A Real-Time Framework for Speckle-Induced Spike Suppression in High-Sample-Rate LDV Measurements

Abstract

Enabling non-contact measurement of vibration and velocity, the laser Doppler vibrometer (LDV) is widely used across diverse fields. However, in LDV measurements, the occurrence of speckle noise is almost inevitable and induces spikes as it undergoes subsequent signal processing steps. Although various signal processing and deep learning-based methods have been investigated to address this issue, the application of those methods is limited in scenarios where numerous consecutive data points are affected by spikes or where spikes are present in succession. This study proposes a three-stage framework that combines the Modified Hampel filter for initial spike suppression, the Temporal Convolutional Network (TCN) for signal reconstruction, and a low-pass filter for final signal smoothing. Through experiments on LDV-measured vibration-velocity signals, we demonstrated that the proposed method operates in real-time under high-sampling-rate conditions and can efficiently reconstruct LDV signals by suppressing spike components.