以營建業勞工儲備心率為基礎之機器學習分類模式與個人化安全管理之研究

Abstract

營建業的施工環境變化快速、多樣性高，勞工常因需要暴露在烈陽高溫、侷限空間或臨空作業等高風險的作業環境，而承受較高的職災風險。工程期間發生的職業災害不僅會造成排程延宕，衍生出額外的工程成本，給罹難家庭帶來無可抹去的傷痛，而且還要面臨被社會檢討的負面名聲成本，增加剩餘員工之心理壓力。因此營造建設產業的工程災害是各界應注意並改善的問題。

本研究以潛盾隧道環境為例，將過去潛盾施工隧道收集的勞工原始工作心率和基本身體條件問卷數據做為資料集建置疲勞預測模型，結合機器學習發展出儲備心率百分比預測系統，擬定兩項研究目標：施工前的安全評估和個人化的安全管理，希望做為往後相關勞工安全管理研究與應用之參考。在施工前綜合安全評估中，比較5種監督式機器學習演算法，在邏輯迴歸、支撐向量機、K鄰近、決策樹和隨機森林中，以決策樹預測模型在按照各%HRR 分類區間數量比例分割 80%訓練集和20%測試集下的準確率、精確率、召回率和F1值的表現最好，分別為 0.7879、0.7740、0.7879、0.7801，接近0.8，具有一定預測可靠度。而在個人化安全管理中，採用 3 項未知數的S曲線方程式(Sigmoid function)可以得到高解釋性和低誤差的當日個人工作累積工時百分比和對應%HRR的擬合曲線，得到的%HRR反曲點最大值即可作為該名工人的疲勞警示值，當日後超過警示值，可適度關心該位勞工的工作負荷情形，調整工班安排減少疲勞發生。研究結果顯示以機器學習作為施工前的安全評估和個人化安全區間的建置方法具有一定程度的可信力和解釋力，可作為日後相關研究的發想與延伸應用參考。

The construction environment in the construction industry changes rapidly and has high diversity. Workers are often exposed to high-risk working environments such as hot sun, confined space, or airport work, and thus suffer from high risk of occupational accidents. Occupational accidents that occur during the construction period will not only cause delays in scheduling, generate additional construction costs, and bring indelible pain to the families of the victims, but also face negative reputation costs from social censorship, and increase the psychological stress of the remaining employees. Therefore, the engineering disasters caused by the construction industry are critical issues that all sectors should pay attention to and improve.

In this study, taking the shield tunnel environment as an example, the original working heart rate and basic physical condition questionnaire data of laborers collected in the past shield construction tunnels were used as data sets to build a fatigue prediction model, and a reserve heart rate percentage prediction system was developed in combination with machine learning. Two research objectives: pre-work safety assessment and personalized safety management, hope to serve as a reference for future research and application of labor safety management. In the pre-work comprehensive safety assessment, comparing 5 kinds of supervised machine learning algorithms: logistic regression, support vector machines(SVM), k-nearest neighbors (KNN), decision tree and random forest. using the decision tree prediction model to predict the proportion of the number of classification intervals according to each %HRR.The accuracy rate, precision rate, recall rate and F1-score under the 80% training set and 20% test set performed best, respectively 0.7879, 0.7740, 0.7879, 0.7801, the result closed to 0.8, with a certain degree of predictive reliability. In personal safety management, the Sigmoid function with three unknowns can be used to obtain a highly interpretable and low-error personal cumulative working hours percentage and %HRR fitting curve of the day, and the obtained %HRR inflection point maximum value is enough As the worker's fatigue warning value, if it exceeds the warning value after the day, you can properly care about the worker's workload and adjust the work shift arrangement to reduce the occurrence of fatigue. The research results show that using machine learning as a method of pre-work safety assessment and construction of personalized safety intervals has a certain degree of credibility and explanatory power, and can be used as a reference for future related research ideas and extended applications.