Non Invasive Fault Detection of Offshore Wind Turbines Using Deep Network-Based Thermogram Features

The offshore regions typically experience greater wind speeds, which makes offshore Wind Turbines (WTs) more efficient. This enhanced output comes with a price, including more maintenance requirements, greater proneness to malfunctions, and difficulties with accessibility. It is challenging to gather the signatures with invasive sensors from offshore WTs for a long time due to the surrounding conditions. Sensors get displaced from their positions, and the acquired data becomes less reliable. The issues with invasive sensors in offshore locations for WTs are addressed in the study by presenting a non-invasive method for fault detection in offshore WTs using thermography. The approach is able to classify 11 different health conditions of WTs such as healthy, single-multiple phase stator faults with different shorting percentages, cooling fan faults, and rotor faults. Serial Based Feature Fusion technique is proposed in which features are extracted from seven Pre-Trained (PT) models and fused to get a feature set with advantages of individual PT model. The problem of high processing time and complexity a Hybrid Feature Selection technique is proposed in which the feature selection is done in two stages along with hyperparameter tuned Shallow Learning (SL) Classifier at the output layer. The algorithm is tested for multiple combinations of DNN and SL approaches. The highest achieved efficacy is 100%. By using feature set with best possible feature, the suggested model is more reliable. Additionally, it eliminates the necessity for segmentation and clustering, which reduces the computational burden and time required for diagnosis.