Research

1. Principled UQ for inverse problems in acoustical signal processing [JASA'23, ASA'23, IEEE MLSP'23]:  Slides Link

For the acoustic problem of direction-of-arrival (DOA) estimation, we want to provide statistically valid and easy-to-obtain uncertainty intervals in the presence of external uncertainty. Conformal prediction (CP) technique produces an interval containing the true value with a statistical guarantee. We demonstrate the usage of CP with multiple machine-learning models trained for DOA estimation. CP is able to produce uncertainty intervals that are statistically valid and can reliably assess the DOA estimation performance in the presence of uncertainty.

1. Uncertainty management for structural health monitoring [JASA'22, ICASSP'22]: Slides Link 

We demonstrate the ability of specialized machine learning techniques to generate robust predictions in the presence of external sources of uncertainty. We show that the ensemble of deep neural networks trained solely on simulated observations with adversarial perturbations is an attractive alternative. This strategy provides improvements in damage detection and localization performance over baseline methods for real-world data with external variations.

2. Reliability assessment for structural health monitoring [under review at NDT & E International]: 

We want to assess the reliability of a guided-wave structural health monitoring system. In a structural monitoring sensing setup, sensors transmit and receive guided waves. Based on the received signals, we want to localize the damage. The challenge is to assess the reliability of localization predictions when external uncertainties are present. We analyze the ability of multiple deep-learning methods to represent the reliability of individual damage localization estimates in the presence of external uncertainty. We are able to obtain a heuristic uncertainty estimate in a simple and scalable manner.

1. Obtaining interpretable insights for grain growth [Frontiers'23, QNDE'22]: Slides Link

Grain growth in polycrystalline materials is a complicated physical process. External heat or pressure application results in a particular grain orientation (texture). Identifying the texture property is crucial to understanding the grain growth mechanism. Grain images obtained from a sensing setup can be used for analysis. We demonstrate the usage of convolutional neural networks to classify grain images from a scanning electron microscope (sensing setup) in an automated manner. Further, the intermediate features of the convolutional network provide potential insights into the grain growth mechanism.