Bearing fault diagnosis based on spatial filtering constraint feature extraction and deep network

Abstract

As the core component of a gas turbine, the health condition of the main bearing has a crucial impact on the safe operation of the gas turbine. However, the distribution inconsistency problem existing in the time series data and the nonlinear coupling effect between the data will affect the extraction of characteristic fault features, leading to a decrease in diagnostic accuracy. To solve the above problems, this paper proposes a bearing fault diagnosis method based on frequency domain distribution filter and deep learning. Specifically, a sinusoidal Mel filter bank is designed to extract the fault features of low-frequency vibration signals based on the distribution principle of the traditional Mel filter bank. Then, considering that the fault information contained in other frequency bands of the vibration signals is also not easy to ignore, an inverse sinusoidal Mel filter bank is constructed to further mine the signature fault features of the vibration signals in the middle and high-frequency bands. Finally, the frequency domain distribution filter is proposed by combining the above two filters to reduce the impact of data distribution inconsistency on the accuracy of fault diagnosis. In addition, a deformable convolutional network is applied to further decouple the fault data and improve the spatial separability of the data features. Experiments with inconsistent data distribution are validated on two public datasets, and the diagnostic accuracy reaches 100%; the engineering reliability of the proposed method is verified on the main bearing of a gas turbine, and the accuracy reaches 99.86%.