Vibration signals from rotating bearings exhibit strong periodicity, very high sampling rates, and subtle spectral variations, which distinguish them from general time-series data. These characteristics make the transformer model based on time-domain ...
Vibration signals from rotating bearings exhibit strong periodicity, very high sampling rates, and subtle spectral variations, which distinguish them from general time-series data. These characteristics make the transformer model based on time-domain inefficient, as they hard to learn rotation-level patterns or require excessively long input sequences. In addition, although anomaly attention detects anomalies by modeling relationships between time points, it is difficult to distinguish normal and abnormal behavior at individual time steps in high-resolution vibration signals. This study aims to address these limitations and develop an early bearing fault detection method optimized for transformer-based anomaly detection. In this study, we propose a preprocessing framework that generates frequency-domain sequences suitable for transformer architectures by segmenting the signal into rotational units and converting each segment into a frequency-domain vector. The proposed method applies a sliding-window processing based on rotation unit, performs a fast Fourier transform (FFT) to obtain its spectral representation, and constructs token sequences that allow anomaly attention to learn long-term and inter-rotation relationships efficiently. This approach captures rotation-level characteristics and long-term dependencies without requiring extremely long sequences, while also reducing computational cost. The model computes anomaly scores using reconstruction error and the discrepancy between prior and series associations, which aligns more closely with both the properties of bearing vibration data and the structural intent of anomaly attention. Experimental results show that the proposed method provides more stable and improved anomaly detection performance compared with the time-domain anomaly transformer, demonstrating its effectiveness for early fault detection in rotating machinery.