A robust and reliable data-driven prognostics approach based on Extreme Learning Machine and Fuzzy Clustering

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