Robustness Analysis of a Fast Virtual Temperature Sensor Using a Recurrent Neural Network Model Sensitivity

Virtual sensing is an emerging field of research that has garnered increasing attention in recent years. In this paper, we focus our attention on recurrent neural networks for time-series forecasting, namely, the Nonlinear AutoRegressive eXogenous model (NARX). The NARX is utilized as a surrogate neural network for simulating heat flow. Our research has investigated the sensitivity of NARX models of varying complexity. The presented results show that the loss function value alone does not indicate the model's sensitivity. We have demonstrated that undertrained models exhibit visible artifacts in their sensitivity, highlighting the model's weak points. From observing how the sensitivity changes over training epochs, we can conclude that the sensitivity increases with more epochs, while its overall shape remains relatively unchanged.