Radio spectrum is a limited national resource whose efficient utilization is of strategic importance. With the rapid advancement of wireless technologies, maintaining spectrum cleanliness and enabling fast and reliable anomaly detection have become critical challenges. Artificial intelligence (AI)-based approaches have recently shown great promise in addressing these issues; however, their effectiveness strongly depends on the availability of high-quality, representative, and annotated datasets. Generating such datasets is a complex task, further complicated by environmental conditions such as weather. Hungary’s nationwide spectrum monitoring network enables continuous observation of frequency bands, thereby providing the opportunity to construct large-scale and sustainable datasets. This study introduces a measurement methodology designed for the FM sound broadcasting in the VHF band (87.5–108 MHz), presents the resulting dataset, and details the annotation process. The published, openly accessible dataset is expected to serve not only as a valuable reference point but also as a benchmark for the international research community, facilitating the development, validation, and objective comparison of AI-driven spectrum monitoring solutions.

Automated anomaly detection is increasingly important in radio spectrum monitoring, where large volumes of continuously generated measurements must be analyzed while anomalous events remain relatively rare. This study investigates reconstruction-based anomaly detection for real-world very-high-frequency (VHF) Band II radio-broadcasting spectrum measurements using autoencoders whose gradient-based parameter optimization was performed exclusively on normal observations. Measurements covering 87.5–108 MHz in VHF Band II were represented as time–frequency images and processed through a unified pipeline comprising preprocessing, representation learning, reconstruction, residual-based scoring, and anomaly decision. The effects of input normalization, image resolution, network depth, latent-space dimensionality, encoder–decoder symmetry, spectrum representation, and reconstruction scoring were systematically evaluated across symmetric, asymmetric, stacked, augmented stacked, and pixel-sorted autoencoder configurations, while exploratory CVAE (convolutional variational autoencoder) experiments were additionally used to examine probabilistic latent regularization. Increasing architectural complexity did not necessarily improve detection performance. The best configuration identified during model development achieved an $F_1$-score of 91.4% on the labeled development subset at its $F_1$-optimized operating point. To obtain an independent assessment of generalization, the finalized pipelines were subsequently evaluated on a separate held-out test subset that was excluded from model, checkpoint, scoring-hyperparameter, and threshold selection. The held-out evaluation showed lower absolute performance and a change in configuration ranking, demonstrating that development-set differences should not be interpreted as definitive evidence of architectural superiority. A controlled reconstruction-scoring ablation further showed that anomaly separability depends materially on the formulation of the observation-level reconstruction score.

This paper presents a study of three machine-learning approaches for anomaly analysis in spectrum-monitoring data from VHF Band II radio broadcasting. Spectrum anomalies are defined as deviations from expected spectral behavior, including unexpected signal emissions, interference, and irregular variations in signal characteristics. A supervised YOLO-based object detector was evaluated for anomaly localization using independently produced expert annotations, while convolutional autoencoder-based clustering and long short-term memory (LSTM) prediction were investigated for exploratory spectral and temporal analysis, respectively. The supervised experiments examined both within-annotator and cross-annotator performance, whereas the unsupervised and temporal approaches were assessed qualitatively. The autoencoder-based approach revealed recurring spectral structures in latent space, and the LSTM model identified deviations from the recent behavior of individual broadcast transmissions. The principal quantitative results showed within-annotator $F_1$ scores ranging from 0.85 to 0.94, compared with 0.34 to 0.77 in cross-annotator evaluations. These findings indicate that the reported anomaly localization performance is sensitive to the reference annotations, highlighting the importance of annotation consistency and independent evaluation in spectrum-monitoring applications.