Convolutional Autoencoder-Based Damage Detection for Urban Railway Tracks Using an Ultra-Weak FBG Array Monitoring System

Urban railway tracks inevitably suffer damage due to the cyclic load of trains, necessitating structural health monitoring (SHM) with the primary purpose of damage detection. Recently, the ultra-weak fiber Bragg grating (UWFBG) monitoring system has been employed for long-term monitoring of urban railway tracks due to its ability to multiplex thousands of sensors on a single optical fiber. The vast amount of data collected imposes the importance of data-driven damage detection methods. Given the lack of labeled damage datasets, unsupervised learning methods are highlighted. This study proposes a damage detection method for urban railway tracks based on an unsupervised deep neural network, referred to as deep convolutional autoencoder (DCAE). The monitored data are first processed to the autocorrelation functions (ACFs) to be aligned across different channels, and then, the multichannel ACFs are used as the inputs of the DCAE model. Finally, the reconstruction error of the DCAE model is employed as the damage index, and field monitoring data are utilized to verify the proposed method. The results show that the proposed damage index is sensitive to track damage, and the precision of damage detection increases with the threshold of reconstruction error, reaching a peak at 1. The method also achieves a maximum F1 score of 0.90.