Due to limited computation and storage resources of industrial internet of things (IoT) edge devices, many emerging intelligent industrial IoT applications based on deep neural networks (DNNs) heavily depend on cloud computing for computation and storage. However, cloud computing faces technical issues in long latency, poor reliability, and weak privacy, resulting in the need for on-device computation and storage. On-device computation is essential for many time-critical industrial IoT applications, which require real-time data processing. In this paper, we review three major research areas for on-device computation, specifically quantization, pruning, and network architecture design. The three techniques could enable a DNN model to be deployed on edge devices for real-time computation and storage, mainly due to the reduction of computation and space complexity. More importantly, these techniques could make DNNs applicable to industrial IoT devices.

由于工业物联网（IoT）边缘设备的计算和存储资源有限，许多基于深度神经网络（DNN）的新兴智能工业IoT应用严重依赖于云计算进行计算和存储。但是，云计算面临着较长的延迟，较差的可靠性和较弱的隐私等技术问题，导致需要在设备上进行计算和存储。设备上计算对于许多对时间要求严格的工业IoT应用至关重要，这些应用需要实时数据处理。在本文中，我们回顾了设备上计算的三个主要研究领域，特别是量化，修剪和网络体系结构设计。这三种技术可以使DNN模型部署在边缘设备上，以进行实时计算和存储，主要是由于减少了计算和空间复杂性。更重要的是，这些技术可以使DNN适用于工业物联网设备。