Abstract
A novel model reduction approach for analyzing structural mechanical response states in automotive systems is introduced, leveraging time-domain signal data from road testing. Initially, finite element modeling and analysis identify peak stress areas in critical structures under typical operating conditions. Subsequently, road load spectrum signal tests extract vibration acceleration and strain signals from these areas, forming the foundation for model reduction training and validation sets. Comprehensive research into machine learning and model reduction techniques is conducted, with a focus on polynomial order in response surface models and kernel functions in Gaussian process models. A hyperparameter tuning and optimization procedure for neural network models is proposed, exploring batch size, hidden layer type, activation function, cells number, epochs number, and number of hidden layers. This ensures that the reduced-order model achieves high fidelity index of 98.6% and 95.2% for the validation and test sets, respectively. The refined model is encapsulated for deployment in monitoring local structural mechanical response states and assessing damage. Notably, increasing cells number necessitates a proportional rise in epochs to fully exploit multi-neuron learning capabilities. Conversely, adding hidden layers may not enhance accuracy, potentially reducing model generalization and increasing training costs.