Rolling Bearing Fault Diagnosis with Distribution Shift Data Using Improved Spatial Distribution Filters and Constraint Feature Extraction

The main bearing is a key component to maintaining the stable and safe operation of the gas generator power end, so fault diagnosis for the main bearing has special significance. To decrease the impact of data distribution shift on feature extraction performance, a symbolic feature extraction method based on improved spatial distribution Mel-Frequency Cepstrum Coefficient (IMFCC) is proposed. In addition, to extract constraint features more fully, fault data is fed into the spatial visual attention mechanism network in the form of patch blocks. Specifically, this paper first uses the Fourier transform to perform spectrum analysis on the original five fault types of vibration signals. Based on analyzing the frequency band range with high discrimination of different fault types, IMFCC filter banks are constructed to enhance the representation ability of symbolic features and improve fault diagnosis performance under data distribution shift. Secondly, the extracted features are used to construct a three-dimensional (3D) spatial sample dataset. The dataset is split into a combination of patch blocks along the spatial dimension. Finally, combining spatial attention network and distribution attention network to extract the spatial constraint feature and distribution constraint feature of patch blocks. Two public rolling bearing datasets are implemented to demonstrate the effectiveness of the proposed method. Subsequently, the proposed method is applied to the fault diagnosis of the main bearing from a waste landfill gas generator in real industrial environment, which proves the feasibility of the method.