A multi-task learning-based automatic blind identification procedure for operational modal analysis

Traditional modal analysis approaches for structural heath monitoring (SHM) have a low implementation efficiency. This study develops an artificial intelligence (AI)-based automatic blind identification procedure for determining the modal parameters of structures. The core of this procedure is to establish a multi-task deep neural network (MTDNN) that can automatically and efficiently extract independent modes from multi-mode vibration responses of structures. Then modal frequencies and damping ratios of structures can be extracted from independent modes via employing the conventional random decrement technique (RDT) and curve fitting approach. The weights between the last two layers of MTDNN represents the corresponding mode shapes. The approach is verified by a five-degree-of-freedom numerical model and then imple-mented to a field test of a long-span cable-stayed bridge in engineering practice. The results indicate the ability of the developed approach to automatically determine the modal parameters of structures with reliable accuracy. In the prediction stage, the modal separation process by using MTDNN takes only about 0.12 s (numerical example) and 0.48 s (practical example). The computational efficiency of the developed approach is significantly higher than that of traditional stochastic subspace identification (SSI) and frequency domain decomposition (FDD) approaches, which provides a promising new solution for online modal parameter identification and modal tracking of structures.