Ensuring high reliability with low operational costs is an important operational requirement for future manufacturing systems (e.g., smart factory, intelligent production lines). Traditional reliability approaches, however, cannot be directly applied, as they require large amount of historical failure data, which are often not available for future manufacturing systems. To fill this gap, this project aims at developing new approaches to improve the reliability and reduce operational costs of future manufacturing systems through online reliability assessment and predictive maintenance planning based on digital twins. First, a digital twin-based model (called digital failure twin) will be developed to simulate the failure behaviour of future manufacturing systems. By combining digital twins with failure models, the digital failure twin could allow evaluating reliability without the need of historical failure data. Then, a Bayesian framework will be developed for online updating of the reliability based on the condition-monitoring data from sensors. Through the mechanism of online updating, the developed method will provide more accurate reliability assessments. Thirdly, we will propose a transfer learning-based approach for remaining useful life prediction and predictive maintenance planning based on the digital failure twin model. The developed approach could significantly reduce the amount of required training data, as a model could be pre-trained based on simulated training data from the digital failure twin, and then fine-tuned based on the online-collected condition monitoring data through transfer learning. Together with our industrial partners (GE Healthcare), we will design real-world use cases of smart manufacturing systems to test the developed approaches. The performances of the developed methods will be validated by comparing them to state-of-the-art methods from literature on 1) the accuracy of reliability assessment and 2) the long-term operational costs