A research of convolutional neural network model deployment in low- to medium-performance microcontrollers

Artificial intelligence internet of things (AIoT) is a technology that came into being under the development of artificial intelligence (AI) and Internet of things (IOT) where deep learning is vigorously promoted and used. Compared with the traditional concept of the Internet of things, the main difference of AIoT technology is that it applies interconnected devices which are embedded with the capacity of neural network model reasoning to the perception layer, this reduce reliance on edge servers (especially for neural network model training or reasoning). Thus, the edge devices of the system will get a more intelligent execution power. For the IOT system structures that have been built at present, most of the interconnection devices in the sensing layer, such as data acquisition nodes or execution nodes, are designed with the low and medium performance microcontroller unit as the processing core. After using the technology such like lightweight neural network and global average pooling, we succeed in deploying the convolutional neural network model to the low and medium performance microcontroller. Thus, the original node can get the reasoning result of neural network model in offline state and use it as a decision element for the operation of the system whit a simple modification of the program.