Deep evolutionary modeling of condition monitoring data in marine propulsion systems

In many complex industrial scenarios where condition monitoring data are involved, data-driven models can highly support maintenance tasks and improve assets' performance. To infer physical meaningful models that accurately characterize assets' behaviors across a wide range of operating conditions is a difficult issue. Usually, data-driven models are in black-box format, accurate but too complex to intelligibly explain the inherent physics of the process and lacking in conciseness. This study presents a deep evolutionary-based approach to optimally model and predict physical behaviors in industrial assets from operational data. The evolutionary modeling process is combined with long short-term memory networks, which are trained on estimations made by the evolutionary physical model and then used to predict sequences of data over a number of time steps. The likelihood of behaviors of interest is assessed by means of the resulting sequences of residuals, and a resulting score is computed over time. The proposed approach is applied to model and predict a set of temperatures related to a marine propulsion system, anticipating anomalies and changes in operating conditions. It is demonstrated that deep evolutionary modeling results are quite satisfactory for prognostics and obtained physical models are practical and easy to understand.