Abstract:
In industrial big-data applications, the ability to transform very complex sensor streams into accurate maintenance information is crucial. The challenge with Remaining Useful Life prediction remains, however, because of the non-linearity of degradation, the fact that operating conditions are often variable, and the balance between model performance and computational time. This paper introduces a predictive intelligence framework for RUL prediction of turbofan engines that is scalable and optimized. The method consists of capped RUL target construction, multivariate sensor preprocessing, baseline predictive modeling, XGBoost learning, and hyperparameter optimization through Random Search, Optuna/Bayesian Optimization, Genetic Algorithm and Particle Swarm Optimization. NASA C-MAPSS FD001 and FD004 were used for the experiments; FD001 was employed for baseline development and FD004 was employed for complex environment validation. The results indicated that the non-linear model was better than the statistical regression. The use of the Random Search optimized XGBoost gave the lowest error RMSE 15.566 and R2 0.682 for FD001 and low but acceptable errors RMSE 16.621 and R2 0.574 for FD004 for XGBoost. The test of scalability showed that the RMSE was not affected by increasing amount of FD004 data, while robustness testing showed sensitivity to sensor noise. The novelty of this work is the reproducible mathematical analytics workflow for optimized and scalable industrial RUL prediction for practical industrial decision support.