Scholarlink

The rapid growth of IoT and sensor-enabled devices presents challenges in data processing, including high transport costs and resource constraints. Efficient on-device processing is essential to reduce cloud reliance and improve near real-time analysis. Language Models (LMs) offer a new intriguing approach to sensor data analysis, enabling interacting with the data and contextual reasoning. To explore this potential, we propose EdgeSense, a framework for experimenting with LMs on edge devices and mobile phones. As an example case study, we developed the EdgeSense/HealthSense mobile app, which classifies user activities using real-time sensor data and LMs. Our evaluation shows that while EdgeSense achieves over 80% accuracy, response generation remains resource-intensive, utilizing two CPU cores and over 40% of the memory. These results highlight both the potential and limitations of LMs for near real-time sensor data analysis on edge devices, emphasizing the need for further optimization.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating the Practicality of Language Models on Edge Devices for Sensor Data Analysis: A Sensible Approach?

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dinesh Kumar Karthikeyan,
  • Roopesh Kumar Shanmugasundaram,
  • Anna-Sofia Paavonen,
  • Niko Mäkitalo

摘要

The rapid growth of IoT and sensor-enabled devices presents challenges in data processing, including high transport costs and resource constraints. Efficient on-device processing is essential to reduce cloud reliance and improve near real-time analysis. Language Models (LMs) offer a new intriguing approach to sensor data analysis, enabling interacting with the data and contextual reasoning. To explore this potential, we propose EdgeSense, a framework for experimenting with LMs on edge devices and mobile phones. As an example case study, we developed the EdgeSense/HealthSense mobile app, which classifies user activities using real-time sensor data and LMs. Our evaluation shows that while EdgeSense achieves over 80% accuracy, response generation remains resource-intensive, utilizing two CPU cores and over 40% of the memory. These results highlight both the potential and limitations of LMs for near real-time sensor data analysis on edge devices, emphasizing the need for further optimization.