Self-Evolving Binarized Neural Framework for Low-Resource Data Streams in Edge Artificial Intelligence of Things
V. Soma Sundari, Baker Karim, Aruna. M, R. Kiruthika, N. Naga Saranya · 2025
In recent years, the deployment of edge Artificial Intelligence of Things (AIoT) systems for real-time bio-signal analysis for continuous Electrocardiogram (ECG) monitoring has demanded ultra-low-power models capable of real-time decision-making. However, the existing simple binarized Convolutional Neural Networks (bCNNs) model remains fundamentally static and incapable of learning from evolving data streams as well as adapting to sensor drift. To address these limitations, this study proposes a novel Self-Evolving Binarized Neural Framework (SELFBIN-AI) designed for streaming, lowresource AI environments. Initially, SELFBIN-AI integrates a quantized multilayer perceptron ($q$MLP) -based encoder for efficient binary image generation from raw sensor input and a modular bCNN architecture for low-power classification. Subsequently, a continual learning engine with dual-loss aggregation is employed to maintain performance across data shifts. Then, the model incorporates a task-aware routing mechanism that selects specialized binary branches based on the signal context. A lightweight on-device scheduler monitors concept drift and triggers adaptation, whereas an optional edgecloud co-learning protocol allows selective knowledge distillation to enhance generalization without compromising the data privacy. Experimental results indicate that the proposed SELFBIN-AI model outperformed in terms of accuracy ($\mathbf{9 8. 9 \%}$), along with dynamic power dissipation of$\mathbf{2 7. 4} \boldsymbol{\mu} \mathbf{W}$when compared to existing bCNN model.