Intelligent Monitoring and Remaining Useful Life Estimation of Industrial Systems

In the last decade, industrial manufacturing and worldwide companies are evolving towards Industry 4.0, which integrates fully connected and optimized cyber-physical systems (CPS). This evolution ultimately results in increased reliability, availability, safety and reduced maintenance costs and poses many challenges in systems' health management. Prognostics and health management (PHM), one of the challenges posed by the revolution of Industry 4.0, aims at extending the life cycle of engineering assets while improving production performances and reducing unplanned downtime losses. Prognostics is the PHM's most critical key feature with future capabilities. It is defined as the estimation of remaining useful life (RUL) during which a critical component, subsystem, or an entire system performs its intended functionalities. An accurate RUL estimation can avoid catastrophic failures, maximize operational availability and reduce maintenance costs. A various number of researches related to prognostics approaches have been proposed and categorized into three classes: (i) Physics-based, (ii) Data-driven and (iii) Hybrid approaches. Each category has its benefits, limitations and issues. This thesis aims at deploying intelligent monitoring approaches to estimate the RUL of critical components and complex systems in the context of Industry 4.0. However, approximating the behavior of a critical component and performing prognostics is more feasible than working directly at an industrial level. Thus, system-oriented PHM approaches are challenging and still quite difficult in practice. Due to the difficulty of representing the mode failure of complex systems using physics-based models derived from first principles, data-driven and hybrid approaches are the most appropriate to address this challenge. Hence, this thesis proposes a novel hybrid and data-driven prognostics approach to estimate the RUL of critical components and complex systems. The approach consists of transforming the data provided by simulations and sensors installed on the component/system into relevant behavioral models that represent the time evolution of the degradation phenomenon. To do this, some issues of data-driven and hybrid approaches are highlighted: 1) Are there enough historical data that can represent the failure mechanism? 2) Is the critical component to monitor accessible in order to attach the sensors around? If it is not, how is it possible to collect the data? 3) Which parameters should be monitored and where should the sensors be placed for data collection? 4) How can the raw data be transformed into a suitable set of indicators that can reflect the evolution of degradation? 5) How can the model be chosen so that it can fit the data while covering all phenomena associated with the degradation mechanism? Such issues constitute the problems addressed in this thesis. The main contributions are as follows: - Data processing step and health indicator construction are improved by introducing a new approach for features' selection called suitability, where features' selection is based on three characteristics, i.e., monotonicity, trendability and predictability. - Health indicator (HI) construction is improved by a new approach for features selection based on Linear degradation models. The main idea is to fit every feature with a linear degradation model and to select those with large slopes for further analysis. - RUL estimation thanks to degradation models. The performance of the prognostics is enhanced thanks to the constructed HIs. - RUL estimation using deep convolutional neural networks (CNN). The performance of the prognostics is enhanced thanks to the extraction of the features using deep networks. - RUL estimation using similarity-based approaches. The performance of the prognostics is enhanced based on the similarity measure between the constructed HI from the training and test datasets. The developed approaches are validated on simulated data from broken rotor bar, on an experimental dataset from an industrial testbed of an Assembly replica plant and on the C-MAPSS NASA datasets.