The precise determination of a bearing’s Remaining Useful Life (RUL) is paramount for upholding equipment operational safety and for the effective execution of predictive maintenance programs. To counteract the deficiencies of established deep learning methodologies, especially their limitations in extracting bearing wear-and-tear features at diverse scales and their less-than-optimal attention frameworks for temporal data, this study introduces an innovative hybrid forecasting architecture: the Multi-scale Time-series Convolutional Transformer (MTCT). The MTCT integrates multi-scale feature fusion with convolutional attention mechanisms. The model employs a temporal multi-scale segmentation strategy to extract time-domain, frequency-domain, and time-frequency domain features, while incorporating an improved Time-Series Convolutional Block Attention Module (TS-CBAM) to dynamically weight critical degradation features. Additionally, a Transformer encoder is utilized to capture long-range dependencies, and a bias correction mechanism is introduced in the loss function to enhance prediction stability. Empirical assessment was carried out employing vibration signals from bearings throughout their complete lifecycle, captured under various operational parameters. This involved contrasting the proposed complete architecture with its constituent variants to ascertain the effectiveness of both multi-resolution feature integration and the attention framework. The outcomes reveal that the presented model offers a significant advancement in capturing degradation progression, demonstrating prediction precision and resilience that surpasses existing advanced methodologies.