空管不正常事件风险信息抽取与识别方法研究  

Research on risk identification methods for air traffic control irregular events using information extraction

作　　者：王洁宁[1] 王帅翔 孙禾[2] WANG Jiening;WANG Shuaixiang;SUN He(College of Air Traffic Management,Civil Aviation University of China,Tianjin 300300,China;Key Laboratory of Wide Area Surveillance and Safety Control Technology for Civil Aviation Flights,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学空中交通管理学院,天津300300 [2]中国民航大学民航航班广域监视与安全管控技术重点实验室,天津300300

出　　处：《安全与环境学报》2025年第4期1444-1454,共11页Journal of Safety and Environment

基　　金：国家自然科学基金项目(U2133207);中国民航大学民航航班广域监视与安全管控技术重点实验室开放基金项目(202204)。

摘　　要：目前,空管各类安全管理信息化平台积累了大量非结构化文本数据,但未得到充分利用,为了挖掘空管不正常事件中潜藏的风险,研究利用收集的四千余条空管站不正常事件数据和自构建的4836个空管领域专业术语词,提出了一个基于空管专业信息词抽取的双向编码器表征法和双向长短时记忆网络的深度学习模型(Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory,BERT-BiLSTM)。该模型通过对不正常事件文本进行信息抽取,过滤其中无用信息,并将双向编码器表征法(Bidirectional Encoder Representations from Transformers,BERT)模型输出的特征向量序列作为双向长短时记忆网络(Bidirectional Long Short-Term Memory,BiLSTM)的输入序列,以对空管不正常事件文本风险识别任务进行对比试验。试验结果显示,在风险识别试验中,基于空管专业信息词抽取的BERT-BiLSTM模型相比于通用领域的BERT模型,风险识别准确率提升了3百分点。可以看出该模型有效提升了空管安全信息处理能力,能够有效识别空管部门日常运行中出现的不正常事件所带来的风险,同时可以为空管安全领域信息挖掘相关任务提供基础参考。This study aims to enhance risk identification in Air Traffic Control(ATC)irregular events by utilizing the extensive unstructured textual data gathered across various ATC safety management platforms.To identify potential risks associated with these events,we initially collected over 4000 records of ATC station irregular occurrences.After tokenizing the data with spaCy,we developed a lexicon comprising 4836 ATC specific terms.We subsequently tagged various named entities with their corresponding parts of speech,categorizing them into ten classes:person,law,facility,product,event,effect,guidance,chain,problem,and measure.We annotated the texts in the ATC irregular event training set with risk categories,conducted information extraction to isolate relevant named entities,and eliminated extraneous information.Building on these resources,we proposed a BERT BiLSTM pre-trained language model.The feature vector sequences generated from the irregular event texts by the BERT model were fed into a Bidirectional Long Short-Term Memory(BiLSTM)network for risk identification analysis.Finally,we conducted a comparative experiment for risk identification in ATC irregular event texts,evaluating a general-domain BERT model,a BERT model retrained following information extraction,and a BERT BiLSTM model trained after information extraction.The experimental results indicate that,compared to the standard BERT model,the accuracy of the BERT model trained with information extraction has improved by 0.6 percentage points.Notably,there were significant enhancements in identifying fatigue-related risks and operational management deficiencies,with F 1 scores increasing by 27.58%and 17.4%,respectively.This demonstrates that ATC domain-specific texts processed through information extraction are more effective for risk identification tasks.Additionally,the enhanced BERT BiLSTM model demonstrated significant improvements over the general-domain BERT model in identifying four types of risks:insufficient personnel skills,fatigue-related risks,in

关 键 词：安全工程 双向编码器表征法 双向长短时记忆网络 空管不正常事件 风险识别 

分 类 号：X949[环境科学与工程—安全科学]