An improved multi-task approach for SHM missing data reconstruction using attentive neural process and meta-learning

Abstract

Missing data due to sensor or transmission failures pose a significant challenge in structural health monitoring (SHM) systems, and data reconstruction methods can effectively address the missing data problem. Most of the traditional approaches typically focus on single-task data reconstruction, requiring repeated applications for each sensor and increasing computational cost. To address this issue, in this paper, we propose a probabilistic deep learning-based approach for multi-task data reconstruction. The multi-task data reconstruction is achieved by a probabilistic learning-based attentive neural process network (ANPN) that uses a common implicit data-driven kernel to learn the relationships among sensors. The meta-learning strategy is employed to train the common kernel in the ANPN. The attention mechanism is incorporated to further improve the reconstruction accuracy by enhancing the learning of the relationship between missing data and observed data. The effectiveness of the proposed ANPN is evaluated using the simulation data from a square pyramid space grid and the field data acquired from the Xiong’an Railway Station. The results show that the proposed ANPN can accurately reconstruct the missing data from multiple sensors within a second, underscoring its computational efficiency and accuracy. Furthermore, the influence of critical parameters (i.e., network depth, feature size, attention mechanism, and data loss ratio) on the reconstruction accuracy and efficiency is comprehensively investigated, and the optimal parameter settings are suggested.