Improving the Performance of Nuclear Quadrupole Resonance Sensing Technologies for in situ Detection of Narcotics and Explosives
2020-01-08T16:26:35Z (GMT) by
Nuclear Quadrupole Resonance (NQR) and Nuclear Magnetic Resonance (NMR) are spectroscopic techniques that offer the ability to characterise samples non-destructively in situ for quality control in industrial processes (e.g., water content in food), and for detection of explosives and narcotics in defence and security sensing applications. Current NMR-/ NQR-based sensing technologies can achieve good performance in a controlled laboratory environment where the effect of external Radio Frequency Interference (RFI) can be mitigated by RFI shielding. However, for in situ sensing applications outside the laboratory, complete physical shielding is often not possible or not practical (heavy, bulky) and therefore alternative methods are needed for NMR-/ NQR-based sensing technologies to be useful.
This EngD project focusses on the development of methods for active elimination of RFI for in situ sensing applications of NMR and NQR. The work aims to develop signal processing techniques that work with varying degrees of physical suppression (RFI shielding), to improve the accuracy and reliability of the NMR-/ NQR-based sensing technologies. The approach being developed is a machine learning process in which RFI is automatically identified using a decision tree model followed by a RFI suppression algorithm to produce the RFI-minimised signal.
This talk describes the development of an experimental testbed for data collection and some recent results achieved by applying the algorithms to measured NQR data subject to simulated burst mode RFI. The performance of the decision tree model is validated against human operator performance data, generated by volunteers rating the degree to which they could confidently declare a NQR signal present or not in RFI-polluted data. The performance of the model is quantified as a Receiver Operating Characteristic (ROC) curve, which plots true positive against false positive for a binary classifier. For the data described here, the decision tree model improved (relative to no RFI removed) the area under curve (AUC) value from 0.58 to 0.906, where AUC = 1 means a 100% detection rate with a 0% false alarm rate.
This work contains material subject to @Crown Copyright (2019), Dstl.