Análise de dados de vibração não-estacionários em aerogeradores usando algoritmos de aprendizagem profunda potencializados por estimadores de densidade kernel

Abstract

Wind turbines are essential components for incorporating sustainable measures in energy production. However, the effectiveness of condition monitoring systems (CMS) is often challenged by stochastic and broadband noise, which can mask harmonic fault components in vibration signals. Seeking to solve this problem, signal filtering methods can be used in conjunction with multiple approaches to pattern recognition techniques. Classical machine learning models (KNN, SVM, ANN) can be used to evaluate the filtering impact on sets of statistical features extracted from the spectrum. In parallel, deep learning techniques, specifically Convolutional Neural Networks (CNNs), can be applied for end-to-end classification directly from spectrograms. This thesis proposes and validates a new spectral preprocessing methodology, based on Kernel Density Estimation (KDE), to selectively remove noise and optimize the separability of fault features. The validation was structured in two phases: a Phase 1 of controlled validation on the VBL database (comparing KDE, DWT, and raw data) and a Phase 2 of real-world application on the EISLAB database (anomaly detection in wind turbines). Both training paradigms were explored: Supervised Learning, where the model learns to map inputs to known fault labels, and Self-Supervised Learning (via SimCLR), focused on learning robust data representations without using labels. The results from Phase 1 (VBL) demonstrated the effectiveness of KDE in optimizing robust models (SVM/ANN) and, more expressively, in CNN models, where the KDE + Supervised CNN combination achieved the highest validation accuracy (98.5%), surpassing the raw data (87.0%) and the DWT (96.0%). In Phase 2 (EISLAB), the application of the KDE filter to real-world wind turbine data again proved superior, increasing the Supervised CNN accuracy from 68.5% to 72.3% and, crucially, raising the F1-Score for the fault class from 0.554 to 0.600. It is concluded that adaptive filtering via KDE is an effective methodological tool, acting as a spectral feature optimizer, significantly increasing the precision and robustness of artificial intelligence-based diagnostic systems. The experiments showed that KDE significantly improves robust models, especially CNNs, achieving the best results both in a controlled environment and with real-world wind turbine data. Thus, the method stands out not only as a filter but as a spectral feature optimizer, enhancing accuracy and reliability in fault diagnosis for rotating systems.