Unsupervised long-term damage detection in an uncontrolled environment through optimal autoencoder

Abstract

Unsupervised damage detection in the presence of both regular environmental variations, such as a daily change in temperature and humidity, and irregular environmental variations, such as rain and snow, remains a critical challenge in guided wave structural health monitoring. This paper proposes an optimal autoencoder-based damage detection strategy to solve this problem. Although the autoencoder and other neural networks have been used to detect anomalies in structural health monitoring, the training data has been also required to be collected from a known intact structure, which increases the time and labor cost for inspection and limits the application of such methods. Instead, our autoencoder is trained with guided wave data that contain uncontrolled regular, irregular, and damage variations. We investigate hyperparameters that will negatively influence autoencoder-based damage detection, including the training time and damage duration. We also propose sparsity, dropout, and weight decay regularization strategies to enhance the robustness of the autoencoder to these hyperparameters. Results show that our optimal autoencoder method can achieve an area under the receiver operating curve score near 0.92 for detecting damage present in the last 16 days, which improves previous local principal components analysis reconstruction methods with a score of 0.88.