An Evaluation Framework for Deep Learning-Based Anomaly Detection in Structural Health Monitoring.

The task of evaluating deep learning algorithms in the context of Structural Health Monitoring (SHM) for damage detection is made particularly challenging by the limited availability of empirical data from damaged structures. Making it impossible to assert whether the trained algorithm would be able to pick up changes due to (unseen) damage. This study puts forth a methodology that employs synthesized anomalies to advance our understanding of the specific conditions under which deep learning algorithms for anomaly detection are succesfull and when they prove to be insensitive to damage.

This research aims to develop a comprehensive deep learning model that utilizes raw data as its input, negating the need for specialized preprocessing or the development of anomaly indices that are constrained to specific types of anomalies. Central to our methodology is the introduction of simulated damage into the data set through various manipulations. This evaluation method could be generalized across diverse sensor types, such as accelerometer data (ACCs) and Fiber Bragg Sensors (FBGs).

As a case study, we focus on accelerometer data, utilizing Power Spectral Density (PSD) as the input for a deep learning-based anomaly detection algorithm. The study employs both attenuation and the addition of harmonics at varying levels and frequencies to mimic anomalies, thereby investigation the model's areas of sensitivity. To empirically validate the approach, an 8-degree-of-freedom simulated system is used, and environmental effect is modelled by a linear stiffness reduction across all degrees of freedom (DOFs). Structural damage is then simulated by altering the stiffness of a specific DOF.

Our results demonstrate a robust correlation between the model’s success in identifying these synthesized anomalies and its capability to detect actual structural damage. This correlation serves as a valuable guide for hyperparameter optimization.