PSI - Issue 62

Israel Alejandro Hernández-González et al. / Procedia Structural Integrity 62 (2024) 879–886 Hernández-González/ Structural Integrity Procedia 00 (2019) 000 – 000

880

2

1. Introduction The effective handling of aging infrastructure represents a significant challenge for most developed nations. A substantial portion of the current infrastructure was built in the post-World War II era, and it is now reaching or surpassing its designated lifespan of 50-100 years. Dreadful incidents like the 2007 collapse of the I-35W bridge in Minnesota, USA (Hao, 2010), the 2018 collapse of the Morandi Bridge in Genoa, Italy (Calvi et al., 2019), or the 2019 collapse of the Nanfang'ao Bridge in Yilan County, Taiwan; vividly highlight the serious dangers associated with insufficient maintenance. These conditions have prompted significant research and development efforts in the field of Structural Health Monitoring (SHM) as a proactive maintenance strategy. Within this context, vibration-based SHM methods using Operational Modal Analysis (OMA) have gained widespread popularity due to their low intrusiveness, completely non-destructive nature, suitability for long-term SHM, and global damage identification capabilities. Artificial intelligence (AI) has become a crucial area of advancement in the contemporary world, holding the promise to transform various facets of industry and society. In the field of SHM, a significant amount of research has been devoted to identifying effective damage indicators. Notably, Giglioni et al. (2023) presented a damage detection approach utilizing autoencoders to reconstruct raw acceleration data. Trained using data from the healthy condition of the structure, their results demonstrated that the residuals between the predictions of the AI and the recorded time series can be used as sensor-based damage-sensitive features. Rosso et al. (2023) investigated the application of subspace-based features to train supervised multi-layer perceptron (MLP) networks for multiclass damage classification, demonstrating great potential for damage localization in a laboratory I-beam. However, the utilization of deep learning for modal identification in dynamic systems has been relatively limited. One noteworthy contribution in the literature is the work of Liu et al. (2023), which presented a two-step AI algorithm for SSI. This algorithm incorporates a first neural network (NN) to determine the order of the dynamic system, followed by a second NN for modal identification. Likewise, Shim et al. (2023) introduced an SSI-based long-short term memory (SSI-LSTM) method for monitoring changes in the modal parameters of structures. Their AI model utilizes raw acceleration data as input and is trained to replicate the modal estimates derived from the covariance-driven-SSI (CoV-SSI) algorithm. Considering the simplicity of BSS techniques and the capacity of AI to formulate rapid identification models, their combination holds substantial promise for overcoming the scalability challenges of OMA. In this light, Shu et al. (2023) recently formulated a multi-task Deep Neural Network (DNN) designed for the automated recognition of independent modes derived from SCA. In their methodology, the mode shapes are directly extracted as the weights between the last two layers of neurons in the NN, while resonant frequencies and damping ratios are determined from the independent components using the random decrement technique (RDT). Inspired by the possibilities of AI to solve the scalability limitations of current OMA techniques, this paper introduces a novel Multitask Learning Deep Neural Network (MTL-DNN) for quick BSS and automated modal identification of structures. In contrast to previously reported AI methods documented in the literature, the present method makes use of a multi-task learning deep NN capable of extracting both the real and imaginary components of the independent modal sources and, as a result, the complex-valued mode shapes. The effectiveness of the developed approach is appraised through a real-world in-operation-bridge: the Méndez-Núñez Bridge in Granada (Spain). 2. Theoretical background 2.1. Second order blind identification Based on the concepts of structural dynamics theory, the governing equation of motion of a dynamic system with n -degrees of freedom (DOF) can be written as follows: ̈( ) + ̇( ) + ( ) = ( ), (1) with M , C and K ∈ ℝ denoting the mass, damping, and siffness matrices of the system, respectively. The vectors F (t) and X (t) ∈ ℝ are the external force and the response displacement vectors, respectively. Dot notation is used to indicate differentiation with respect to time. Assume that there are m ∈ℕ ∗ sensors deployed on the structure, recording