Aero-engine bearing anomaly detection framework based on adaptive wavelet transform network and neighborhood angle factor analysis

The reliability of aero-engines, the core power units of modern aircraft, is paramount for flight safety. However, the harsh operating environment—characterized by high speeds and variable loads—makes rolling bearings susceptible to faults that are difficult to detect using traditional methods. Existing signal processing techniques rely heavily on manual parameter tuning, while purely data-driven deep learning models often struggle to extract stable features under fluctuating operating conditions. To address these challenges, this paper proposes AWTN-NAFA, a novel unsupervised anomaly detection framework integrating an Adaptive Wavelet Transform Network (AWTN) and Neighborhood Angle Factor Analysis (NAFA). The AWTN module embeds the physical mechanism of the fast discrete wavelet transform into a Convolutional Neural Network (CNN) architecture. By employing learnable filters and activation functions, AWTN adaptively extracts compact, stable, and discriminative latent representations that remain robust against domain shifts. Subsequently, the NAFA module performs anomaly detection on these features. Unlike traditional distance-based detectors sensitive to feature scale, NAFA exploits the geometric properties of the data manifold, utilizing the angular variance of nearest neighbors to construct a robust, adaptive threshold. Extensive experiments on a real-world aero-engine bearing dataset demonstrate that AWTN-NAFA significantly outperforms state-of-the-art baselines, achieving superior detection accuracy and robustness across diverse complex operating conditions. © 2026 Elsevier Masson SAS.