利用自然語言處理技術學習過去施工排程以達成自動施工規劃

Abstract

幾十年來，建築、工程和施工(AEC) 行業一直遭受延誤和成本超支的困擾。為了按時並在預算範圍內完成項目，AEC 行業在很大程度上依賴於施工項目進度。然而，這些進度表仍然是手動創建和更新的，這會花費大量時間，導致錯誤，並導致規劃和調度不善，這是當今項目延誤的主要原因之一。因此，克服這些挑戰需要一種自動化的施工進度管理方法，該方法從現有和以前的數據庫中提取資訊和知識，並學習施工過程以改進施工規劃和進度安排。本研究旨在開發一種基於學習的施工進度模板，自動從歷史專案規劃和進度記錄中學習施工知識以生成進度模板。在最近的研究中，自然語言處理(NLP) 技術可以通過提供從專案文檔、模型和過去的專案知識中提取和解釋數據的能力，徹底改變施工規劃和調度。具體來說，我們提出了基於構建活動的向量和實體表示的基於自動學習的進度開發，其中使用三種不同的語言模型（基於BERT 的uncased、DistilBERT 和RoBERTa）對排序和邏輯依賴知識進行建模。我們從各個建築項目獲得的16 個施工進度表的多樣化數據集對本研究所提出的基於學習的進度建構方法進行了詳盡的測試和驗證。結果表明，我們的方法可以有效地組織施工進度活動並準確預測活動的順序及其在不同工作包中的邏輯聯繫。它支持規劃人員建立利用歷史數據的知識管理系統，從而改進未來專案中重複情境的決策。

The Architecture, Engineering, and Construction (AEC) industry has been suffering from delays and cost overruns for decades. To complete projects on time and under budget the AEC industry heavily relies on construction project planning and scheduling. However, these schedules are still created and updated manually, which takes a lot of time, causes mistakes, and leads to poor planning and scheduling, one of the main reasons for project delays today. Consequently, overcoming these challenges requires an automated construction plan schedule management method that extracts information and knowledge from existing and previous databases as well as learns the construction process to improve construction planning and scheduling. This study developed a learning-based construction planning which, automatically learns the construction knowledge from historical project planning and scheduling records. In recent studies, Natural Language Processing (NLP) technologies that can revolutionise construction planning and scheduling by providing the ability to extract and interpret data from project documents, models, and past project knowledge-based. Specifically, we present Automatic Learning-Based Construction Planning based on a vector and entity representation for construction activities where the sequencing and logic dependency knowledge is modeled with three different language model (BERT-based uncased, DistilBERT and RoBERTa). Our learning-based language models are exhaustively tested and validated on a diverse dataset of 16 construction schedules obtained from various building projects. The results indicated that our method is effective in formalizing construction planning activities and accurately predicting the sequences of activities and their logic dependency in different work items. It supports the planners to establish knowledge management systems that leverage historical data, leading to improved decision making for repetitive scenarios in future projects.