Probabilistic Deep Learning With Bayesian Networks for Predicting Complex Engineering Systems' Remaining Useful Life: A Case Study of Unmanned Surface Vessel

Abstract

Remaining useful life (RUL) serves as a key indicator of system health, and its accurate and timely prediction supports informed decision-making for efficient operation and maintenance. This is essential for complex engineering systems (CESes) such as unmanned surface vessels (USVs), where the human operators have limited opportunity to intervene during the operation. This paper proposes a framework for predicting the RUL of the CESes. The proposed framework employs a probabilistic deep learning (PDL) approach to predict the component's RUL and an equation node-based Bayesian network (BN) to predict system RUL (SRUL) at any future time-step. The component-level RUL method is validated using the NASA's Commercial Modular Aero-Propulsion System Simulation (c-mapss) dataset, and then the proposed framework is demonstrated with a USV case study. The results are evaluated using a set of quality metrics. By making use of the condition-monitoring sensor data, component reliability data, and models that account for the complex causal relationships between components, the proposed framework can provide near real-time predictions of the RUL with uncertainty of a CES, thus supporting its informed decision-making during the operation.