<p>Fault diagnosis in rotating machinery faces significant challenges in strong noise environments. Especially under extremely high noise intensity and unknown noise types, existing methods struggle to maintain accuracy. We propose the Improved Residual Attention Convolutional Neural Network (IRA-CNN) to address strong noise problem. IRA-CNN integrates the interconnected multi-branch structure and the mixed attention mechanism specially designed for vibration signals. Unlike previous studies that only consider Gaussian noise and signal-to-ratio &gt; − 6, we evaluate the model's noise robustness by extensive experiments across three datasets, three noise types, and six noise intensity levels. The results reveal that the noise type significantly impacts model performance which has often been overlooked in previous studies. IRA-CNN outperforms state-of-the-art models in both accuracy and generalization. These findings establish a highly effective solution for fault diagnosis in challenging strong noise environments.</p>