Scholarlink

<p>A real-time embedded smoke classification system is proposed that integrates a multi-sensor electronic nose with a lightweight 1D-Convolutional Neural Network (1D-CNN). The architecture employs six low-cost metal oxide semiconductor (MOS) gas sensors to capture dynamic temporal profiles of five visually similar but chemically distinct indoor smoke types: joss stick, mosquito coil, benzoin resin, burning paper, and cigarette. Sensor signals are pre-processed through z-score normalization and Savitzky–Golay filtering, then segmented into 30-second windows to form structured inputs for classification. The 1D-CNN model, optimized using TensorFlow Lite for edge deployment, achieved a classification accuracy of 96.4% and an F1-score of 96.3% on a Raspberry Pi 4B, with an end-to-end decision latency (including preprocessing and I/O) below 160&#xa0;ms per 30‑s window; the quantized 1D‑CNN inference itself requires ~ 5–6&#xa0;ms on Raspberry Pi 4B. SHAP (SHapley Additive exPlanations) analysis provided interpretable insights into individual sensor contributions, reinforcing the effectiveness of sensor fusion. Deployment trials validated stability, responsiveness, and feasibility for embedded applications. These findings establish a viable pathway for intelligent, resource-efficient smoke detection in next-generation environmental monitoring and smart building systems. This system enables low-cost, scalable deployment in smart homes and industrial safety applications, addressing a critical gap in fine-grained smoke discrimination.</p>

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

Lightweight edge AI for real-time smoke classification using a multi-sensor E-nose

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • N. Vivekanandan,
  • K. Rajeswari,
  • Sanjay Salve

摘要

A real-time embedded smoke classification system is proposed that integrates a multi-sensor electronic nose with a lightweight 1D-Convolutional Neural Network (1D-CNN). The architecture employs six low-cost metal oxide semiconductor (MOS) gas sensors to capture dynamic temporal profiles of five visually similar but chemically distinct indoor smoke types: joss stick, mosquito coil, benzoin resin, burning paper, and cigarette. Sensor signals are pre-processed through z-score normalization and Savitzky–Golay filtering, then segmented into 30-second windows to form structured inputs for classification. The 1D-CNN model, optimized using TensorFlow Lite for edge deployment, achieved a classification accuracy of 96.4% and an F1-score of 96.3% on a Raspberry Pi 4B, with an end-to-end decision latency (including preprocessing and I/O) below 160 ms per 30‑s window; the quantized 1D‑CNN inference itself requires ~ 5–6 ms on Raspberry Pi 4B. SHAP (SHapley Additive exPlanations) analysis provided interpretable insights into individual sensor contributions, reinforcing the effectiveness of sensor fusion. Deployment trials validated stability, responsiveness, and feasibility for embedded applications. These findings establish a viable pathway for intelligent, resource-efficient smoke detection in next-generation environmental monitoring and smart building systems. This system enables low-cost, scalable deployment in smart homes and industrial safety applications, addressing a critical gap in fine-grained smoke discrimination.