Please login first
A Generative Adversarial Network based autoencoder for Structural Health Monitoring
* 1 , 1 , 1 , 2 , 1 , 1 , 1
1  Politecnico di Milano
2  CentraleSupélec, Université Paris Saclay
Academic Editor: Frank Werner


Civil structures and infrastructures such as buildings, bridges, tunnels and dams play a crucial role in our society. Their safety and health are threatened by different factors: aging, progressive accumulation of damage and alteration of working and environmental conditions with respect to the design ones, are just a few examples. Rebuilding these systems when damage has grown too large, would not be economically feasible; therefore, various approaches for damage detection have been recently developed. In this context, Structural Health Monitoring (SHM) has become an active field of research, aiming to detect, locate and quantify the damage, namely the degradation of the strength and stiffness properties of the structural system. Compared to the widespread visual inspections and non-destructive testing, a global monitoring approach based on continuous vibrational measurements provides numerous advantages: first, it does not require any a-priori knowledge on the position of the damage; second, it can provide a quantitative estimate of the structural health that is seldom furnished by traditional methods. In this work we propose a new SHM approach leveraging on a Generative Adversarial Network (GAN) based autoencoder. The characteristic feature of the offered neural network architecture is the capability to reconstruct the sensor recordings, leading to an informative and disentangled latent space associated to the damage class. The main novelty of the approach is represented by the capability to generate plausible signals for different damage states, based only on undamaged recorded or simulated structural responses, thus without the need to rely upon real recordings linked to damaged conditions.

Keywords: Structural Health Monitoring; Machine Learning; Neural Networks; Generative Adversarial Network; disentangled latent space