A hybrid fault diagnosis method for rolling bearings based on GGRU-1DCNN with AdaBN algorithm under multiple load conditions

Abstract

Abstract The fault diagnosis of rolling bearings is a critical aspect of rotating machinery, as it significantly contributes to the overall operational safety of the mechanical equipment. In the practical engineering environment, the complex and variable working conditions, along with the presence of overlapping noise, contribute to intricate frequency information in the acquired signals and their highly time-dependent characteristics, which makes it difficult to extract the available fault features hidden in the signal. Based on this, a hybrid fault diagnosis method named GGRU-1DCNN-AdaBN is introduced, which combines improved gap-gated recurrent unit network (GGRU), one-dimensional convolutional neural network (1DCNN), and adaptive batch normalization (AdaBN). The proposed approach involves several parts to enhance fault diagnosis accuracy in vibration signals under constant load conditions and variable load conditions. Firstly, the end-layer structure of the traditional GRU is replaced with a one-dimensional global average pooling layer to aggregate the influence components of defects and reduce model training parameters. Secondly, the fusion of different types of frequency and sequence features is achieved by combining 1DCNN, addressing the limitation of a single network’s feature extraction capability and the loss of temporal features in a cascaded hybrid model. Subsequently, the fused features are input into a softmax multi-classifier to obtain fault type identification results. Lastly, the GGRU-1DCNN method is further improved by incorporating the AdaBN algorithm, enhancing the model’s domain adaptive capability under variable load conditions and noisy environments. The method is validated using datasets obtained from Case Western Reserve University, aero-engine bearings, Xi’an Jiaotong University, and the Changxing Sumyoung Technology. The findings suggest that the proposed method demonstrates superior accuracy and robustness in fault diagnosis, as well as excellent generalization capability and universal applicability.