Exploring the Boundaries of Resource-Constrained AIoT: Tiny Machine Learning for Condition Monitoring using 8-bit and 32-bit Microcontrollers

Tiny machine learning (TinyML) emerges as an innovation that gives more autonomy to edge devices, which are typically low-powered and low-memory. With most developments focused on 32-bit devices, this work queries the performance and effectiveness of 8-bit devices to provide equal, if not more, benefit in terms of cost and power consumption, in exploiting all available resources in a small device. The research interest was deploying a multilayer perceptron model for multivariate time series prediction in 8-bit and 32-bit devices with evaluation of the limitations of neural network modelling in 8-bit microcontrollers. A gas-based early fire detection TinyML device was designed as a use case, to monitor the condition of an indoor environment. With only 32 kilobytes flash memory, the findings suggest that artificial intelligence (AI) tasks generally requiring a larger power source and network endpoints bridge could now be realized on-device at a fraction of computing resource utilization. Given the challenges of sustainability in electronics, this study hopes to contribute to reconcile reducing energy-hungry AI models that would otherwise be difficult to run. A proving ground is the smallest off-the-shelf microcontroller available capable of TinyML applications, and that crucial consideration of each factor in the implementation can help optimize costs and materials.