Estratégias de aprendizado para detecção de falhas aprimoradas por kernels quânticos

Abstract

This master’s thesis investigates the use of quantum kernel functions in hybrid classical–quantum models for fault detection in wind turbines. Parameterized quantum circuits were used to map input data into high-dimensional Hilbert spaces. The quantum circuit architectures analyzed include ZZFeatureMap, Re- alAmplitudes, and EfficientSU2, each implemented with four entanglement strategies: Linear, Full, Circular, and Shift-Circular-Alternate. These were com- pared with classical kernels—Linear, Polynomial, Radial Basis Function, and Sigmoid—within a Support Vector Machine framework. The dataset, comprising 54 features collected by turbine sensors, was reduced via Principal Component Analysis to 4, 8, and 16 components based on cumulative variance. The RealAm- plitudes circuit with Full entanglement and 16 components outperformed the Radial Based Function kernel in standard machine learning metrics. ROC-AUC curves and confusion matrices showed no overfitting, reinforcing the potential of quantum kernels in industrial fault detection.