Speaker
Description
Maintaining production equipment in manufacturing is critically important. As the number of monitored machines increases, the computational cost of conventional machine learning approaches grows rapidly. In vibration and acoustic analysis, methods combining autoregressive (AR) model coefficients with one-class support vector machines (OCSVMs) are widely used for anomaly detection. This study extends our previous work on single-anomaly detection to simultaneous multi-anomaly scenarios, investigating the advantages and limitations of quantum kernel methods across both settings.
Five miniature pieces of semiconductor factory equipment were prepared, with ten anomaly types intentionally induced and recorded via microphone. Audio data were segmented, preprocessed, and fitted to AR models; feature vectors were constructed from AR coefficients. Anomaly detection was performed using OCSVMs with quantum kernels and evaluated using F0.5, F1, and F2 scores. A modified Mahalanobis metric space was employed to visualize and distinguish anomaly types.
For individual anomaly detection (10 types), quantum kernels achieved near-perfect F0.5, F1, and F2 scores. For simultaneous two-anomaly detection (5 selected pairs), F2 scores exceeded 0.8, demonstrating the advantage of quantum kernel methods over classical RBF kernels. Mahalanobis visualization confirmed that anomaly points were separably distributed. However, simultaneous anomalies did not appear at the simple superposition of individual positions, suggesting nonlinear acoustic interactions within the measurement environment.
Future work will focus on improving separation of proximate anomaly clusters and characterizing these interactions. This poster synthesizes findings from both detection regimes and discusses open challenges toward practical deployment.
| Topic | Quantum kernel methods and their applications |
|---|