Early Fault Detection Based on Wind Turbine SCADA Data Using Convolutional Neural Networks

Abstract

Early fault detection in wind turbines using the widely available SCADA data has been receiving growing interest due to its cost-effectiveness. As opposed to the large variety of fault detection methods based on high resolusion vibration data, the use of 10-minute SCADA data alone does not require any additional hardware or data storage solutions and would be immediately implementable in most wind farms. However, the strong variability of these data is challenging and requires significant improvements of existing methods to ensure early and reliable fault detection and isolation. Here we suggest to use Convolutional Neural Networks (CNNs) to enhance the detection accuracy and robustness. We demonstrate the superiority of the CNN model over standard fully connected neural networks (FCNN) using examples for faults with very different time dependent characteristics: an abruptly evolving and a slowly degrading fault. We show that the CNN is able to detect the faults earlier and with a higher accuracy and robustness of prediction than the FCNN model. We then extend the CNN model to a multi-output CNN (CNNm) which provides early fault detection based on a multitude of output variables simultaneously. We show that with the same training time and a similar detection quality as the single output CNN, the CNNm model is an ideal candidate for a practical and scalable fault detection algorithm based on already available 10-minute SCADA data for wind turbines.