Addressing Class Imbalance in Aeroengine Fault Detection

Data-driven approach to fault detection in aeroengine has received considerable attention owing to the availability of engine sensor information for locating and classifying faults. In building models for aeroengine fault detection, class imbalance is a prominent issue. This is due to the fewer number of faults from operational flights and the cost of acquiring them. Notable resampling techniques have been proposed in addressing this class imbalance dilemma. The process involves small incremental iteration of resampling ratios (i.e., the ratio of the number of samples in the minority class over that of the majority class) on a machine learning classifier in a broad search-space, until an acceptable performance is reached. Finding the ideal resampling ratio that guarantees an optimal model performance is particularly necessary in cases where the class samples are distributed in close neighbourhood. However, the process incurs considerable resource expense. This study undertakes an investigation into resampling techniques for tackling class imbalance in aeroengine fault detection. Four ensemble tree learners were considered, while examining the influence of different resampling ratios on the model performance. In determining the best resampling ratio, we propose a lightweight, two-step approach that iteratively locates the search-space that guarantees the optimal model performance. The experimental findings showed a decrease of up to 83.5% in the computational cost in determining the resampling ratio, along with a notable improvement of up to 5.5% in model performance.