Automated Hyperparameter Tuned DNN for Predicting Uncertain Oscillations

EasyChair Preprint no. 13492

Abstract

This study proposes an automated hyperparameter-tuned deep neural network (DNN) designed to predict uncertain oscillations in systems, which are typically challenging for mobile machines and heavy equipment like excavators and spacecraft. The DNN utilizes a regression model, enhanced through automated hyperparameter tuning, to predict the behavior of these oscillations effectively. The key innovation involves using a feedforward neural network (FFN) configured through a random optimization algorithm to optimize variables such as the learning rate, number of layers, dropout, and activation function. The proposed DNN model demonstrates precise prediction of uncertain mass oscillations with high accuracy, quantified by a relative mean absolute error of 0.010, using simulation data from the Exudyn software.This advancement holds significant potential for enhancing control and structural health monitoring of machines and equipment, suggesting that further research could extend these methods to broader applications in multibody systems, control algorithms, and structural health monitoring.

Keyphrases: Automated hyperparameter, Data-driven surrogates, Deep Neural Network, machine learning, Uncertain oscillations

@Booklet{EasyChair:13492,
  author = {Qasim Khadim and Grzegorz Orzechowski and Emil Kurvinen and Aki Mikkola and Johannes Gerstmayr},
  title = {Automated Hyperparameter Tuned DNN for Predicting Uncertain Oscillations},
  howpublished = {EasyChair Preprint no. 13492},

  year = {EasyChair, 2024}}