An efficient fixed-time increment-based data-driven simulation for general multibody dynamics using deep neural networks

In this study, we propose an efficient fixed-time increment-based numerical scheme for data-driven analysis of general multibody dynamics (MBD) problems combining deep neural network (DNN) modeling and principal component analysis (PCA), avoiding local fluctuation of the time transient response that occurred in other works. Output results of the transient dynamics simulation can be expressed as general displacement, velocity, and acceleration, which can also be represented in a reduced dimension by the PCA. This data set is expressed in a fixed-time increment based format, leading to a large data set that is advantageous for training to construct an efficient DNN meta-model. In addition, the number of samples is also significantly reduced. As a result, the training cost is dramatically reduced compared to the simulation without PCA despite a smaller number of samples being used. To demonstrate the performance of the proposed scheme, we solve three benchmark problems: a double pendulum, damped spherical elastic pendulum, and vibrating transmission. From the results, it was found that, when the proposed scheme is used, the training time can be drastically reduced while maintaining high accuracy.