Search

Landmines are a major international problem, and humanitarian demining requires reliable detection of suspicious objects. Currently, commonly used landmine detection technologies combine metal detectors and ground-penetrating radar. Although these technologies are mature, there is a need to reduce false alarms with technologies that directly detect landmine explosives. Nuclear quadrupole resonance (NQR) is a radio frequency (RF) technology that can remotely detect certain compounds such as explosives. Although NQR is a type of magnetic resonance phenomenon, NQR uses the internal electric field gradient of a crystalline material, rather than an external static magnetic field, to initially align the nuclei, allowing for developing portable, low-cost detectors. A unique feature of NQR is that the detection is specific and less susceptible to false alarms since each material has a different resonance frequency and spectrum. Therefore, NQR is a promising approach to detect so-called "minimum metal" landmines, as it can look directly for their explosive content. This thesis describes the development and the signal processing implementation of a portable NQR device for anti vehicle landmine detection at King's College London. The aim of the project is to address the issue of the cost and complexity of the spectrometer electronics and especially the transmit power amplifier traditionally employed for NQR systems and investigate the improvements that machine learning techniques make to the low signal-to-noise ratio (SNR) of NQR signals. We have successfully built a hardware prototype of a portable NQR sensor for research department composition X (RDX), a type of explosive, and implemented the machine learning (ML) algorithm to improve the accuracy of the determination. This device can transmit pulses at the resonant frequency of RDX, acquire NQR signals from RDX samples, and determine the presence or absence of the target using the ML classifier stored in the device. Specifically, the portable NQR measurement system suitable for landmine detection was developed by using a low impedance circuit, different from the conventional 50 Ω circuits used so far. The highly efficient dual-supply class-D power amplifier, which is connected to the series resonance circuit, can be powered by batteries. The receiver amplifier was designed to maximize the SNR at the resonance frequency of the RDX, and a transmit-receive switching circuit was fabricated to combine it to the power amplifier. The signals of the RDX samples were then actually acquired both in the laboratory and in field trials to show the effectiveness of the device. This thesis also presents an approach to enhance the detection of buried landmines by applying ML to NQR signals from the developed hardware prototype. The conventional method for detecting NQR signals is based on comparing the spectral intensity of the acquired time waveform after fast Fourier transform (FFT) with a predefined threshold value. However, strong RF interference and a low SNR are important challenges for accurate detection in landmine detection. To tackle these problems, we have applied and tested various ML techniques to NQR signals acquired by our system in laboratory experiments and field trials. Results suggest that ML methods can indeed improve the detection accuracy of the NQR device. Importantly, the trained classifiers can be implemented with our device's field-programmable gate array (FPGA) architecture and can run with little time penalty compared with simpler but less-efficient FFT-based energy detection methods.