Detecção de Folga em Cunhas de Geradores Hidrelétricos via Monitoramento de Parâmetros Vibratórios

Abstract

The integrity and efficiency of hydroelectric generators are directly dependent on the stator bar fastening system, composed of wedges and ripple springs. This assembly compensates for thermal expansion and geometric tolerances; however, its loosening constitutes a critical failure. Insufficient clamping allows vibrations under electromagnetic forces, accelerating insulation degradation and intensifying eddy current losses. Consequently, the generator experiences a reduction in its energy conversion capacity. To address this issue, this work proposes a methodology for classifying the tightness level of magnetic wedges through the monitoring of vibratory parameters. The approach integrates experimental and numerical procedures using a simplified system (200 mm wedge and 1 mm top ripple spring). Initially, the dynamic characterization of the assembly was performed on an experimental bench under white noise excitation to identify its natural frequencies. In parallel, a Finite Element Method (FEM) model was developed and calibrated via parametric optimization of mechanical properties. For the wedge, a Young’s modulus (E) of 35 GPa and a Poisson’s ratio (ν) of 0.2 were obtained; for the spring, E = 40 GPa and ν =0.2. With the calibrated model, transient simulations generated 100 acceleration signals for each condition: “loose” (clamping force from 5 to 11 N) and “tight” (11 to 15 N). The final stage involved the training and validation of three classification algorithms using experimental data: k-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Autoencoder. The results demonstrated high similarity between synthetic and real signals. The SVM algorithm showed the best performance with 95% accuracy, followed by KNN (92%) and Autoencoder (83%), confirming the potential of the methodology for detecting looseness in wedges.