A CROSS-TEMPORAL METHOD FOR FUSING HETEROGENEOUS STREAMING DATA VIA A MODIFIED ATTENTION MECHANISM WITH REAL TIME-GAP MODELING

Abstract

Background. Monitoring systems generate heterogeneous data that arrive at different times. Some methods assume that regular timing or merge signals only at the end can mix up past signals with current events and miss short bursts. The authors of the work conducted a study to make the fusion process time-aware and reduce these errors. Materials and methods. This study introduces a mechanism that accounts for time-gap (Δt), weighting recent events more. The model operates backward in time to avoid using future data. It’s designed so the attention score decreases as the time gap grows, and minor timestamp variations cause limited score changes. The method was tested on a simulated stream combining sensor time series and text notes. The delay between a spike and a keyword is varied. Baselines are the same architecture without the (Δt) term. Results. The proposed method offers quicker warnings and reduces false alarms, while keeping similar discrimination levels compared to baseline methods. Examples are presented that highlight which moments and phrases contribute to each decision. Conclusions. The introduction of a time-gap aware attention together with causal masking improves fusion across heterogeneous streaming data and supports earlier and more transparent decisions in monitoring tasks.