Toward a framework for risk monitoring of complex engineering systems with online operational data: A deep learning-based solution

A mathematical architecture is developed for system-level condition monitoring. This architecture is built toward performing end-to-end operation risk and condition monitoring. The streaming monitoring data is given to the architecture as the input and system-level and component-level operation health states are computed as the output. This architecture integrates fault trees as the system-level modeling method and Deep Learning (DL) as the components condition monitoring method. A number of different deep learning models are trained using both operation and maintenance data for the components. Then, the fault tree fuses the continuous components’ assessments to provide system-level health insight. The applicability of this architecture is tested by implementing it on a real-world mining stone crusher system. This approach is extendable to dynamic risk assessment of complex engineering systems. However, DL models should be used with caution for safety-critical applications. We show that having DL models with high accuracy is not enough for trusting their predictions. We discuss the calibration of DL-based condition monitoring models and demonstrate how they can improve the trustworthiness and interpretability of DL models in risk and reliability applications.