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

Area of Science:

• Engineering
• Artificial Intelligence
• Computational Science

Background:

• Virtual sensing is a growing research area.
• Recurrent neural networks are effective for time-series forecasting.
• Nonlinear AutoRegressive eXogenous (NARX) models are a type of recurrent neural network.

Purpose of the Study:

• To investigate the sensitivity of NARX models with varying complexity for heat flow simulation.
• To determine if loss function values alone indicate model sensitivity.
• To identify potential weaknesses in undertrained NARX models.

Main Methods:

• Utilized NARX models as surrogate neural networks for heat flow simulation.
• Analyzed the sensitivity of NARX models across different levels of complexity.
• Examined the impact of training epochs on model sensitivity and artifacts.

Main Results:

• Loss function value alone is insufficient to indicate NARX model sensitivity.
• Undertrained NARX models exhibit artifacts in their sensitivity, revealing model weaknesses.
• Model sensitivity generally increases with more training epochs, while its pattern remains consistent.

Conclusions:

• Model sensitivity is a crucial factor in virtual sensing applications.
• Careful consideration of training extent is necessary to avoid artifacts in NARX models.
• NARX models can be effectively used for heat flow simulation with appropriate training.