This paper presents a lightweight fault diagnosis framework for bearing defects, integrating time-frequency analysis, deep learning, and model compression techniques to address challenges in resource-constrained environments. The proposed method combines the S-transform for high-resolution time-frequency representation with MobileNet as an efficient backbone network, enabling robust feature extraction from complex vibration signals. To enhance deployment on edge devices, knowledge distillation is employed to compress the model, significantly reducing computational complexity while maintaining diagnostic accuracy. Additionally, domain adaptation is considered to mitigate domain shift issues, ensuring robust performance across varying operating conditions. Experimental results demonstrate the framework’s effectiveness, achieving high diagnostic accuracy with reduced computational overhead, making it a practical solution for real-time industrial applications. The proposed approach bridges the gap between advanced deep learning techniques and practical industrial requirements, offering a scalable and efficient solution for bearing fault diagnosis.