Defining Energy Indicators for Impact Identification on Aerospace Composites: A Physics-Informed Machine Learning Perspective
Natália Ribeiro Marinho, Richard Loendersloot, Frank Grooteman, Jan Willem Wiegman, Uraz Odyurt, Tiedo Tinga · 2025
Read Defining Energy Indicators for Impact Identification on Aerospace Composites: A Physics-Informed Machine Learning Perspective by Natália Ribeiro Marinho, Richard Loendersloot, Frank Grooteman, Jan Willem Wiegman, Uraz Odyurt, Tiedo Tinga (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01592.txt.
Introduction
Natália Ribeiro Marinho and five colleagues address a safety problem in aircraft parts made of composite materials. A low-speed impact can crack the inside of a panel while the surface looks untouched. Estimating the impact's energy from sensor signals is the first step to locating damage, but the inverse problem is noisy, data-poor, and full of tangled candidate features.
Their answer is a structured workflow for choosing input features, guided by engineering knowledge. Signals are mined in the time, frequency, and time-frequency domains, then filtered for statistical significance, correlation, reduced dimension, and robustness to noise. The surviving energy indicators feed a neural network. It is trained and validated on experimental impact data. The result is a principled input design, instead of the field's usual ad hoc feature picking.
Catalog record
| CREATOR | Natália Ribeiro Marinho, Richard Loendersloot, Frank Grooteman, Jan Willem Wiegman, Uraz Odyurt, Tiedo Tinga |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 69 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-01592
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Defining Energy Indicators for Impact Identification on Aerospace Composites: A Physics-Informed Machine Learning Perspective Natália Ribeiro Marinho, Richard Loendersloot, Frank Grooteman, Jan Willem Wiegman, Uraz Odyurt, Tiedo Tinga Abstract Energy estimation is critical to impact identification on aerospace composites, where low-velocity impacts can induce internal damage that is undetectable at the surface. Current methodologies for energy prediction are often constrained by data sparsity, signal noise, complex feature interdependencies, non-linear dynamics, massive design spaces, and the ill-posed nature of the inverse problem. To address these challenges, this study introduces a physics-informed framework that embeds domain knowledge into machine learning through a dedicated input space. The approach combines observational biases, which guide the design of physics-motivated features, with targeted feature selection to retain only the most informative indicators. Features are extracted from time, frequency, and time–frequency domains to capture complementary aspects of the structural response. A structured feature selection process integrating statistical significance, correlation filtering, dimensionality reduction, and noise robustness ensures physical relevance and interpretability. Exploratory data analysis further reveals domain-specific trends, yielding a reduced feature set that captures essential dynamic phenomena such as amplitude scaling, spectral redistribution, and transient signal behaviour. Together, these steps produce a compact set of energy-sensitive indicators with both statistical robustness and physical significance, resulting in impact energy predictions that remain interpretable and traceable to measurable structural responses. Using this optimised input space, a fully-connected neural network is trained and validated with experimental data from multiple impact scenarios, including pristine and damaged states. The resulting model demonstrates significantly improved impact energy prediction accuracy, reducing errors by a factor of three compared to conventional time-series techniques and purely data-driven models. Overall, the framework advances predictive performance, interpretability, and diagnostic confidence by embedding domain-knowledge insights through feature-level design. Its strength lies in the integration of observational biases with targeted feature selection for an effective impact identification method. [orcid=0000-0002-1042-3571] [orcid=0000-0002-1113-8203] [orcid=0000-0001-8036-3845] [orcid=0000-0002-1066-7234] [orcid=0000-0003-1094-0234] [orcid=0000-0001-6600-5099] keywords Impact identification ,Physics-informed machine learning ,Aerospace composites ,Structural health monitoring † † titlenote: This work is part of the PrimaVera Project, which is partly financed by the Dutch Research Council (NWO) under grant agreement NWA.1160.18.238. † † corresponding: Corresponding author: [email protected] 1 Introduction Aerospace composite structures are susceptible to Barely Visible Impact Damage (BVID), which poses a critical threat to both structural integrity and operational safety. These internal damages are typically undetectable by conventional visual inspection and can significantly degrade load-bearing capacity [ 82 , 11 ] . To mitigate these risks, Structural Health Monitoring (SHM) systems have been developed to capture dynamic responses using embedded sensors. These systems enable the estimation of impact energy and the detection of internal damage, supporting timely maintenance, enhancing operational reliability, and proposing design improvements [ 36 , 91 , 87 ] . A central challenge in this context is the reconstruction of impact energy from measured responses, which constitutes an ill-posed inverse problem. The relationship between sensor signals and unknown impact conditions is typically non-unique, highly sensitive to noise, and underdetermined. These difficulties intensify under practical constraints, such as sparse sensor layouts, uncertainties in measurements, and non-linear structural behaviour at high excitation levels. Consequently, conventional methods frequently produce unstable or non-physical estimates unless constrained by regularisation or supported by computationally intensive formulations [ 90 , 55 ] . To ensure reliable performance under these conditions, estimation strategies must effectively manage incomplete and noisy data, preserve numerical stability, and support real-time execution within embedded systems. Research efforts sought to address these challenges using numerical, analytical, data-driven, and hybrid approaches. In the numerical and analytical domains, Liu and Wang [ 55 ] developed a finite-element model for woven laminates subjected to low-velocity impact. Their model reconstructed force histories with reasonable accuracy but required intensive computation and relied exclusively on simulated virtual accelerations. Correas et al. [ 14 ] introduced an analytical spring-mass model for stiffened panels that shows strong agreement with numerical simulations. However, the model employed simplified loading assumptions, idealised boundary conditions and lacked experimental validation. Data-driven methods have also been employed to model complex and non-linear relationships between structural responses and unknown impact conditions. Tabian et al. [ 83 ] employed a Convolutional Neural Network (CNN) to localise and quantify impacts on composite structures. Similarly, Ghajari et al. [ 77 ] successfully reconstructed impact force histories from time-series data. Although these approaches exhibit effectiveness in controlled environments, they typically demand large, high-quality datasets and struggle with generalisation to new or noisy scenarios. Their dependence on data representativeness and low interpretability further constrained practical deployment. Additional developments have explored probabilistic and hybrid techniques. Yan et al. [ 86 ] introduced a Bayesian regularisation framework combined with an unscented Kalman filter to improve noise robustness for impact identification. Although the method improved stability, it remained sensitive to sensor configuration and limited to low-energy regimes. Huang et al. [ 37 ] employed transfer learning in a deep learning model, reaching acceptable accuracy but requiring extensive tuning and access to reference force histories. Zhang et al. [ 89 ] integrated experimental testing with high-fidelity simulations on honeycomb panels, yielding detailed insights but incurring high computational costs and sensitivity to parameterisation. Despite their contributions, probabilistic and hybrid methods remain computationally demanding, difficult to scale, and dependent on sensor configurations and reference data, which limits their practical adoption. Overall, these approaches highlight critical gaps in existing literature. Many methods rely on idealised boundary conditions or simplified impact scenarios, which reduce the reliability of the estimation and limit generalisation. Others depend heavily on large datasets and remain vulnerable to noise, sensor placement, and parameterisation. In addition, the ill-posed nature of the inverse problem, combined with low interpretability, limits its implementation in operational environments. Above all, most approaches neglect the explicit incorporation of physical insight into the input space. As a result, models often fail to distinguish between measurement artefacts and the actual system dynamics, thereby compromising both predictive accuracy and physical relevance. In response to these limitations, Physics-Informed Machine Learning (PIML) [ 44 , 16 , 45 ] , also referred to as Physics-Enhanced Machine Learning (PEML) [ 49 , 13 ] or grey-box modelling [ 15 , 32 ] , has emerged as a promising approach. By embedding physical knowledge into data-driven frameworks, PIML improves robustness, interpretability, and generalisation. Importantly, it reflects the knowledge–data balance, where physical insight can compensate for limited or noisy data. Although this strategy has been applied in other engineering domains, its use for impact identification in aerospace composites remains unexplored. More broadly, input space design has received limited attention across disciplines, despite being a critical step in linking measurement signals to governing dynamics and ensuring reliable energy estimation. This study addresses this gap through a hybrid framework that integrates observational bias within a PIML architecture. The framework builds on the principle that measurement signals carry information about the input excitation [ 65 ] . Observational bias is then addressed by utilising this information to construct the input space from physics-motivated features that capture the structural dynamics. These features form a multi-domain candidate set extracted from time, frequency, and time–frequency descriptors. Each domain contributes complementary information, and together these descriptors provide a complete and balanced characterisation of the impact event that no single domain can achieve. Although prior studies in