Unsupervised anomaly detection in non-linear mechanical systems for structural health monitoring

Abstract This work investigates the application of operational modal analysis (OMA) and unsupervised variational autoencoders (VAEs) for damage and anomaly detection in structural health monitoring. The underlying structural model is a multi-degree-of-freedom (MDOF) Bouc-Wen-Baber-Noori hysteretic system, which captures the highly nonlinear behavior typical of degrading and pinching effects in real-world structures. By simulating and analyzing the responses of this nonlinear MDOF system under stochastic excitation, we assess the effectiveness of data-driven approaches, including VAEs trained on healthy states, for detecting changes in system dynamics. The example system was specifically chosen to represent anomalies that manifest predominantly in the nonlinear components of the system, while leaving the linear part unchanged. Special attention is given to the sensitivity, interpretability, and ease of obtaining extracted features from the different methods. The results show that VAEs offer advantages over the well established COVariance-driven stochastic subspace identification OMA approach when applied to markedly nonlinear data.