Spectral and Topological Representation Framework for Deep SVDD-Based Anomaly Detection in Hpc Systems

Operational stability assurance depends on highperformance computing (HPC) systems, which produce complicated multivariate telemetry data. This paper develops an unsupervised anomaly detection system through Deep Support Vector Data Description (Deep SVDD) with topology-based and spectral characteristics appended. The structural information recovery of topological data analysis through persistent homology operates globally, but the discrete wavelet transform enables precise detection of temporally localized frequency patterns. A convolutional neural encoder produces compact hypersphere embeddings that combine features obtained from processed data with raw metrics. The system detects anomalies through any departure from its standard representational format. Empirical tests on Marconi100 Tier-0 supercomputer telemetry data show regular node-wise and global ROC AUC scores over 0.85, demonstrating a notable performance improvement exceeding 9% compared to previous state-of-the-art methods. A visual presentation of latent space structures together with the distribution of anomaly scores helps explain the model's performance during the processing of data disrupted by temporal gaps and workload variations. The presented framework delivers precise anomaly detection through its unified methodology, which unites multiscale analysis with topology and space-time examination while maintaining operational high-performance computing (HPC) compatibility.