Non-stationary vibration data analysis using deep learning algorithms boosted by kernel density estimators

Wind turbines are essential for incorporating sustainable measures in energy production. Harmful effects caused by vibrations in turbine shafts and associated components can be mitigated or anticipated via spectral analysis of the non-stationary vibration signals collected from each component of interest. This study aims to propose a vibration signal filtering technique using a kernel density estimator and windowing technique to remove noise from non-stationary signal harmonic components. The tool was evaluated using three distinct classifiers (ANN, KNN, and SVM), which demonstrated improvements from 98.75% to 100%, 73.13% to 79.38%, and 81.88% to 91.25%, respectively, after applying KDE filtering.