Efficient condition monitoring of off shore wind turbines using deep networks

The wind turbines (WT) are installed both onshore and offshore. The offshore WTs are more productive as compared to the onshore WTs as the availability of the wind speed is much higher offshore, but this productivity comes at the cost of an increase in faults, a higher maintenance requirement, and difficulty in accessibility. It is very difficult to gather the signatures with invasive sensors from the offshore WT due to the surrounding conditions. The position of the sensors changes, leading to unreliable signatures. A non-invasive sensor-based infrared thermography technique has been proposed to solve this problem. In this work, an approach for non-invasive condition monitoring of offshore WT for inter-turn short circuit (ITSC) faults has been developed using deep neural network (DNN)-based features of thermograms. Pre-trained deep learning (DL) models such as Alex Net have frequently been used for the diagnosis of motor faults. These pre-trained DL models are tested on particular data, which has a different classification problem and does not provide a reliable and high-efficacy model. Along with the previous drawback, these models also suffer from high processing times. To solve these problems, in this work, pre-trained models are modified with the incorporation of the feature selection technique and a shallow learning classifier at the output layer. The proposed methodology is able to classify the ITSC, rotor, and cooling fan faults with an accuracy of 99.7%.