Virtual-Real Spatial-Temporal Dual Layer Transformer for virtual sensor state perception

In practical application scenarios such as air quality, traffic and mechanical processing, sensors are often constrained by spatial capacity, geometric structures, extreme environments and other factors, making it impossible to place them in critical monitoring areas. To address this issue, a novel virtual sensor state perception generalization framework, the Virtual-Real Spatial-Temporal Dual Layer Transformer (VR-STDT) model is proposed. It constructs a spatial-temporal correlation model between real sensors and unobservable virtual sensors, to solve the problem of missing information in sensor-restricted zones. Considering the “stop-start” single-operation system with a short time window and high sampling frequency, a historical similar attention mechanism and a convolution-based time patching mechanism are proposed to effectively solve the contradiction between low resolution and information loss. Finally, verification was carried out in practical application scenarios, such as the kitchen particle concentration diffusion experiment platform and the machine tool spindle temperature experiment platform, and then the open urban air quality data set was used for auxiliary verification. The results show that the proposed model achieved an average performance improvement of 10.20 % over existing inter-node spatial-temporal prediction models.