Aero-engine gas path anomaly monitoring based on unsupervised structure and deep slow feature representation

Abstract Gas path anomaly monitoring holds a crucial position in aero-engine health management due to the dynamic nature of gas path parameters, data imbalance, and the lack of labels, presenting significant challenges. To address these issues, this study proposes a novel method for dynamic anomaly monitoring in aero-engines utilizing Kernel Slow Feature Analysis (KSFA) and Deep Support Vector Data Description (Deep SVDD). In this approach, the original gas path parameter values undergo preprocessing using the KSFA algorithm to extract pertinent features indicative of gradual changes in gas path status. The Deep SVDD model, employing a one-dimensional Convolutional Neural Network (1D-CNN) fused with a feature attention mechanism, is iteratively trained to identify the optimal hypersphere. The Health Indicator (HI) is then determined by quantifying the distance between the test set and the hypersphere's center, enabling a quantitative assessment of the aero-engine's performance degradation. Experimental findings demonstrate that this method outperforms alternative evaluation techniques by effectively tracking the aero-engine's degradation process and anticipating engine anomalies, showcasing its practical value in engineering applications.