Multi-Damage Detection in Composite Space Structures via Deep Learning

Modern satellites are becoming increasingly larger, lighter, and more sophisticated. Flexible solar arrays and deployable structures enable unprecedented capabilities, but they also introduce a new challenge: even a tiny impact from orbital debris can compromise structural integrity without producing immediately visible effects. In the harsh environment of space, where maintenance is often impossible, identifying these hidden damages as early as possible becomes essential. Our recent research effort explores how AI can transform structural health monitoring into an autonomous capability, allowing spacecraft to continuously assess their own condition while carrying out normal operations.

Instead of relying on visual inspections or complex post-processing techniques, the proposed approach observes how a spacecraft naturally behaves during standard attitude manoeuvres. Every movement generates vibrations that propagate through the satellite structure, carrying subtle information about its physical condition. To exploit these signals, we developed a complete digital framework that combines physics-based simulation with deep learning. A detailed finite element model of a satellite equipped with flexible solar panels is used to reproduce multiple damage scenarios caused by orbital debris. These structural variations are then incorporated into a dynamic spacecraft model capable of simulating realistic in-orbit manoeuvres and generating sensor measurements under both healthy and damaged conditions. Rather than searching for obvious anomalies, the system learns how each damage configuration subtly modifies the temporal evolution of structural vibrations.

An important aspect of the work is the comparison between two different sensing technologies. One monitoring architecture relies on distributed accelerometers, while the other uses piezoelectric patches capable of measuring the electrical response generated by structural deformation. Instead of assuming one solution is inherently superior, both sensor networks are evaluated under the same operating conditions. Their measurements are collected during simulated spacecraft manoeuvres and transformed into time-series datasets that become the input for the learning algorithm. This enables a direct assessment of how effectively each sensing strategy captures the information required to distinguish different damage locations.

Because the available information is inherently sequential, the proposed methodology adopts a recurrent neural network. Unlike approaches that require handcrafted features or extensive preprocessing, the network learns directly from raw temporal sensor signals, discovering the dynamic patterns associated with each structural condition. The objective goes beyond determining whether damage exists. The model is trained to recognize multiple possible impact locations, transforming structural health monitoring into a multi-class classification problem capable of identifying where the damage has occurred rather than simply detecting its presence. To build an informative training dataset, the candidate damage locations are selected through modal strain energy analysis, concentrating the learning process on the areas that are structurally most critical.

The experimental results demonstrate that both sensing configurations allow the deep learning model to identify damaged regions using only short sequences of measured data. This suggests that continuous onboard monitoring could become feasible without requiring large observation windows or computationally expensive signal processing. The broader significance extends well beyond this specific application. As future spacecraft become increasingly autonomous and missions last longer in more demanding environments, intelligent monitoring systems capable of continuously interpreting structural behaviour will become a key component of mission resilience. Rather than waiting for failures to manifest themselves through degraded performance, spacecraft could detect the earliest signs of structural degradation, supporting timely mitigation strategies and laying the foundation for future autonomous maintenance and repair operations in orbit.