A digital twin-based fault diagnostic method for subsea control systems.

• A new multi-physics modeling architecture was proposed for digital twin (DT) physics modeling. • A novel DT-based diagnosis method using DT-generated fault data to train the deep learning diagnostic model was proposed. • An online model modification technique was introduced to mine labeled training samples from measured SCS monitoring signals. A digital twin (DT) based framework is proposed for data-driven fault diagnosis in a subsea control system (SCS). A novel modeling technique, the physics informed temporal convolution network (PITCN), is first developed by combining a traditional physics-based simulation with collected sensor signals (e.g., pressure and flowrate). The DT is then used to generate simulated signals under different operation and fault conditions, for the purpose of training the convolutional neural network (CNN) based data-driven fault diagnostic model. In addition, an online model modification technique is proposed to label the SCS real-time data used for continuously training the PITCN and CNN during the SCS production period. Experimental results showed the proposed diagnostic framework is superior to traditional CNN based diagnostic methods, as measured by diagnostic accuracy, particularly when labeled sample volumes are limited. The proposed online model modification improved diagnostic accuracy from 91.87% to 97.5% using real-time collected data. [ABSTRACT FROM AUTHOR]