资源受限的机械振动WSN层次分解CNN边缘计算方法

Hierarchical decomposition of CNN for resource-constrained mechanical vibration WSN edge computing

用于机械振动监测的无线传感器网络节点的微控制器需要进行复杂的边缘计算,然而硬件资源受到限制。卷积神经网络作为一种性能优越的深度学习算法,若将其运行在MCU上可增强边缘WSN节点的计算能力。本文提出了一种不修改CNN模型的层次分解方法,解决了难以在资源受限的MCU上运行不轻量化CNN的问题,实现了机械振动WSN节点的计算能力增强。首先通过设计文件结构用于分解并存储CNN模型参数,然后提出内存管理方法并推导随机存取存储器的消耗过程,最后提出参数定位方法准确高效地读取模型参数。实验表明仅使用1.76 KB RAM与2.14 KB Flash,在3.15 ms内便可实现高准确率的边缘计算识别任务。

The microcontroller of wireless sensor network(WSN)nodes used for mechanical vibration monitoring requires intricate edge computing,yet face limitations in hardware resources.Convolutional neural network(CNN),as a high-performance and commonly used deep learning algorithm,can enhance the computational capabilities of edge WSN nodes when run on microcontroller units(MCUs).This paper proposes a hierarchical decomposition method for CNN models without modification,addressing the challenge of running non-lightweight CNN on resource-constrained MCU and enhancing the computational capabilities of mechanical vibration WSN nodes.First,a file structure is designed to decompose and store CNN model parameters.Subsequently,a memory management method is proposed,and the consumption process of random-access memory is derived.Finally,a parameter localization method is introduced to accurately and efficiently retrieve model parameters.Experiments demonstrated that with only 1.76 KB of RAM and 2.14 KB of Flash,high-precision edge computing recognition tasks can be accomplished within 3.15 ms.