A robust & reliable Data-driven prognostics approach based on extreme learning machine and fuzzy clustering.
Kamran Javed · INRIA a CCSD electronic archive server · 2014
Prognostics & Health Management (PHM) aims at extending the life cycle of an engineering asset, while reducing exploitation and maintenance costs. For this reason, prognostics is considered as a key process with future capabilities. Indeed, accurate estimates of the Remaining Useful Life (RUL) of equipment enable defining further plan of actions to increase safety, minimize downtime, ensure mission completion and efficient production. Recent advances show that data-driven approaches (mainly from machine learning) are increasingly applied for fault prognostics. They can be seen as black-box models that learn the system behavior directly from Condition Monitoring (CM) data, use that knowledge to infer its current state and predict future progression of failure. However, approximating the behavior of critical machinery is a challenging task that can result in poor prognostics. As for understanding some issues of data-driven prognostics modeling, consider the following points. 1) How to effectively process raw monitoring data to obtain suitable features that clearly reflect evolution of degradation? 2) How to discriminate degradation states and define failure criteria (that can vary from case to case)? 3) How to be sure that learned-models will be robust enough to show steady performance over uncertain inputs that deviate from learned experiences, and to be reliable enough to encounter unknown data (i.e. operating conditions, engineering variations, etc.)? 4) How to achieve ease of application under industrial constraints and requirements? Such issues constitute the problems addressed in this thesis and have led to develop a novel approach beyond conventional methods of data-driven prognostics. Main contributions are as follows. - The data-processing step is improved by introducing a new approach for features extraction using trigonometric and cumulative functions, where features selection is based on three characteristics, i.e., monotonicity, trendability and predictability. The main idea of this development is to transform raw data into features that improve accuracy of long-term predictions. - To account for robustness, reliability and applicability issues, a new prediction algorithm is proposed: the Summation Wavelet-Extreme Learning Machine (SWELM). SW-ELM ensures good prediction performances while reducing the learning time. An ensemble of SW-ELM is also proposed to quantify uncertainty and improve accuracy of estimates. - Prognostics performances are also enhanced thanks to the proposition of a new health assessment algorithm: the Subtractive-Maximum Entropy Fuzzy Clustering (S-MEFC). S-MEFC is an unsupervised classification approach which uses maximum entropy inference to represent uncertainty of unlabeled multi-dimensional data and can automatically determine the number of states (clusters), i.e., without human assumption. - The final prognostics model is achieved by integrating SW-ELM and S-MEFC to show evolution of machine degradation with simultaneous predictions and discrete state estimation. This scheme also enables to dynamically set failure thresholds and to estimate the RUL of monitored machinery. Developments are validated on real data from three experimental platforms: PRONOSTIA FEMTO-ST (bearings test-bed), CNC SIMTech (machining cutters), C-MAPSS NASA (turbofan engines) and other benchmark data. Due to realistic nature of the proposed RUL estimation strategy, quite promising results are achieved. However, reliability of the Prognostics model still needs to be improved which is the main perspective of this work.