A Multi-Interval Method for Discretizing Continuous-Time Event Sequences

Prognostic and health management (PHM) techniques are being developed to support health assessments of increasingly diverse and complex systems. Causal models, such as Dynamic Bayesian Networks (DBNs), are used in PHM to visualize dependencies between system components as well as to propagate new information for updated system health assessments. DBNs rely on operational data from multiple sources to update their estimates. DBNs are discrete models, but many data sources (including sensors and system monitors) provide near-continuous streams of data. Information from these data needs to be implemented in discrete models like DBNs to update health assessments. The discretization technique most commonly used has challenges responding to changing data needs due to dynamic environmental and operational mode shifts; this paper proposes a new approach for responsively converting continuous operational data feeds into discrete information intervals to support building DBNs for PHM of complex engineering systems (CESs). The proposed discretization method is demonstrated on a PHM model for a nuclear reactor experiencing a scram failure after an accident. Results from the case study indicate models using this approach for identifying information intervals may be more responsive to changing data needs. Multi-interval models reduce computational requirements for updating assessments and storing data, and provide greater insight for PHM assessments and maintenance decisions in new operational environments. Results from the case study showed that the proposed approach addressed the limitations of the current discretization method. Adjusting how much information the model receives when the system transitions into an accident scenario makes models flexible in different operational environments and responsive to current needs for increased data availability. These models can support better PHM assessments based on data from the current operational environment, leading to improved decisions for maintaining and prolonging system life.

Automated summary

Prognostic and health management (PHM) techniques are being developed to support health assessments of increasingly diverse and complex systems. Causal models like Dynamic Bayesian Networks (DBNs) visualize dependencies between system components and update system health assessments. A new approach is proposed to convert continuous operational data feeds into discrete intervals for building DBNs in PHM.