An Improved Domain Adaptive Method for Early Weak Fault Diagnosis of Bearings Based on Sample Compactness and Anomalous Feature Filtering

The health monitoring of mechanical equipment is evolving towards intelligence, and early weak fault diagnosis of bearings based on transfer learning can achieve fault warning under different working conditions and equipment. However, extracting weak features is difficult and corresponding knowledge is lacking for fault diagnosis. The erroneous correspondence generated by anomalous features in the source and target domains can negatively affect the transfer process. In this paper, an early weak fault diagnosis method based on sample compactness and abnormal feature filtering is proposed. This method is used to effectively extract features from weak faults and filter anomalous features during domain adaptation. Initial noise reduction of the bearing signals is first performed by envelope spectral analysis, after which the vibration signals are converted into grey-scale images by means of the Gramian angular field. After enhancing weak faults through the above steps, Gabor convolutional neural network (Gabor-CNN) is used as the backbone network to extract weak fault features. The target domain data is assigned pseudo labels by the backbone network. Then sample compactness is proposed to measures the distribution of initial target domain samples with the same pseudo labels to determine the effectiveness of model classification, and the network model is optimized in reverse. Finally, the centroid average distribution and filterable conditional distribution are defined to suppress the negative impact of anomalous features in the domain adaptation process. The proposed method achieves high accuracy in cross condition experiments and transfer experiments from artificial faults to degraded faults between different devices.