Predictive Offloading in Fog Manufacturing for Computational Pipelines using Multi-task Learning

In smart manufacturing, it is significant to integrate the computation service with the manufacturing process to support real-time process controls and data analytics. A suitable computing architecture to handle the influx of data generated from the manufacturing process is Fog manufacturing. In Fog manufacturing, the Fog-cloud collaborative architecture is enabled through a distributed computing platform to facilitate responsive, scalable, and reliable data analysis in manufacturing networks. However, effective utilization of the Fog-cloud computing service requires optimal offloading strategies due to limited computational and bandwidth resources in Fog manufacturing. Therefore, a predictive offloading method that can properly deploy each computation task based on the predicted run-time metrics (e.g., time-latency) is desired. However, the run-time metrics collected in Fog manufacturing are heterogeneous in nature and cannot be modeled through conventional predictive analysis. This is because the computational flow and the data sources vary among different Fog nodes. To overcome this issue, in this paper, a multi-task learning model based predictive offloading method is proposed to assign the computation tasks based on their predicted run-time metrics in Fog manufacturing. The proposed method is evaluated on a Fog manufacturing testbed. The results show that the predictive offloading method can adequately predict the run-time metrics, and further effectively offload the computation tasks to maximize the run-time performance of the computation service.