A novel method of wind turbine bearing fault diagnosis based on gradient projection sparse reconstruction and fractional Fourier transform and improved convolutional long short-term memory

Abstract

The bearing is one of the most critical components of a wind turbine, and its condition determines whether the wind turbine can work normally or not. And in recent years, the increasing number of wind turbines leads to more and more data collected for bearing condition monitoring, which brings a great challenge to data transmission and preservation. In addition, the vibration signal is one-dimensional, and directly using one-dimensional signals can avoid the complex calculations caused by increasing the dimensionality of the data. In this article, a new fault diagnosis method for wind turbine bearings is proposed using one-dimensional data as input. The method is based on sparsely reconstructed gradient projection and fractional Fourier transform (GPSR-FRFT) for feature extraction of the signal, and a modified convolutional long short-term memory (CNN-LSTM) is used for fault classification of the data. Firstly, the signal is sparsely expressed by the GPSR algorithm, and the original signal is compressed and observed using compressed sensing so that the observed signal can be transmitted instead of the original signal, this process can reduce data samples to save storage space. The original signal is recovered by the inverse solution of compressed sensing at the terminal, and this step can reduce noise interference. Secondly, the time-frequency characteristics of the signal are obtained using the FRFT transform. In addition, in order to solve the problem of insufficient learning ability and low accuracy of convolutional long short-term memory (CNN-LSTM), an improvement was made by adding a jump connection layer to the convolutional layer to improve the training speed and learning ability. The resulting signal is input to an improved CNN-LSTM network, and experimental analysis proves the superiority of the proposed method.