A feature extraction and machine learning framework for bearing fault diagnosis

Wind power generation has been widely adopted due to its renewable nature and decreasing capital cost per kW. However, existing equipment ages rapidly, leading to higher failure rates, greater operation and maintenance costs, and worsening safety conditions, calling for improved condition monitoring and fault diagnosis for wind turbines. Past methods utilize physical models, but they are only successful in laboratory environments. As increasing data are becoming available, there are methods applying machine learning without careful discrimination, leading to low accuracy. To solve this problem, first this paper proposes to conduct unsupervised learning to understand data properties, e.g., structural density. Subsequently, the sensitivity analysis is conducted to extract the significant features and to avoid overfitting. The sensitivity of various features that are characteristics of wind turbine bearings may vary significantly under different working conditions. During such a process, the piece-wise properties are studied to improve supervised learning. By combining the properties of data and regression, a three-stage learning algorithm is proposed to refine and learn the most useful information for turbine bearing fault diagnosis. The proposed framework is validated by using real data from diversified data sets for nonstationary vibration signals of bearings.