Correntropy-Induced Weighted GAN for Bridge Anomaly Detection

Bridges play an essential role in transportation and economics. Since the loss is significant in the event of bridge damage, the monitoring of bridge structure health is important. One of the important tasks of bridge health monitoring is the bridge anomaly detection. The bridge anomaly detection is a challenging problem since the real-world bridge data collected by the sensors are usually contaminated by large non-Gaussian noises. To address this issue, we propose a correntropy-induced weighted generative adversarial network (CWGAN) for bridge anomaly detection. Specifically, we first propose a correntropy-based generator, which has an encoder–decoder structure with a memory structure, and optimize it based on the maximum correntropy criterion. Then, derived from the correntropy of the reconstruction errors, we specially design a weight module to obtain a weight variable that measures the contamination-free degree of the corresponding sample. Afterward, we take advantage of the weight module and transform the generator and discriminator modules of CWGAN into weighted versions to achieve robustness against contamination and obtain smaller reconstruction errors for noncontaminated normal data. We integrate and iteratively perform the reconstruction and weighting processes to achieve better anomaly detection performance. After the model optimization, we detect the anomalies with reconstruction errors. We conduct experiments on a model bridge to demonstrate the effectiveness of the proposed CWGAN.