Image-Based Structural Health Monitoring Methodology for Corrosion Diagnosis and Prognosis

Corrosion is a serious concern for structural integrity since it can lead to premature structural failures. Nevertheless, its monitoring is particularly challenging due to inherent features of the corrosion process: it is strongly dependent on environmental conditions, materials used, and their coupling. Generalised corrosion phenomena are visible at visual inspection and lead to a reduction of the resistant thickness of components and structures. Instead, localised corrosion results in visually difficult-to-detect damages that can lead to fatigue cracks. In addition, these two corrosion types are not mutually exclusive: a component may present corrosion pits even if it is macroscopically subjected to generalised corrosion. However, a probability-based relationship between generalised corrosion evolution and corrosion pits’ presence and geometrical features is missing. Therefore, it is important to monitor generalised corrosion phenomena to assess the structural integrity of structures and components given the reduction in their resistant section and the possible presence of corrosion pits. Within this context, this study presents a methodology for corrosion-based structural assessment by integrating image-based corrosion diagnosis, measurements of environmental and corrosion rate related parameters, and a filtering technique to perform corrosion diagnosis and prognosis. The image-based corrosion diagnosis relies on a Convolutional Neural Network (CNN) that has been trained to automatically perform semantic segmentation on images of a corroded helicopter component. In this way, it is possible to have discrete-in-time observations of the actual corrosion level of the component when the helicopter is not on missions. The implemented CNN is not only able to distinguish between corroded and uncorroded regions but also between two different corroded regions of interest due to different materials. The CNN was trained with manually segmented images from which corrosion indexes have been extracted. The latter have been related to (i) environmental parameters and (ii) corrosion rate related parameters. In this way, two models have been obtained offline to predict corrosion evolution. Eventually, a Particle Filter (PF) was implemented to adapt the models to the observations of the corrosion level by the CNN. The proposed framework integrates image processing, sensor measurements, and filtering techniques for structural assessment. The PF guarantees the adaptability to real-time observations of the corrosion evolution models developed offline, improving the reliability of the methodology. While the presented framework has been applied to a helicopter component, it can be easily applied to other systems. For instance, being based on image processing, its image-based nature makes it well-suited for monitoring offshore or hardto- access structures using drones or fixed cameras. Eventually, this study provides a step towards a more accurate and data-driven corrosion prognosis, enhancing the accuracy of structural integrity assessments.