<p>In industrial applications, the vibration signals of rolling bearings are often subjected to strong noise interference, variations in operating conditions, and fluctuating rotational speeds, resulting in high signal complexity and challenging fault diagnosis. Recent studies have leveraged the synergy between the Transformer's multi-head self-attention mechanism and convolutional networks to enhance feature extraction. However, these approaches often introduce excessive model complexity, leading to high computational costs and limiting their deployment in real-world industrial scenarios. To address these challenges, this paper proposes a lightweight multi-scale bidirectional self-attentive diagnosis method (MBSADM). First, a multi-scale attention mechanism is designed to effectively capture discriminative features across different scales of vibration signals. Second, a multi-scale feature extraction module integrates multi-scale dilated convolution blocks with the multi-scale attention mechanism, enabling a multi-local receptive field with reduced computational overhead and fewer model parameters. Finally, to fully exploit temporal dependencies, we introduce a bidirectional Transformer that leverages a reverse mechanism to construct sequence representations containing spatially inverted information, thereby enhancing the temporal modeling capability of extracted features. Extensive experiments under strong noise, different load, and fluctuating speed conditions demonstrate the robustness and superior classification performance of the proposed MBSADM. Compared to five state-of-the-art fault diagnosis methods, MBSADM achieves higher diagnostic accuracy and demonstrates stronger industrial applicability, making it a promising solution for real-world bearing fault detection.</p>