SAIL: A Scalable Wind Turbine Fault Diagnosis Platform: A Case Study on Gearbox Fault Diagnosis
Bahojb Imani, Maryam
MetadataShow full item record
Failure of a wind turbine is largely attributed to faults that occur in its gearbox. Maintenance of this machinery is very expensive, mainly due to large downtime and repair cost. While much attention has been given to detect faults in these mechanical devices, real-time fault diagnosis for streaming vibration data from turbine gearboxes is still an outstanding challenge. Moreover, monitoring gearboxes in a wind farm with thousands of wind turbines require massive computational power. In this paper, we propose a three-layer monitoring system: Sensor, Fog, and Cloud layers. Each layer provides a special functionality and runs part of the proposed data processing pipeline. In the Sensor layer, vibration data is collected using accelerometers. Industrial single chip computers are best candidates for node computation. Since the majority of wind turbines are installed in harsh environments, sensor node computers should be embedded within wind turbines. Therefore, a robust computation platform is necessary for sensor nodes. In this layer, we propose a novel feature extraction method which is applied over a short window of vibration data. Using a time-series model assumption, our method estimates vibration power at high resolution and low cost. Fog layer provides Internet connectivity. Fog-server collects data from sensor nodes and sends them to the cloud. Since many wind farms are located in remote locations, providing network connectivity is challenging and expensive. Sometimes a wind farm is offshore and a satellite connection is the only solution. In this regard, we use a compressive sensing algorithm by deploying them on fog-servers to conserve communication bandwidth. Cloud layer performs most computations. In the online mode, after decompression, fault diagnosis is performed using trained classifier, while generating reports and logs. Whereas, in the offline mode, model training for classifier, parameters learning for feature extraction in sensor layer and dictionary learning for compression on fog servers and decompression are performed. The proposed architecture monitors the health of turbines in a scalable framework by leveraging the distributed computation techniques. Our empirical evaluation of vibration datasets obtained from real wind turbines demonstrates high scalability and performance of diagnosing gearbox failures, i.e., with an accuracy greater than 99%, for application in large wind farms. ©2019, Springer Nature Switzerland AG.
Due to copyright restrictions full text access from Treasures at UT Dallas is restricted to current UTD affiliates (use the provided Link to Article).