This study introduces a novel decompose and distilled informer (D2-Informer) framework designed to address the challenge of anomaly detection in highfrequency, long-sequence vibration data from aircraft engines. By employing the variational mode decomposition (VMD), the model effectively decomposes complex vibration signals into intrinsic frequency components, isolating noise and enhancing the detection of key features. Building upon this, the Wav2Vec-inspired convolutional module compresses the decomposed signals, reducing sequence length while retaining critical temporal patterns. This operation ensures that essential features are preserved, aiding in precise anomaly classification. The integration of the informer network enables the capture of long-range temporal dependencies using an advanced self-attention mechanism, optimizing the model’s ability to detect anomalies even in highly imbalanced and complex datasets. The D2-Informer framework excels in both short-term and long-term sequence processing, ensuring high precision and recall in real-world fault diagnosis applications. This model is computationally efficient, requiring minimal retraining, making it suitable for deployment in various industrial contexts. The experimental results demonstrate that the D2-Informer not only outperforms traditional anomaly detection methods but also adapts seamlessly to diverse datasets, offering a robust solution to high-dimensional data challenges without significant computational overhead.

Rapid and accurate multi-class fault diagnosis is essential for high-end equipment because different fault categories require different maintenance responses. This study aims to develop a lightweight and discriminative diagnostic framework that can identify multiple fault categories from non-stationary vibration signals while reducing redundant computation. High-frequency vibration signals provide direct condition information, but long sequences, noise, nonlinear dynamics, and non-stationary behavior make raw-signal classification unreliable. From an entropy-based information-processing perspective, the key issue is to separate informative fault modes from redundant fluctuations and enlarge inter-class distinctions in the probabilistic decision space. This study proposes Fastformer, an integrated framework for vibration-based fault identification. Empirical Mode Decomposition first converts each signal into Intrinsic Mode Functions to reduce modal mixing and preserve fault-related oscillatory components. The resulting components are processed by an encoder-oriented Q/K/V dot-product scoring mechanism, which constructs compact spatiotemporal embeddings without adopting a complete Transformer architecture. Validation-guided pruning removes low-contribution attention responses, while a Margin-Enhanced Fault Softmax classifier optimized with a cross-entropy-based objective strengthens category separation. By combining stable decomposition, lightweight attention scoring, pruning, and probabilistic margin learning, Fastformer achieves faster and more stable convergence. On the XJTU-SpurGear dataset, Fastformer obtains precision, recall, F1-score, and AUC values of 1.000. Additional validation on the HUST bearing dataset further shows that Fastformer achieves the best overall performance among the compared methods, with an AUC value of 0.9596.