Physics-guided quench detection framework for HTS coils based on augmentation of artificial quench data: a comprehensive benchmarking investigation on 75 different machine learning and dimensionality reduction techniques

Abstract High-temperature superconducting (HTS) technology is a cornerstone technology poised to revolutionise critical sectors, including fusion energy, sustainable aviation, and high-efficiency renewable power generation. Nevertheless, the immense potential of these systems could be critically undermined by the risk of quench—a rapid, initially localised transition from superconductivity to high-resistivity normal state, potentially posing a significant threat to operational safety and system longevity. Conventional quench detection techniques, which rely on simple transitions above voltage or temperature thresholds, could be inadequate for critical applications of superconducting technology. They struggle with delayed response times, susceptibility to noise, and an inability to discern subtle precursor signals, potentially leading to either missed events or costly false alarms. To address these challenges, this paper presents a novel physics-guided, artificial intelligence (AI)-driven diagnostic framework that goes beyond simplistic time-domain analysis. We introduce the Inter-Harmonic Amplitude Ratio (IHAR) as a highly discriminative, frequency-domain signature derived from raw measurement signals, offering exceptional robustness against noise and operational fluctuations. To extract the most critical quench indicators from IHAR data, advanced dimensionality reduction (DR) techniques were used to optimise the feature set. A comprehensive benchmark of 15 advanced machine learning classifiers was then performed on the results of 5 DR techniques, which essentially means that 75 different diagnostic combinations have been covered. A framework based on an adaptive boosting (AdaBoost) classifier together with linear discriminant analysis (LDA) DR technique demonstrated superior performance, achieving a detection accuracy of 0.9861 for identifying quench events within a few milliseconds. To validate its effectiveness for quench detection and its feasibility for potential future industrial applications, the framework’s robustness was confirmed through a noise disturbance study, where the AdaBoost + LDA model maintained high accuracy (0.8056) even with signal noise of up to a signal-to-noise ratio of 0 dB. Furthermore, it demonstrated excellent generalisation by achieving the highest classification accuracy (0.875) using data from a different pancake coil. By enabling reliable identification of quench events, this AI-powered framework significantly enhances the operational safety of HTS systems. This work paves the way for the widespread adoption of superconducting technologies in mission-critical applications, from electric aircraft to fusion reactors, ensuring they operate not just with high efficiency but with high reliability.