A spectrum-guided asymmetric window-based fault prediction method

To address the challenges of limited anomalous samples and the difficulty of jointly modeling multi-scale temporal patterns in unlabeled log data, a spectrum-guided asymmetric window framework is proposed for system fault prediction. Specifically, the dominant periodic characteristics of time series are first identified using the discrete Fourier transform (DFT), based on which a Bayesian optimization strategy is developed to adaptively determine the window lengths of the teacher and student models. A teacher–student architecture is then established, where the teacher model integrates an Autoencoder (AE) with a Gaussian Mixture Model (GMM) to generate continuous anomaly intensity-based soft labels, while the student model adopts a Bidirectional Long Short-Term Memory (BiLSTM) network to capture long-range temporal dependencies. Furthermore, a temporal boundary alignment mechanism is designed to facilitate effective cross-scale knowledge distillation between the teacher and student models. Extensive experiments conducted on the SMD, PSM, SMAP, and MSL datasets show that the proposed method achieves an average F1-score of 88.49%, consistently outperforming several state-of-the-art approaches.