Missing data imputation framework for bridge structural health monitoring based on slim generative adversarial networks

Abstract

In structural health monitoring (SHM) systems, sensors are important to collect structural responses to assess the load-resistant capacity and health status of structures. However, data loss of these sensors is sometimes inevitable due to communication outages and malfunction, which will result in incorrect diagnosis of structural health status. Conventional lost data imputation methods fundamentally have low implementing efficiency owing to incompact neurons network structures. This paper proposed a slim generative adversarial imputation network (SGAIN) to recover the missing deflection data for bridge SHM systems. This framework uses slim neural networks with a generator-discriminator architecture to capture the valuable information from the non-missing parts of fault sensor and other normal sensors. Based on analysis on the long-term measured deflection data under the ambient and dynamic excitation of a highway-railway dual-purpose bridge, the proposed method shows the efficient and accurate imputation performance under the input scenarios with different correlation levels. The comparative results show the proposed SGAIN significantly outperforms the conventional GAIN model for all three scenarios with different missing rates. The execution velocity of SGAIN model also outperforms GAIN with 10–20% increment for three type scenarios. Such improvement is significantly valuable and it is believed that SGAIN could also have a good imputation capability when transferred to other valuable scenarios.