Speaker
Description
Semi-supervised anomaly detection in multivariate time series requires models capable of learning normal system dynamics while remaining efficient enough for deployment in resource-constrained environments. We investigate two lightweight quantum machine learning architectures that address this challenge through complementary paradigms: Quantum Reservoir Computing (QRC), which exploits the nonlinear dynamics of a fixed transverse-field Ising reservoir to generate expressive temporal representations without optimizing the quantum substrate, and a quantum-enhanced Transformer (iQTransformer), where variational quantum circuits replace the linear query, key, and value projections within the self-attention mechanism. Both models are trained exclusively on normal operating data and perform anomaly detection through forecasting-based residuals and representation-aware anomaly scoring. Their performance is evaluated on the GECCO benchmark using both point-wise and event-based metrics and compared with equivalent classical baselines. The results indicate that quantum-enhanced feature representations consistently improve the discrimination between normal and anomalous operating regimes while maintaining compact architectures with substantially fewer trainable parameters. In particular, the QRC framework achieves competitive performance using only a lightweight classical readout, whereas the iQTransformer provides enhanced predictive capabilities with approximately half the trainable parameters of its classical counterpart. These findings suggest that lightweight quantum architectures constitute a promising direction for scalable and efficient multivariate time-series anomaly detection, combining expressive temporal modeling with reduced training complexity.
| Topic | Quantum algorithms for machine learning applications |
|---|