Adaptative data-driven approach for the remaining useful life estimation when few historical degradation sequences are available
Résumé
Fault prognostics is the estimation of the Remaining Useful Life (RUL) of a component until failure. It is a main part of the predictive maintenance strategy that can help to enhance the reliability and availability of industrial systems while reducing unscheduled downtime and maintenance cost. Applying fault prognostics to industrial systems is difficult because very few sequences about degradation are available. Hence, this paper proposes a new approach able to perform the fault prognostics when few historical degradation data are available. In offline, a library of Health Indicators (HIs) and library of adaptive models are defined. Then, the best HI-model pairs according to the accuracy of the a priori sequences are selected. In online, several RULs are predicted for the new sequence using the selected pairs. After that, the final RUL is computed by merging the RULs using a weighted mean. The approach is validated using a degradation scenario selected from an aircraft engine degradation dataset (CMAPSS dataset). The obtained results are promising compared to some well-known similar state-of-the-art approaches.
Mots clés
Fault prognostics
Remaining Useful Life
artificial intelligence
failure
machine learning
predictive maintenance strategy
reliability
availability
unscheduled downtime
maintenance
degradation
fuzzy logic
Health Indicator
library
adaptive model
degradation scenario
aircraft engine degradation
Degradation
Adaptation models
Predictive models
Aircraft propulsion
Historical Sequence
Health Indicators
Industrial Systems
General Linear Model
Artificial Neural Network
Systemic Conditions
Good Accuracy
Recurrent Neural Network
Power Factor
Exponential Model
Sensor Measurements
Support Vector Regression
Failure Time
Quadratic Regression
Sensor Values
Means In Order
Threshold In Order
Exponential Trend
Degradation Speed
Double Exponential Model
Failure Threshold
Turbine Engines
Prediction Error
Highest Accuracy
Present Time
Neural Network
Time Index
Origine : Fichiers produits par l'(les) auteur(s)