Convolutional neural networks (CNN) are widely applied to predict the remaining useful life (RUL) of rolling bearings from time‑frequency representations (TFR) of vibration signals. During degradation, characteristic structures in TFRs align mainly along the frequency or time axis, making it difficult for conventional isotropic kernels to capture such directional patterns. Global average pooling (GAP) averages across all axes and may hide the locations of salient activations. To address these issues, we introduce a factorized‑axis convolutional gated recurrent unit (GRU) model. The architecture employs multi‑scale anisotropic convolutions and a dual‑axis convolutional block attention module to enhance directional features and highlight important time‑frequency regions. Dynamic adaptive pooling (DAP) adaptively aggregates information along the time‑frequency axes from the extracted feature maps, after which the GRU captures temporal dynamics in the latent representations and Monte Carlo dropout provides predictive uncertainty estimates. Experiments on two public bearing datasets show that the proposed model consistently outperforms existing RUL prediction methods across operating conditions. Ablation studies reveal that the factorized‑axis design yields lower mean errors than isotropic convolutions; DAP brings clear gains on one dataset and matches GAP on the other, underscoring the value of anisotropic feature extraction and adaptive aggregation for TFR‑based RUL prediction.
Review