There is a need for estimating the profile of a dynamic input motion that shakes a solid by using vibrational measurement data on the surface of the solid. By using the reconstructed input motion, engineers can replay the dynamic responses of the solid. The real-world application of this problem can be characterizing the vibrational sources in a spacecraft during its flight and reconstruction of seismic responses in critical infrastructure shaken by a strong earthquake.An existing method is based on full-waveform inversion (FWI) to reconstruct unknown dynamic input motions through a heterogeneous solid from a surface response. Although FWI has proven to be fairly accurate and precise with the reconstruction process, the process of its reconstruction is prohibitively time-consuming to produce results. This paper attempts to recreate a dynamic input motion more accurately through time-efficient machine learning approaches that significantly reduce the processing time.The adaptability of neural networks to learn from the data provided and the computational prowess present today has made machine learning highly desirable, especially for regression problems like the one tackled in this paper. The two Artificial Neural Networks (ANNs) approaches attempted in this paper are Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs). DNNs have the capability to learn intricate input and output relations with millions of learnable parameters built into multiple hidden layers. One-dimensional (1D) CNNs have also produced stellar results in this paper.This paper goes into depth by explaining how DNNs and CNNs can be adapted to learn a complex inversion process by learning arrangements between input-layer and output-layer datasets provided to the network and have the network predict targeted waveforms in the output layer with high certainty. Both neural networks have shown promising results in the reconstruction of a dynamic input motion and reduced the reconstruction process time from 15 hours to a fraction of a second with high accuracy.The ANNs are trained based on 18-second measured displacement wave signal data in the input layer and predicts dynamic input signal data in the output layer. The performance of ANN is evaluated based on how accurately the ANN can predict targeted dynamic input signal data in the output layer. The ANN was tested on 2000 unseen datasets which produces a mean error of 2.4 to 3.4% while precisely reconstructing the targeted output-layer signal. The highlight of this paper is that both DNNs and CNNs successfully reconstruct a real-life dynamic input signal with near-perfect precision. The paper also conducts brief parametric studies on why certain parameters and hyperparameters were chosen to construct the specific DNN and CNN models. According to the results of this paper, ANNs stand out at predictive modeling to address dynamic-input identification problems like the one presented in this paper.

Artificial Neural Network Approaches for the Identification of Dynamic Input Motions in a Heterogeneous Solid

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.