System-informed neural networks with deep ensembles for frequency detection enhancement

We propose a deep learning-based frequency analysis framework, referred to as the system-informed neural network (SINN), which integrates the corresponding linear system model. The SINN architecture is based on the adaptive learned iterative soft shrinkage algorithm and incorporates the system model into its loss function. This design ensures good generalization, fast processing time, and solutions that align with the system model, similar to physics-informed neural networks. To enhance performance, we utilize deep ensemble-based uncertainty quantification. This involves training multiple SINNs with different initial weights on distinct training datasets, allowing them to produce diverse predictions for the same test sample. The proposed scheme is evaluated using in situ acoustic data containing 43 frequency components and compared against results from established frequency analysis methods.