Partial Distributed Deep Learning Inference Model for Image based Edge Device Cluster

In recent years, the development of AI technology has driven the development of network edge applications such as smart manufacturing, smart factories, and smart cities. Deep neural networks are increasingly being deployed in smart manufacturing edge application scenarios to carry out edge detection tasks. However, there is a gap between the expensive DNN computing network and the limited computing power of the edge devices. To deploy compute-intensive neural networks on resource-constrained edge devices, traditional approaches rely either on offloading workloads to a remote central cloud or on optimizing compute on local edge devices. However, the method of uploading tasks to the central cloud is significantly affected by the unreliability and delay of the WAN, and the sensitive information of users is easy to be leaked during transmission, and the local computing method is also limited by the computing power. The latest machine learning models have great demands on memory, computing, and energy, which poses challenges to distributed computing and data aggregation for heterogeneous and resource-constrained edge devices. In order to address these challenges, this paper proposes solutions for different image types in intelligent manufacturing scenarios. When images are processed in the production process of intelligent manufacturing, this paper adopts the method of horizontal partitioning of the DNN model to perform DNN-based inference tasks on the edge device cluster. In order to reduce the consumption of network bandwidth, a dynamic priority scheduling algorithm is used to optimize the distribution of inference tasks at the edge nodes to improve the inference speed. Finally, experiments show that the edge node optimization scheduling method proposed in this paper has a good effect on the performance of running time, communication cost and energy consumption of edge devices.