underlying failure behavior and dependent failures. This study provides an understanding of the inherent risk significance of dependencies among multiple units, and can also work as the basis for the reliability of multi-unit systems where causal dependencies play a relevant role.
