Transfer path analysis using deep neural networks trained by measured operational responses

We present an operational transfer path analysis (OTPA) formulation based on deep learning to solve structural problems. Deep neural networks (DNNs) with fully connected or convolutional layers model the transfer function from interfacial joints to the responses of points of interest in terms of forces, thereby eliminating the cross-coupling effects of conventional OTPA methods. Using an operational dataset, phase and cross-spectrum augmentation procedures were applied to train DNNs by reference to the required path contributions. A test structure with two plates and three transfer paths was used to experimentally validate the OTPA framework. The operational responses were quantified and used to train DNNs that engage in OTPA; we evaluated various hyperparameters. The path contributions were obtained from DNNs that had been trained to follow the required OTPA procedure and compared to those of classical transfer path analysis (TPA). Experimental identification of the path contributions revealed that the new OTPA method was as accurate as the classical TPA method and had good overall task efficiency.