A Hybrid Empirical Mode Decomposition (EMD)-Support Vector Machine (SVM) for Multi-Fault Recognition in a Wind Turbine Gearbox

The increasing complexity of the wind turbine planetary gearbox and the ever-worsening operating conditions of off-shore wind turbines has made gearbox vibration-based fault diagnosis increasingly challenging. This quagmire has shifted research emphasis in recent times from traditional time-domain feature extraction to the use of more advanced signal processing tools. In this study, a hybrid empirical mode decomposition (EMD)-support vector machine (SVM) is proposed for multi-fault recognition in a wind turbine gearbox. The study investigates the impact of the SVM kernel and the EMD window size on the performance of the SVM model. The study results show that the size of the EMD window and the SVM kernel all impact the performance of the SVM model. The optimal EMD window was 500,000 samples. The radial basis function (RBF) and sigmoid kernel-based SVMs all scored 100% for accuracy, specificity, and sensitivity; however, the RBF kernel-based SVM resulted in the least training time, making it the best for the study. The high performance of the models makes them suitable for use as diagnostic tools for wind turbine gearboxes, allowing for precise fault diagnosis with minimum classification errors.

I. Introduction

Vibration monitoring in its most basic form employs statistical features derived from the raw vibration waveform to train conventional machine learning models (e.g. k-nearest neighbour, decision tree, random forest, support vector machine, and so on) for fault recognition [1]. The performance of the downstream model relies heavily on the quality of features extracted. Gearbox vibration signals are intricate, nonstationary, and nonlinear. When extracted from sensors mounted on real-world wind turbines, the signal is further complicated due to noise. Consequently, time-domain features extracted directly from the raw vibration waveform are not sufficiently discriminative to distinguish a faulty gearbox component from a healthy one. To solve this problem, advanced signal processing algorithms including discrete wavelets transform, continuous wavelets transform, variational mode decomposition, empirical mode decomposition, among others are used to extract meaningful features from the raw vibration waveform. The latter is then fed to the downstream model for fault recognition. In recent years, the use of empirical mode decomposition (EMD) in hybrid with conventional machine learning models has garnered significant interest.