LSTM networks detect and localize structural damage in large space platforms by learning vibration patterns from sensor data.
The Rise of Large and Flexible Space Structures
In recent years, the evolution of space structures has led to the development of increasingly large, lightweight, and flexible platforms. Deployable antennas, extended trusses, and large solar arrays enable high performance but also introduce new vulnerabilities: even small local damages can compromise stability, pointing accuracy, and ultimately mission success. In this context, structural health monitoring is no longer a secondary activity but a core function for the autonomy and safety of space systems.
A Data-Driven Perspective on Structural Health Monitoring
The work addresses this challenge by adopting a fully data-driven approach based on the analysis of dynamic responses measured by sensors distributed across the structure. The underlying idea is simple yet powerful: even when local damage does not visibly alter the global behavior of the system, it leaves a signature in vibration and acceleration time series. Capturing these signatures requires models capable of learning complex temporal dependencies and distinguishing subtle patterns embedded in noise.
Simulating Damage in a Realistic Space Platform
To develop and validate the method, a large flexible space platform representative of an Earth observation mission is modeled, equipped with extensive antennas and solar panels. Multiple damage scenarios are simulated by introducing localized faults in critical structural elements. For each configuration, the dynamic behavior of the system is reconstructed during realistic attitude maneuvers, generating time series that replicate measurements from accelerometric sensors placed at strategic locations.
Learning Structural Signatures with LSTM Networks
At the core of the analysis are deep neural networks based on Long Short-Term Memory architectures, specifically designed to handle multivariate time series. These models can retain relevant information over extended temporal windows, making them particularly well suited to discriminate between very similar structural responses. The networks are trained to solve the problem as a sequence-to-label classification task, associating each time series with the corresponding structural configuration, either healthy or damaged.
Preprocessing Time Series to Enhance Damage Detection
A key aspect of the work concerns data preprocessing. The sequences are appropriately truncated to focus on the most dynamically informative phases and normalized with respect to a nominal undamaged model, thereby amplifying the differences induced by faults. This step proves crucial in improving learning effectiveness and performance robustness.
Accurate Damage Localization with Limited Sensor Data
The results show that the system can accurately identify both the presence and the location of damage, even when using a limited number of sensors. The analysis also highlights how information quality and the selection of temporal windows have a direct impact on performance, confirming that it is not the quantity of data that matters most, but their dynamic relevance.
Authors
P. Iannelli, F. Angeletti, P. Gasbarri, M. Panella, A. Rosato
