請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74715
完整後設資料紀錄
DC 欄位 | 值 | 語言 |
---|---|---|
dc.contributor.advisor | 吳焜裕(Kuen-Yuh Wu) | |
dc.contributor.author | Shao-Zu Huang | en |
dc.contributor.author | 黃紹祖 | zh_TW |
dc.date.accessioned | 2021-06-17T09:06:24Z | - |
dc.date.available | 2022-03-12 | |
dc.date.copyright | 2020-03-12 | |
dc.date.issued | 2020 | |
dc.date.submitted | 2020-01-06 | |
dc.identifier.uri | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74715 | - |
dc.description.abstract | 隨著大數據和人工智慧(AI)技術的進步,暴露評估在作業環境設計中推估潛在風險的應用也隨之發展,以確保在設計階段就考量勞工之職業健康。為因應作業場所在設計及分析其暴露情境時所牽涉的各種參數與變項,整合複雜的資訊組合提供了將作業場所設計流程自動化的動力。自動化作業場所設計可被視為自動規劃(Automated planning, AI Planning)的問題,根據先定義之設計目標,依照暴露濃度預測結果回推最適合的設計參數。但實施此系統在現今仍面臨許多挑戰,如 1) 採用黑箱模型進行高風險決策時,可能導致不公平偏見的擔憂; 2)收集適用於學習演算法模型的高品質暴露數據; 以及 3)決策時如何考量互相牴觸之目標以及未來情境的不確定性。因此,本研究目的為探討目前建立AI作業場所設計系統的當前挑戰,並提出縮小其差距的方法。本研究分為三個部分:1)使用Stoffenmanager®和貝氏統計模型進行設計階段風險評估的案例研究,提出現有工具的機會和限制; 2)利用群聚分析法,描述且量化暴露情境間的相似程度,以增加可使用的數據的樣本數,以改善暴露預測; 3)以穩健決策框架,利用學習演算法搜索最佳決策策略的模擬研究。根據研究結果,本研究提出了一個自動化作業場所設計系統的架構,以建議如何結合數據導向的AI技術與傳統的知識導向模型以改善設計決策過程。 | zh_TW |
dc.description.abstract | Advances in big data and artificial intelligence (AI) technology are encouraging improvements in applications that interpret existing occupational exposure data for predicting potential exposures in new workplaces, to ensure the protection of workers early in the design stage. As workplace designs and their resulting exposure scenarios are correlated with a large variety of features, the amount of details and potential permutations of combinations can easily surpass the human capacity for conducting a comprehensive analysis that carefully considers each feature and their implication of the risk outcome. This provides a strong incentive for an automated, if not computer-assisted decision support system for making occupational health by design assessments. An automated workplace design process can be viewed as a problem of AI planning, in which the planning system must synthesize the best design strategy given a description of the workplace scenario and the desired objectives. However, several challenges arise for the implementation of such system, namely 1) the concern of using black-box models for high-stake decisions that may lead to unfair biases; 2) challenges in obtaining high-quality data for model training; and 3) challenges in optimizing decisions under competing objectives and deep uncertainties of future states.
The objective of this study was therefore to identify current challenges of achieving an AI-based automated workplace design system and propose methods for reducing the gap. Specifically, the study was organized into three parts: 1) a case study of design stage risk assessment using Stoffenmanager® and Bayesian modeling to identify opportunities and shortcomings of current tools; 2) a cluster analysis method for characterizing similar exposure scenarios to increase the size of usable data for improving exposure estimates; 3) a simulation study using robust decision-making framework to search for optimal decision strategies using unsupervised learning techniques, namely many-objective evolution and rule inductions algorithms. Based on the results, an automated workplace design system was proposed to recommend how data-driven AI techniques can be applied along with traditional knowledge-driven models to improve the decision process. | en |
dc.description.provenance | Made available in DSpace on 2021-06-17T09:06:24Z (GMT). No. of bitstreams: 1 ntu-109-D04841012-1.pdf: 5734504 bytes, checksum: 673acd90267d2738527024d916502832 (MD5) Previous issue date: 2020 | en |
dc.description.tableofcontents | 致謝 ii
中文摘要 iii Abstract iv Table of Contents 6 List of Figures 9 List of Tables 11 Chapter 1 Introduction 1 1.1 A BRIEF BACKGROUND ON RECENT DEVELOPMENT OF EXPOSURE MODELING IN OCCUPATIONAL HEALTH 2 1.1.1 Control Banding and Source-Receptor Exposure Models 3 1.1.2 Combining measurement data with exposure modeling 4 1.2 ALGORITHMIC DECISION SYSTEMS AND ARTIFICIAL INTELLIGENCE 5 1.2.1 Algorithmic Decision Systems & Artificial Intelligence: a shifted definition of “intelligence” 6 1.2.2 Growing concerns: black box models 9 1.2.3 Challenges in collection of high-quality data 10 1.2.4 Deep uncertainties of future trends 12 1.3 SUMMARY 13 Chapter 2 Objectives 17 Chapter 3 Material and Methods 19 3.1 DATA COLLECTION 20 3.2 MECHANISTIC MODELING: STOFFENMANAGER® 21 3.2.1 Computational Experiment: Simulation model based on older versions of Stoffenmanager® 22 3.3 BAYESIAN MODELING 28 3.3.1 Specifying prior distributions using Stoffenmanager® 29 3.3.2 Bayesian model diagnostics 30 3.4 CLUSTER ANALYSIS: DETERMINATION OF DEGREE OF SIMILARITY USING GOWER DISTANCE AND K-MEDOIDS CLUSTERING ALGORITHM ANALYSES 30 3.4.1 Gower Distance 31 3.4.2 k-Medoids Clustering Algorithm 32 3.4.3 Incorporating existing exposure data with Stoffenmanager® as prior distribution 33 3.4.4 Model Evaluation: Lack of Agreement and Bias 35 3.5 MANY-OBJECTIVE ROBUST DECISION MAKING 37 3.5.1 Robust Decision-Making 37 3.5.2 Characterizing Robustness 38 3.5.3 Problem formulation 39 3.5.4 Identify Optimal Strategies: Many Objective Evolution Algorithms 43 3.5.5 Identify Vulnerabilities: Scenario discovery by patient-rule induction method (PRIM) 46 3.6 SOFTWARE AND TOOLS USED 47 Chapter 4 Results and Discussion 49 4.1 PART. 1 EVALUATING CURRENT TOOLS FOR OCCUPATIONAL HEALTH BY DESIGN: A CASE STUDY OF HEALTH RISK ASSESSMENT OF PHOTORESISTS USED IN AN OPTOELECTRONIC SEMICONDUCTOR FACTORY 49 4.1.1 Hazard identification of photoresists 51 4.1.2 The exposure scenario 52 4.1.3 Results 54 4.1.4 Discussion 66 4.1.5 Summary 70 4.2 PART. 2 EXAMINING SIMILARITY OF EXPOSURE SCENARIOS TO INFORM EXPOSURE MODELING ESTIMATES: A DEMONSTRATION USING CLUSTER ANALYSIS AND BAYESIAN MODELING 73 4.2.1 Results 74 4.2.2 Discussion 85 4.3 PART. 3 OCCUPATIONAL HEALTH BY DESIGN WITH ARTIFICIAL INTELLIGENCE: USE OF EXPLORATORY MODELING FOR AUTOMATED WORKPLACE DESIGN 91 4.3.1 Results 91 4.3.2 Discussion 98 Chapter 5 Conclusion and Recommendation 107 References 111 Appendix 1 5.1 PART 1 1 5.1.1 Dose-response and reference values 1 5.1.2 BMDS 3.0 Analysis for Cresol and Phenol 5 5.1.3 The Stoffenmanager® Web Application 10 5.1.4 Bayesian Model 15 5.1.5 Part 1. Appendix References 18 5.2 PART 2 20 5.2.1 Comparisons of inverse-gamma, gamma and log-normal distributions as the prior distributions for the variance components 20 5.2.2 Part 2. Appendix References 26 5.3 PART 3 27 | |
dc.language.iso | en | |
dc.title | 職業衛生設計: 建立人工智慧應用之先期研究 | zh_TW |
dc.title | Occupational Health by Design: A Development Towards Applications of Artificial Intelligence | en |
dc.type | Thesis | |
dc.date.schoolyear | 108-1 | |
dc.description.degree | 博士 | |
dc.contributor.coadvisor | 鄭尊仁(Tsun-Jen Cheng) | |
dc.contributor.oralexamcommittee | 林菀俞(Wan-Yu Lin),陳志傑(Chih-Chieh Chen),何俊傑(Jiune-Jye Ho),陳主智(Chu Chih Chen),陳振和(Jeng-Ho Chen) | |
dc.subject.keyword | 職業衛生設計,人工智慧,StoffenmanagerR,群聚分析,穩健決策, | zh_TW |
dc.subject.keyword | Occupational Health by Design,Artificial Intelligence,StoffenmanagerR,Cluster analysis,Robust decision-making, | en |
dc.relation.page | 125 | |
dc.identifier.doi | 10.6342/NTU202000025 | |
dc.rights.note | 有償授權 | |
dc.date.accepted | 2020-01-06 | |
dc.contributor.author-college | 公共衛生學院 | zh_TW |
dc.contributor.author-dept | 環境與職業健康科學研究所 | zh_TW |
顯示於系所單位: | 環境與職業健康科學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-109-1.pdf 目前未授權公開取用 | 5.6 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。