A novel dynamic spatio-temporal graph based condition monitoring framework for consistency retention of digital twin

A responsive consistency retention strategy is crucial for the engineering application of digital twin (DT). The condition monitoring technique based on graph theory can provide an overall reliability assessment and thus guide DT model updating. However, most existing studies constructed graph topology merely based on data information without incorporating prior engineering knowledge, which restricts the performance of such approaches. To tackle this limitation, a novel graph construction paradigm based on the mechanism of performance degradation and fault propagation is developed in this study. On this basis, unsupervised learning is further combined to form a dynamic spatio-temporal graph based condition monitoring framework for DT consistency retention. Specifically, the spatial dependencies of multi-sensors are quantified based on the evolution of the fault-related frequency band, and then multidomain features are assigned to each graph node. After that, the spatio-temporal graph set is fed to a dual-decoder graph autoencoder to extract the essential features of normal conditions, where a domain adaptation module is introduced to eliminate environmental effects. Hypothesis testing is conducted at last to inspect the machine state over time and make the final decision. Validation and comprehensive comparison experiments were carried out on two engineering scenarios with different scales (component and system level). The Numenta Anomaly Benchmark (NAB) was employed to evaluate the effectiveness of the proposed approach and the results revealed the great potential of the proposed framework for DT consistency retention.