Adaptive Generalized FRESH Filtering for Separating Composite Fault Signals Under Varying Speed Conditions

Extracting feature signals is a crucial step in mechanical fault diagnosis. However, separating feature signals of composite faults under varying speed conditions poses a challenging problem. In existing approaches, inner product matching methods often exhibit poor noise robustness, while methods relying on statistical feature extraction struggle to achieve real-time fault signal separation under variable speeds. To address these challenges, this article introduces a novel adaptive filtering structure—adaptive generalized FRESH (AG-FRESH) filtering—for fault feature extraction. This filter leverages the generalized frequency-shift (FRESH) correlation between different frequency components of angle-time cyclostationary (AT-CS) signals to effectively separate composite weak features. The method requires only knowledge of the cyclic frequency of features to achieve blind extraction of fault signals. Moreover, through adaptive filtering, this approach supports real-time deployment and adapts to instantaneous statistical characteristic variations in vibrations. Numerical simulations and experimental analyses confirm the superiority of the proposed method in extracting high-frequency and subtle features, particularly when a separate analysis of composite faults is necessary.