Fault diagnosis strategy for the current source section of a field-cycling nuclear magnetic resonance instrument

This paper proposes a fault diagnosis strategy to address catastrophic failures in all powercomponents of the current source of a field-cycling nuclear magnetic resonance (FC-NMR)instrument. The current source, implemented with a single power MOSFET operating in linearmode, is prone to thermal instability and degradation under high-current conditions, posingsignificant risks to system reliability. Due to the continuous conduction inherent in linear-modeoperation, fault signatures in the MOSFET could be subtle and difficult to distinguish fromnormal operational variations, making diagnostic methods relying on switching transientsineffective in this context. To overcome these limitations, an active fault diagnosis framework isintroduced to enhance fault detection and localization. This framework combines test signalinjection with data-driven artificial intelligence classifiers. Three algorithms—ResNet, aconvolutional neural network (CNN), and a nearest neighbor with dynamic time warping(NN-DTW), used as a benchmark—are evaluated using hybrid datasets derived from simulationprogram with integrated circuit emphasis (SPICE) simulations and experimental fault injections.The methodology employs time-domain signals measured at key circuit nodes, avoidingcomputationally intensive preprocessing steps. Simulation and experimental results demonstrateclassification accuracies of 100% for ResNet and NN-DTW, and 95.2% for CNN, withprediction times under 20 ms for neural networks. The proposal successfully diagnoses botheasy-to-detect faults, validated through simulation, and hard-to-detect faults, confirmedexperimentally. The entire fault diagnosis process is completed in under 15 s, making it suitablefor in-field monitoring of FC-NMR systems.