Diagnóstico de falhas em aerogeradores usando sinais de vibração: uma abordagem integrada de classificação e previsão utilizando wavelets e aprendizado de máquina.

Abstract

Motivated by the high costs of corrective maintenance, the risks of unavailability, and the operational complexity associated with variable-speed operation in wind turbines, this work presents a systematic approach for bearing fault diagnosis and forecasting based on vibration-signal analysis, integrating signal processing, feature extraction, and machine-learning algorithms. Distinguishing among different bearing fault types is inherently challenging due to low signal-to-noise ratios under real operating conditions, modulation effects induced by variable speed, the presence of competing excitations from the drivetrain, and, most importantly, spectral overlap between fault signatures, which limits class separability and makes the discrimination of intermediate conditions more sensitive to data imbalance and operational variability. To address these limitations, a reproducible and implementation-oriented methodological workflow is proposed, in which key stages and design choices are explicitly organized in sequence: data acquisition, preprocessing, reconstruction, feature extraction and selection, model optimization, and fault classification. The preprocessing stage relies on wavelet decomposition and thresholding techniques to reduce noise while preserving informative components associated with impacts and relevant modulations. After signal reconstruction, 17 time- and frequency-domain features are extracted and used as inputs to five supervised models (Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), and Decision Tree (DT)), as well as to classifier committees formed by combining the most promising models. The methodology is applied to real wind-turbine signals in the forecasting stage and to experimental bench data in the classification stage. In the latter, the best configuration is obtained with the Db1 wavelet associated with Universal thresholding and the RF+KNN+MLP committee, achieving an accuracy of 0.983, thereby highlighting the potential of the proposed workflow to deliver high performance using relatively simple modeling components. In parallel, the forecasting task is formulated in two complementary settings: a multiclass formulation aimed at estimating the criticality zones defined by ISO 20816-21 and a binary formulation aimed at anticipating Zone D. In the multiclass setting, although overall accuracy remains high at a 1-day horizon (around 0.80–0.84), balanced accuracy typically ranges from 0.60 to 0.64, underscoring the difficulty of discriminating intermediate zones under imbalanced conditions and making explicit a practical limitation inherent to the problem granularity. As the horizon increases to 7, 15, and 30 days, a progressive degradation of performance is observed; in this context, applying Db1 with Universal thresholding (level 2) does not yield systematic gains and may increase sensitivity to the training set for some models, indicating a tendency toward overfitting. In the binary formulation, by contrast, the models exhibit more consistent performance when separating critical from non-critical conditions, with high accuracies at a 1-day horizon (typically between 0.82 and 0.89 in the wavelet-based scenario) and stable performance at 7, 15, and 30 days, showing greater alignment with the operational goal of anticipating critical conditions. Overall, by explicitly characterizing multiclass separability limitations and structuring a systematic pipeline, this thesis simplifies methodological decision-making, improves reproducibility, and delivers a technically robust and implementable solution for bearing diagnosis and criticality forecasting in wind turbines.