A Scalable Multi-Task Model for Virtual Sensors

Virtual sensors replace expensive physical sensors by predicting target signals from available measurements. Existing approaches require application-specific models with hand-selected inputs, cannot leverage task synergies, and lack consistent benchmarks. Time series foundation models are computationally expensive and limited to predicting their input signals, making them incompatible with virtual sensors. We introduce the first multi-task model for virtual sensors. It simultaneously predicts diverse virtual sensors, exploits synergies, and learns relevant input signals for each target, eliminating expert selection while adding explainability. Across three standard benchmarks and an application-specific dataset with over 18 billion samples, our architecture reduces computation time by up to 415x and memory requirements by 951x while maintaining or improving predictive quality over unified baselines. Compared with isolated single-sensor models, it generates superior predictions at similar inference speed and scales to hundreds of virtual sensors with nearly constant parameter count.