隨著工業 4.0 和邊緣運算技術的快速發展，智慧醫療與工業監控等場景對 即時資料處理和智慧化分析提出了高效能、低功耗與資源優化的嚴苛需求。然而，傳統卷積神經網路 (CNN) 因其高運算量與記憶體需求，難以滿足資源受限邊緣裝置的應用場景。為解決此一挑戰，本研究提出了一種結合 ReActNet 架構與 XGBoost 分類器的硬體加速器，專注於一維時間序列資料的處理。 本設計以 1D CNN 的輕量化特性為基礎，通過二值卷積技術顯著降低運算 負擔，嘗試取得低功耗與高準確度之間的平衡，並且為智慧醫療中的生理 訊號監測以及工業監控中的設備異常檢測提供了創新的解決方案，展現出應用於資源受限邊緣裝置的潛力。;With the rapid advancement of Industry 4.0 and edge computing technologies, applications in smart healthcare and industrial monitoring demand highly efficient, low-power, and resource-optimized solutions for real-time data processing and intelligent analysis. However, traditional convolutional neural networks (CNNs), with their significant computational and memory requirements, struggle to meet the constraints of resource-limited edge devices. Addressing these challenges, this study introduces a hardware accelerator combining the ReActNet architecture with an XGBoost classifier, specifically designed for processing one-dimensional time-series data. Built upon the lightweight characteristics of 1D CNNs, the proposed design leverages binary convolution techniques to significantly reduce computational overhead, aiming to achieve an optimal balance between low power consumption and high accuracy. This innovative approach offers promising solutions for real-time physiological signal monitoring in smart healthcare and anomaly detection in industrial monitoring, demonstrating substantial potential for deployment in resource-constrained edge devices.
