Combining Edge and Cloud Computing for Monitoring a Fleet of Wind Turbine Drivetrains Using Combined Machine Learning Signal Processing Approaches

Abstract

Today, we are at the beginning of Industry 4.0. Machines are increasingly sensorized and equipped with internet connection. Especially the emergence of 5G is a game changer in this regard. It becomes possible to send data at high speeds to cloud computing data-centers. However, streaming all data is deemed to be unfeasible. It is more advantageous to use the additionally available bandwidth to drastically increase the number of connected sensors. Thus, on-board processing of the data directly at the edge is necessary. Preprocessed data is then sent to a cloud data-center for processed using computationally intensive algorithms. An application that can leverage this new context is condition monitoring (CM). By balancing the edge-processing with central processing the detection potential and insights in system behavior can increase. Advanced features are extracted at the edge. Long-term trending, anomaly detection and learning approaches are performed centrally together with more computationally intensive algorithms. This paper illustrates an integrated monitoring approach for wind turbines exploiting this Industry 4.0 context. Our combined edge-cloud processing approach is documented. We show edge processing of vibration data captured on a wind turbine gearbox to extract diagnostic features. Focus is on statistical indicators. Reallife signals collected on an offshore turbine are used to illustrate the concept of local processing. The NVIDIA Jetson platform serves as edge computation medium. Furthermore, we show an integrated failure detection and fault severity assessment at the cloud level. Health assessment and fault localization combines state-of-the-art vibration signal processing on high frequency data (10kHz and higher) with machine learning models to allow anomaly detection for each processing pipeline. Again this is illustrated using data from an offshore wind farm. Additionally, the fact that data of similar wind turbines in the farm is collected allows for exploiting system similarity over the fleet. Again this is illustrated using data from an offshore wind farm.