Abstract: In the complex space radiation environment, commercial off-the-shelf (COTS) semiconductor devices are increasingly susceptible to micro single-event latch-up (μ-SEL). This phenomenon disrupts localized circuit functionality, leads to abnormal current flows, and can even result in the complete failure of electronic systems. Current detection methods for μ-SEL primarily rely on contact-based measurements of device current and the use of ground truth labels associated with specific faults. However, these methods face significant limitations, such as the need for hardware modifications and the reliance on labeled data, making them unsuitable for the practical demands of complex space environments. To address these challenges, this paper proposed a non-contact μ-SEL detection method based on machine learning. A non-contact signal acquisition platform was developed using a near-field electromagnetic probe and a software-defined radio (SDR). This platform was used to collect electromagnetic signals radiated from a chip undergoing laser-induced μ-SEL. The modulation characteristics of the radiated signals were analyzed, and IQ modulation was employed for signal acquisition. This approach preserves the integrity of the signal while reducing the complexity of data processing. After signal preprocessing, a comprehensive dataset was generated for further analysis. In terms of feature extraction, multiple dimensionality reduction algorithms, including PCA, MDS, PP, t-SNE, Isomap and UMAP, were applied to reduce data dimensionality and remove noise. Among these algorithms, PCA demonstrates the best performance in retaining key features. Furthermore, five supervised machine learning algorithms, including KNN, decision trees, and random forests, were utilized to evaluate the dataset. Experimental results show that the supervised algorithms achieved a maximum detection accuracy of 99.88%, validating the feasibility and high efficiency of the proposed non-contact approach. In addition, to overcome the dependency on labeled data in supervised methods, this study introduced unsupervised detection algorithms, focusing on the optimization of K-means, spectral clustering, and DBSCAN. Multi-parameter optimization strategies were employed to determine the optimal parameters for each algorithm. These included evaluating different distance metrics for K-means, employing an enumeration method to optimize the gamma parameter for spectral clustering, and using the K-distance graph to determine the optimal eps value for DBSCAN. These optimizations significantly improve the performance of the unsupervised detection algorithms. Notably, DBSCAN achieved a detection accuracy of 92% without requiring labeled data or additional hardware modifications, demonstrating the robustness and practicality of the proposed method in complex environments. In conclusion, the proposed machine learning-based non-contact μ-SEL detection method provides a novel technical pathway and practical solution for efficient and accurate μ-SEL detection. This approach offers significant potential for improving the reliability and robustness of space electronic systems in challenging environments.