In this paper, a finite element analysis of a cantilever structure supported by a bolt joint was performed. The frequency response function is obtained by applying the model order reduction method to the finite element model used when performing finite element analysis. The frequency response function is used to identify changes in dynamic characteristics as the clamping force changes. Applying this to deep learning, it trains a network that predicts bolt tightening force in the initial bolt loosening area. The accuracy of the prediction was evaluated through the evaluation index using the evaluation data for the bolt fastening force prediction network.
Create a finite element model of a cantilever beam supported by eight bolts. Since the frequency response analysis using the initial model requires excessive computational resources and time, a reduced model is created using a model order reduction method based on the Krylov subspace method. Thereafter, the size and phase change of the frequency response function generated using the system matrix considering the nonlinearity according to the change in clamping force at a plurality of output points is compared with the initial steady state and quantified. Generate a heatmap of the form and use it as input data for a convolutional neural network. The input data includes training, validation, and evaluation data and is randomly generated within a domain using the Mersenne Twister technique. The learning of the convolutional neural network was performed in the MATLAB environment, and after the learning was completed, the prediction of the fastening force with a relative error within 5% was performed based on the evaluation index output using the evaluation data of the convolutional neural network.