工程監造人員疲勞即時監測與預警系統之研究

Abstract

營建業職業災害死亡率長年居高不下，其中「疲勞」被視為關鍵影響因子之一，尤以監造人員的疲勞風險長期未獲充分重視，且目前較少可實際應用之監測與預警系統架構。為此，本研究建構一套即時疲勞監測與預警系統，透過心率手環蒐集監造人員之心率資料，並以環境感測器結合 Raspberry Pi Pico W 量測工地之溫濕度與噪音。感測數據透過 LoRa 與 4G LTE 技術，結合 LoRaWAN、Socket、MQTT 等多種通訊協議進行即時傳輸，經本地端處理後寫入資料庫，並以資料庫作為資料串流中樞節點，進一步導入深度學習模型進行即時疲勞預測，並將預測結果回寫資料庫，以支援後續視覺化顯示與警示系統運作。

模型訓練方面，本研究針對40位來自不同工程案場之監造相關人員進行生理監測，並採用多種資料增強策略建立九種訓練資料集，藉此比較不同資料組合對模型效能之影響。最終疲勞分類任務準確率達約 77%，回歸任務之 R-square 約為 0.777，展現良好的預測準確性與泛化能力。系統亦結合 Apache Superset 建立即時資料視覺化平台，即時呈現心率變化、模型預測結果與環境參數。同時，整合 SMTP Server 與資料庫建立自動發信之預警機制，當疲勞等級預測超過臨界值時，能即時通知相關人員採取應對措施，降低因疲勞未即時處理所導致之職業災害風險。

本研究最終於工務所與工地現場進行兩次實地測試，分別驗證系統於室內與室外環境中之傳輸穩定性、模型預測效能、視覺化更新與警示機制運作情形。結果顯示，室內環境下系統運作穩定，但環境參數對疲勞監測的影響相對有限；而工地現場則雖易受 4G LTE 訊號影響，產生資料遺漏與延遲情形，但環境變化明顯，具備長期監測價值。

綜上所述，本研究成功整合即時通訊、深度學習、資料庫管理、視覺化分析與自動化預警技術，建置具即時性與實務應用潛力之營建現場疲勞監測系統，未來可進一步克服現有限制，擴展至多元工地場域，強化營建業之職業安全管理效能。

The construction industry has long suffered from a high rate of occupational fatalities, with fatigue identified as one of the key contributing factors. In particular, the fatigue risks faced by supervisory personnel have been historically under-recognized, and there remains a lack of practical systems for monitoring and early warning. To address this gap, this study develops a real-time fatigue monitoring and alert system. Heart rate data from supervisory personnel is collected using wearable wristbands, while environmental parameters—such as temperature, humidity, and noise levels—are measured through sensors connected to a Raspberry Pi Pico W. The collected data is transmitted in real time via LoRa and 4G LTE technologies, integrated with multiple communication protocols including LoRaWAN, Socket, and MQTT. After local processing, the data is written into a central MySQL database, which serves as the core of the data stream pipeline. A deep learning model is then applied to perform real-time fatigue prediction, with results written back to the database to support further visualization and alert functions.

For model training, physiological data was collected from 40 supervisory personnel across various construction sites. Multiple data augmentation strategies were employed to create nine different training datasets, allowing for performance comparison across different configurations. The final classification model achieved an accuracy of approximately 77\%, while the regression task attained an R-squared value of 0.777—indicating strong predictive accuracy and generalizability. The system also incorporates Apache Superset to establish a real-time visualization platform, which displays heart rate trends, predicted fatigue levels, and environmental conditions. Additionally, an automated email alert system was developed using an SMTP server integrated with the database. When predicted fatigue levels exceed a predefined threshold, the system sends real-time notifications to relevant personnel, enabling timely interventions and reducing the likelihood of fatigue-related incidents.

Two field tests were conducted to evaluate system performance: one indoors at a site office and one outdoors at a bridge construction site. These tests verified the system’s transmission stability, model performance, visualization responsiveness, and alert functionality. Results showed stable operation in the indoor environment, although environmental parameters had limited impact on fatigue detection. In contrast, the outdoor setting exhibited greater variability in environmental conditions but occasionally experienced data loss or delay due to unstable 4G LTE signals—highlighting areas for future improvement.

In conclusion, this study successfully integrates real-time communication, deep learning, database management, data visualization, and automated alerting to establish a practical and responsive fatigue monitoring system for construction sites. The system demonstrates strong potential for deployment in diverse field environments and can significantly enhance occupational safety management in the construction industry.