Towards Edge Anomaly Detection: Lightweight Multi-Resolution Spectral Features from High-Bandwidth Vibration Data

Reliable anomaly detection in rotating machinery depends not only on algorithms but also on the quality of the sensed signal. In vibration analysis, the sampling rate fundamentally governs the visibility of fault signatures. High rates capture critical high-frequency anomalies but at the cost of data volume and processing load, whereas low rates enhance frequency resolution but obscure higher-frequency content. Although wideband sensing offers clear advantages, it is often viewed as impractical on embedded platforms with limited computational resources. To address this challenge, This study introduces a multi-resolution frequency-domain feature extraction pipeline that leverages sequential downsampling and fixed-size FFTs to capture both low and high frequency fault signatures. By explicitly enriching the machine learning input space with complementary low and high frequency information, this method improves class separability and potentially strengthens downstream method. The method have been implemented on a microcontroller, achieving real-time operation while consuming only one-third of a one-second measurement window with negligible numerical error. These findings demonstrate that multi-resolution spectral features not only unlock the benefits of wideband vibration sensing on embedded platforms but also provide machine learning models with richer and more discriminative inputs, enabling lightweight anomaly detection at the edge.