PSI - Issue 77

Francisco Afonso et al. / Procedia Structural Integrity 77 (2026) 584–592 F. Afonso et al. / Structural Integrity Procedia 00 (2026) 000–000

585

2

1. Introduction

Structures gradually deteriorate over time, impacting performance or even causing failure, which, if not properly monitored, can be unexpected and potentially harmful. Vibration-based Structural Health Monitoring (SHM) exploits the link between gradual changes in a structure’s physical parameters and its vibration characteristics, which serve as a primary indicator of a structure’s health [1]. This also applies to smaller structures, such as machines whose vibrations during operation indicate problems in their components [2]. In practice, vibration measurements can be obtained through contact or non-contact techniques [3, 4]. Contact methods are robust, well-established and widely used in both academic and industrial sectors [3, 5]. Non-contact methods, including electromagnetic, acoustic and optical approaches, avoid introducing mass in the system, which could change its vibration frequencies during operation, and enable remote monitoring. With advances in the computer vision field, optical vibration measurements have become increasingly competitive as they can o ff er superior accuracy, simple setups and cost-e ff ectiveness [3]. There are many di ff erent implementations of optical methods in vibration monitoring; some common computer vision approaches include digital image correlation (DIC) and marker tracking [6]. Marker tracking methods follow a target placed on the subject under test and register its displacement over time with good results, this data can then be processed into acceleration values [7], or used as is. Pan et al. (2023) [8] developed two deep-learning real-time methods for obtaining vibration measurements using target tracking, the methods were tested on a steel structure on a shake table, accomplishing high speed and accuracy compared to traditional sensors such as linear variable di ff erential transducers (LVDT). Yan et al. (2022) [9] obtained the natural frequency of a bridge by measuring the displacement over time of stationary reference points using a UAV (Unmanned Aerial Vehicles) and CNNs (Convolutional Neural Networks). DIC enables high-accuracy full-field measurements by comparing images before, during and after defor mation or motion, by following the behaviour of a speckle pattern that is usually applied to the surface [10, 8]. Seo et al. (2022) [11] implemented high-speed photography with DIC to measure blast-induced displacement and vibrations in underground mine rock pillars, with good results. Additionally, Frankovsky´ et al. (2022) [12] used DIC to measure metal plate displacement, and extracted modal parameters such as resonant frequencies, modal shapes and damping coe ffi cients in close agreement with numerical simulations. Conventional frame cameras, while well established, capture images at a fixed rate and thus record many pixels that are unchanged between frames. This is problematic when high frame rates are required, resulting in a large amount of information output [13]. Event-based (neuromorphic) cameras o ff er adi ff erent approach by asynchronously capturing pixel-wise brightness changes. As a result, neuromorphic sensors enable a very high temporal resolution and low latency, as well as a high dynamic range and low power consumption [14]. These characteristics suggest a strong potential for vibration monitoring. In the research conducted by Ishikawa et al. (2025) [15], an event-camera was used to measure vibration and detect anomalies during DC-motor operation, and Liu et al. (2025) [16] achieved robust performance on their neuromorphic setup coupled with a spiking neural network for fault diagnosis. In this paper, we test the feasibility of a USB event camera for vibration measurements, comparing it to a tri-axial accelerometer and a frame camera, whose images were processed using image tracking and DIC. Chapter 2 presents the experimental setups and components; Chapter 3 details the measurement procedures and results; while Chapter 4 presents the conclusions and future works.

Nomenclature

rps Rotations per second DIC Digital Image Correlation SHMStructural Health Monitoring LVDTLinear Variable Di ff erential Transducers UAV Unmanned Aerial Vehicles CNN Convolutional Neural Network

rpm Rotations per minute fps Frames per second