Frequency response similarity-based bolt clamping force prediction method using convolutional neural networks

This paper proposes a convolutional neural network (CNN)-based method with which to predict bolt clamping force using the frequency response of bolted structures. The dynamic characteristics of the bolted structure change with the bolt clamping force, which is predicted using a CNN trained with massive frequency response data. Big data required for training the CNN is constructed using prestressed frequency response analysis according to the clamping force of individual bolts. The numerical efficiency is increased using the Krylov subspace-based model order reduction (MOR) method. The frequency response for each set of bolt clamping forces calculated from the MOR method is converted into form of the magnitude and shape (MS) similarity spectrum by using the MS similarity function. Finally, an MS similarity map is generated by stacking the MS similarity spectrum at several output points. A CNN that is trained using massive MS similarity maps as training data, is used to predict the clamping force of bolted structures. To validate the efficiency and accuracy of a trained CNN in practical applications, the prediction results of the trained network in terms of computation time and accuracy were compared according to the size of the training input data.