Machine learning algorithms for delaminations detection on composites panels by wave propagation signals analysis: Review, experiences and results

Abstract

Performances are a key concern in aerospace vehicles, requiring safer structures with as little consumption as possible. Composite materials replaced aluminum alloys even in primary aerospace structures to achieve higher performances with lighter components. However, random events such as low-velocity impacts may induce damages that are typically more dangerous and mostly not visible than metals. The damage tolerance (DT) approach is adopted for the fatigue design of aircraft, but fracture mechanisms and propagation of failure prediction in composite structures are much more challenging. Consequently, the DT approach is still costly for these types of structures. It can be achieved only through expensive experimental testing and a drastic reduction of allowable stress levels and maintenance intervals by applying scattering factors due to the uncertainties involved in their original estimations. Structural health monitoring (SHM) systems deal mainly with sensorised structures providing signals related to their “load and health status” to reduce maintenance and weights. At the same time, the use of Deep Neural Networks (DNNs) based on strategic engineering criteria, for instance, may represent an effective and efficient analysis tool to promote faster data analysis and classification. In the field of aircraft maintenance, this approach may lead, for example, to a faster awareness of an aircraft/fleet situation or predict failures. Deep learning-based networks provide automatic feature extraction at different levels of abstraction. With the universal function approximation property of neural networks, it learns the inverse mapping from input space (signals) to target space (damage classes). Starting from the well-established Structural Health Monitoring (SHM) technologies, a network of distributed sensors embedded throughout the structure could be used for real-time structural monitoring and data acquisition. Structural data will constitute an enormous amount of information that can be adequately filtered with the help of specific DNNs designed and trained for the structural context and aimed to classify and identify significant parameters. The authors have collaborated for some years to collect wave propagation signals through experimental tests and validated numerical models of healthy and damaged composite structures, and developed machine learning algorithms (mainly dense and convolutional neural networks) aimed at signal classification and analysis for damage detection and localization. This paper presents a brief review of relevant works about SHM employing Machine Learning methodologies and summarizes the most promising approaches developed during the last years jointly by the two research groups and presents a critical analysis of obtained results and subsequent future activities.