Compound Defect Diagnostics of Swind Turbine Gearboxes Via Variational Mode Decomposition, Hjorth Parameters and Machine Learning

Abstract

This study aims to develop a robust and automated diagnostic framework for identifying and classifying compound defects in WT gearboxes. The proposed methodology adopts a multi-phase approach. Vibration signals corresponding to various compound defect scenarios across multiple gearbox stages are initially collected from a WT gearbox system. From these signals, Hjorth parameters (activity, mobility, complexity), representing time-domain features, are computed for all gearbox health conditions. Simultaneously, the signals undergo variational mode decomposition (VMD) processing to extract intrinsic mode functions (IMFs). A comprehensive set of statistical features encapsulating time-frequency domain characteristics is subsequently derived from these IMFs. Data augmentation is employed to enhance the robustness and discriminatory power of the dataset, leading to the formulation of an enriched feature set by integrating the Hjorth parameters (time-domain features) with the statistical features (time-frequency domain features). The feature set is subjected to machine-learning-based multi-class classification to diagnose compound gearbox faults. The study systematically evaluates the diagnostic performance of the augmented feature dataset and benchmarks it against individual feature sets. The findings underscore the effectiveness of the proposed methodology in accurately diagnosing compound faults with reduced dependency on complex signal post-processing techniques, thereby advancing the state of fault diagnostics in WT gearboxes.