DDNN based data allocation method for IIoT

Abstract

Abstract With the complete application of artificial intelligence in the field of industrial production and manufacturing and the rapid development of edge computing, industrial processing sites often need to deploy machine learning tasks at edges and terminals. We propose a data allocation method based on Distributed Deep Neural Networks (DDNN) framework, which allocates data to edge servers or stays locally for processing. DDNN divides deep learning tasks and deploys pre-trained shallow neural networks and deep neural networks at local or edges, respectively. However, all data is processed locally, and the failure is sent to the edge server or the cloud. It will lead to excessive pressure on local terminal equipment and long-term idle edge servers, which cannot meet industrial production’s real-time requirements on user privacy and time-sensitive tasks. In this paper, the complexity and inference error rate of machine learning model, the data processing speed of local equipment and edge server, and the transmission time are comprehensively considered to establish the system model. A joint optimization problem is proposed to minimize the total data processing delay. The optimal solution is derived analytically, and the optimal data allocation methhod is given. Simulation experiments are designed to verify the method’s effectiveness and study the influence of key parameters on the allocation method.