A method for predicting dynamic responses of bolted connections based on time-frequency analysis and channel-space collaborative attention CNN

Efficient prediction of nonlinear bolted-connection responses is important for repeated parameter studies because frictional contact, preload, and cyclic loading make nonlinear finite-element (FE) analysis costly. This study develops an FE-based time-frequency surrogate for a threaded bolted assembly. Displacement responses generated under variations in friction, tightening displacement, and reciprocating-load amplitude are transformed into normalized short-time Fourier transform spectrograms. An encoder-decoder CNN with parallel channel and spatial attention reconstructs the unobserved response segment from the observed segment. Performance is assessed using consistent image-domain metrics, residual maps, case-wise resampling, and lifecycle cost accounting. The reconstructed spectrograms preserve the dominant energy bands and their evolution across the investigated conditions. After training, the surrogate provides rapid response estimates for parameter screening and iterative design within the sampled FE domain.