Mitigating High Dimensionality and Few Training Data with a Quantum Kernel Learning Approach

• Mauro Nooblath UFAM / SENAI CIMATEC
• João Bessa UFAM
• Rosiane de Freitas UFAM

Resumo

In this work, we investigate the integration of quantum algorithms as kernel functions in Support Vector Machines (SVMs) to classify highly imbalanced datasets, using a case study on anomaly detection in wind turbine sensor data. A hybrid quantum-classical computational strategy was developed and benchmarked against four classical kernels, demonstrating superior or competitive performance in several scenarios. The evaluation focused on the model’s generalization ability to detect minority classes in predictive maintenance datasets. The main contributions of our research include: (i) the impact of kernel complexity on minority class detection, (ii) the comparative performance of quantum and classical models in identifying fault types, and (iii) the influence of different quantum entanglement strategies on SVM performance. Simulation results using the Qiskit (Quantum Information Software Kit) framework highlight how superposition and entanglement can effectively map data into higher-dimensional spaces, improving hyperplane separation and enhancing fault detection.