Fault diagnosis of wind turbine gearbox based on long term and short term memory network

Wind energy is a kind of clean and renewable energy, which develops rapidly in recent years, and the installed capacity continues to increase. Wind turbines are usually distributed in coastal areas, mountainous areas and other areas with abundant wind energy but poor natural conditions. Statistics show that gearbox is one of the main fault sources of wind turbine. Because it works under the condition of high speed and high load, long-term operation is prone to bearing wear, gear failure and other failures, which brings huge economic losses to the wind farm. Therefore, the study of reliable and effective gearbox fault prediction method is of great significance to improve the operation reliability and reduce the maintenance cost of wind turbine.

Gearbox is composed of bearings, gears and other components. In the process of its operation, all kinds of mechanical components have complex coupling relationship, so the various state signals often present nonlinear characteristics and contain a lot of noise, which makes it difficult to extract the early fault features of gearbox. For this reason, scholars at home and abroad have proposed a variety of different feature extraction methods, such as wavelet packet analysis, local time-frequency entropy, multi-layer denoising automatic encoder and so on. These methods require the secondary installation of vibration sensors on the wind turbine, and the cost is high; and the SCADA system is integrated into the wind turbine, so there is no need to install additional monitoring system, which can greatly save the cost.

Therefore, wind turbine condition monitoring based on SCADA operation data has become a research hotspot in the field of wind power in recent years. For example, the artificial neural network method is used to model the SCADA data to realize the early fault detection of the gearbox bearing; the Bayesian data-driven model is established based on the SCADA data to monitor and evaluate the health status of the wind turbine; Combining principal component analysis with extreme learning machine, the temperature monitoring model of wind power main bearing is established, and the fault of high temperature of main bearing is predicted in advance. The above methods have achieved good results in the condition monitoring of key components of wind turbine. However, these methods are still shallow learning models in essence, and they have limited modeling ability for complex condition monitoring data, and can not well mine the internal structure information and correlation of data, and their modeling accuracy needs to be further improved.