<p>Autonomous manufacturing relies on intelligent and automatic fault diagnosis to enable predictive maintenance and minimize manual intervention. Rolling bearings, as critical components widely used in manufacturing equipment, pose a significant challenge for fault diagnosis in actual production environments, particularly under strong noise conditions. Traditional signal processing methods heavily depend on expert experience and often suffer from poor generalization, while deep learning models lack interpretability and experience substantial performance degradation in noisy environments. To address these limitations, an attention-driven hybrid model combining improved ensemble noise-reconstructed empirical mode decomposition (IENEMD) and parallel multiscale convolutional neural network (CNN) is proposed for automatic fault diagnosis of rolling bearings under noisy conditions. Specifically, raw vibration signals are automatically decomposed into intrinsic mode functions (IMFs) using IENEMD, effectively leveraging and reducing inherent noise in the signals. These raw signals and IMFs are then simultaneously processed by an attention-driven parallel multiscale CNN, which adaptively amplifies informative components to generate robust fault features. Finally, an adaptive global average pooling module identifies bearing faults with high accuracy. Three comprehensive case studies, involving eight comparative methods and nine different signal-to-noise ratio (SNR) levels, validate the superior performance of the hybrid model. Moreover, the model’s working mechanism is elucidated using Hilbert envelope analysis, enhancing the trustworthiness of emerging deep neural networks and bridging the gap between these advanced techniques and traditional signal processing methods.</p>