Sharing the Edge: Multiple Concurrent Neural Workloads on a Single Microcontroller in a Smart HVAC Vent

Open Access
Article
Conference Proceedings
Authors: Yousef Abdou, Hazem Elshorbagy, Alaa Abdou
Abstract

Edge devices increasingly run multiple inference workloads at once, but a microcontroller offers no free lunch for multitasking: every model competes for the same memory, cycles, and physical environment, including the space the user speaks into. This paper examines that competition through the second prototype of ZenSys, a smart-vent system sharing one ESP32-S3 across a wake-word detector, a keyword spotter, and continuous thermal-occupancy inference. On 2,000 labeled thermal frames, an adaptive background-subtraction blob detector outperformed a compact CNN (F1 0.92 vs. 0.80) at a fraction of the memory and latency, a result we attribute to a single-room corpus rather than to any inherent weakness of learned inference. Small early voice trials point to a similar inversion: shorter commands held up better under noise, wake-word detection appeared weaker for female speakers, and actuator vibration shrank usable voice range to 3–10 cm whenever the fan ran. For the occupant, the binding constraint was mechanical rather than computational.

Keywords: Edge AI, Embedded Machine Learning, Keyword Spotting, Thermal Occupancy Sensing, Voice User Interface, Human–Machine Interaction, Internet of Things

DOI: 10.54941/ahfe1008389

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← Intelligent Design Architectures for Human-AI Teaming in Safety-Critical OperationsMetacognitive Scaffolding for Novice Designers in Generative AI-Assisted Poster Design: A Formative Study of Premature Convergence →
View all articles in Human Factors in Design, Engineering, and Computing →