Energy Efficient Predictive Maintenance in Resource Constraint Industrial IoT devices using Smoothed Mish Activation Function

Abstract

Traditionally, Industrial IoT devices collect sensor data to a cloud platform where ML/DL processing is to be done. Edge computing has the advantage of reducing latency, improved battery performance, safe transmission and reducing vulnerability. These benefits are particularly significant considering the limited resources on IoT devices, such as only a few kilobytes of RAM, and the critical importance of energy savings in industrial applications. Arduino Uno uses 2kB of RAM and 32 kB of read only flash memory. Optimal performance is required for ML inference on EDGE devices to get good accuracies and low latency. This research paper is based on memory optimization in resource constrained devices, focusing on a novel approach that combines the Bonsai tree architecture with a smoothed Mish activation function. While the Bonsai tree architecture has been explored in prior works, this research contributes to the refinement of the activation function, a crucial component in achieving superior test accuracy within limited memory constraints. Although many algorithms have been developed with good accuracy but can’t be used in optimal power consumption with low latency and resource constraint like fewer RAM and ROM. A smooth Mish activation function in Bonsai tree algorithm is proposed to enhance accuracy with lowest ML model size, where Bonsai tree is a single,shallow,sparse tree where data is projected in low dimension space. Lowest ML model size consumes least power. We observed the smoothed Mish activation function in the Bonsai tree algorithm outperform with low latency and power consumption. This paper aims to elucidate the methodology, experimental results, and implications of the modified activation function, shedding light on its potential to revolutionize memory-efficient edge computing in the Industrial IoT network.