Deep neural network can give contributions of input: a feasibility study of transfer path analysis

抄録

Deep learning is emerging in various areas such as structural noise and vibration problems. In this study, feasibility whether a deep learning model can exactly represent the transfer function of a structure is tested especially in viewpoint of training dada and model structure. For the feasibility test, a transfer path analysis model using an artificial neural network structure was developed and trained with various augmented data as well as original input/output responses. The original input/output responses were collected from a finite element model of a test structure in the frequency domain. By changing the reference phase and by multiplying the complex conjugate of the input and output responses to the input/output pairs, the collected data was augmented. Comparing the performance of the trained deep learning model with classic transfer path analysis model, to correctly represent the contribution of each input to output response it is essential that the phase and the crossspectrum augmented should be included during the deep learning model training process.