PREDICTIVE MAINTENANCE OF AIRCRAFT ENGINES USING MACHINE LEARNING APPROACHES

Predictive maintenance involves a clear detection of the machine degradation to avoid unnecessary downtime and lower the expenses of operation. This paper introduces a predictive maintenance model that is built on the basis of NASA CMAPSS dataset to categorize the operating conditions as healthy, warning, and imminent failure condition. A hypergraphbased method of class balancing is proposed to maintain the structural relations in the minority samples. The permutation importance is used when selecting the sensors to keep the most informative ones. Three ensemble classifiers, namely AdaBoost, Random Forest, and XGBoost are
compared in terms of various metrics. XGBoost achieves the highest accuracy of 84.67%, followed by Random Forest with 83.83%, and AdaBoost with 82.17%. XGBoost yields the strongest F1 performance for the Healthy and Warning states, while Random Forest provides the highest F1-score for the Imminent Failure state. SHAP and LIME are combined to have universal and specific interpretability. The suggested methodology provides accuracy, along with efficiency,and explainability appropriate to industrial predictive maintenance systems in the real world.