First artificial intelligence-based non-invasive framework for data-driven kink defect detection in HTS coils for power applications

High-temperature superconductor (HTS) coils and windings are fundamental structural elements underpinning a wide range of superconducting technologies and are the crucial components for next-generation superconducting devices across electric transportation and energy systems fields, including fusion reactors for sustainable energy generation, aviation propulsion systems aiming for substantial efficiency gains, and renewable electric power generators integrating superconducting technologies. Nevertheless, the manufacturing processes of these HTS windings commonly introduce localized mechanical stresses that lead to subtle initial bending defects, so-called "kinks", potentially evolving into a risk of catastrophic failure, and significantly compromising operational reliability. Technically, conventional kink detection techniques, such as optical microscopy, magnetic imaging, and metallurgical techniques, performed deployment challenges of invasive inspections, low inspection efficiency, and insufficient sensitivity to subsurface anomalies when handling complex coil geometries. To address these challenges, this study develops the first non-invasive electrical detection method that autonomously identifies incipient kink defects, leveraging an innovative artificial intelligence (AI)-enabled framework that combines frequency domain analysis with adaptive machine learning classifiers. To experimentally validate the proposed framework, controlled kink defects were induced into a Bi-2223 superconducting coil under precise experimental conditions, replicating realistic deformation scenarios encountered during coil winding processes. Voltage signals from coils exhibiting induced kink defects and pristine coils were collected and transformed into discriminative spectral features through effective frequency-domain analysis. These spectral characteristics are further refined by pinpointing essential components, thus achieving effective data dimensionality reduction. Subsequently, a K-Nearest Neighbors classifier, enhanced by adaptive distance metrics and robust cross-validation protocols, was employed, attaining a remarkable detection accuracy of up to 98.9%. Critically, acknowledging practical engineering constraints such as cost efficiency, compatibility with industrial monitoring units, and lower sampling requirements, the developed method successfully maintained an acceptable detection rate. Ultimately, this study establishes the proposed analysis paradigm as a proof-of-concept of a non-invasive diagnostic tool for identifying early-stage mechanical defects. While further validation case studies across different coil designs and operating conditions is required for full industrial generalization, the present study highlights the potential relevance of data-driven electrical diagnostics for improving the manufacturing and maintenance capabilities of HTS coil winding processes. This model contributes to enabling the next generation of superconducting technologies, with prospective applications in electric transportation and energy systems.