Among several fundamental concepts within Industry 4.0, the concept of digital twin plays a central role because it bridges the assets in the physical world with the models and analytics at the virtual space to provide companies with complete digital footprint of their equipment for the entire life cycle, leading to deep understanding of equipment health conditions and optimal maintenance schedules. The majority of currently available digital twins are based on physics based models. Development of digital twins upon physics based models requires in-depth knowledge about every details within physical systems, which may not be available in industrial practice. Despite the fact that physics based models offer good interpretability, the difficulty in model identification and adaptation to changing conditions precludes the wide application of physics based models in digital twins. With the recent advances in deep learning and data acquisition techniques, it is natural to consider data-driven models to develop digital twins.
An industrial furnace is chosen in study as a pilot example to demonstrate the effectiveness of data-driven digital twins. A novel probabilistic sequence-to-sequence learning model is developed to capture the complex dynamics within the furnace. The proposed method outperforms a range of alternative approaches in simulating the industrial furnace as a digital twin. In addition, the stochastic nature of a sophisticated environment can be taken into account for optimal decision making and management through the probabilistic outputs from the digital twin. Moreover, the data-driven digital twin can be adapted to new or unseen conditions using transfer learning techniques. This presentation will discuss development of digital twins using deep learning models and potential application of such data-driven digital twins in the context of reliability and asset management.