Input parameter specific uncertainty quantification of deep learning models using monte carlo dropout

Low-data environments, common in industrial use-cases, exhibit significantly elevated levels of both model and data uncertainty in deep learning applications. Data generation in such settings is generally financially expensive. Therefore, strategic data collection plays a critical role in improving the performance of deep learning models. Effective model uncertainty quantification plays an essential role in identifying input domains with poor model confidence, enabling more efficient data collection strategies. While Monte Carlo dropout is a well-known method for model uncertainty quantification, this study explores a novel input parameter specific application of Monte Carlo dropout to sequentially applied input channel layers rather than the entire model. This approach allows for a more detailed uncertainty quantification focused on specific input parameters, yielding additional insights into model robustness and model confidence across different input domains. A case study on condition monitoring in manufacturing using the Temporal Fusion Transformer demonstrates how these insights can guide smarter data generation strategies, ultimately improving model performance in low-data, high-cost environments.