حسین حق بین

Obtaining sufficient labeled data for machinery fault diagnosis in real industrial settings remains a major challenge. Lightweight deep learning models often struggle to deliver reliable diagnostic performance under noisy conditions. To address these issues, a novel lightweight CNN framework, termed Depth Separable Interaction CNN (DSICNN) is proposed. At its core, the Depth Separable Interaction Convolutional Module (DSICM) integrates depthwise convolution, pointwise convolution, and global average pooling to enhance feature extraction and receptive field expansion, while substantially reducing parameter count and computational cost. Experimental results on the CWRU bearing dataset demonstrate that the proposed DSICNN outperforms the baseline CNN and some other CNN based models in diagnostic accuracy and robustness, while preserving a lightweight architecture suitable for practical industrial deployment.