|
Abstract:
|
Synthetic opioids (e.g., fentanyl) are responsible for recent surges in overdose deaths in the United States. The majority of these drugs are synthesized abroad and shipped or transported into the US. To help stem the flow of synthetic opioids into the US, our team is developing a nuclear quadrupole resonance (NQR) detector to screen packages for synthetic opioids. NQR is a chemical analysis technique that uses radio frequency pulses (which penetrate simple packaging) to detect materials non-invasively. However, the raw NQR measurements take the form of complex-valued time series that are frequently contaminated by highly nonstationary noise, complicating the use of existing detection techniques. In this work, we develop and apply statistical and machine learning approaches for denoising that operate in the complex domain. To address robustness to novel noise sources or changes to the signal that may arise during deployment, we evaluate these approaches on out-of-distribution data. Our results suggest that our proposed complex-valued networks (based on recent neural network architectures for real-valued speech enhancement) perform well while providing fast test-time predictions.
|
