1 College of Aeronautical Engineering, Civil Aviation University of China, Tianjin 300300, China
2 China Southern Airlines Co., Ltd., Engineering Technology Branch, Zhengzhou 411000, China
Show less
文章历史+
Received
Published
Revised
2025-03-05
2026-01-18
2025-10-24
Online Date
2026-09-20
PDF (3883K)
摘要
随着民用航空业的快速发展,民机运维数据的深度解析成为提升飞行安全与维修效能的关键路径。针对非结构化运维文本中要素实体识别与用户意图理解的双重需求,提出了一种多任务协同分析框架:首先,通过RoBERTa(robustly optimized BERT approach)模型进行预训练捕捉语义信息,接着通过双向长短期记忆网络(bidirectional long short-term memory,BiLSTM)捕获时序依赖特征,并利用条件随机场(conditional random field,CRF)优化标签转移约束并输出,实现对维修记录中部件、故障、操作等实体的高精度识别;其次,面向智能问答系统的语义理解需求,对比分析BERT、RoBERTa及Qwen-Max预训练语言模型在运维意图识别任务中的适应性,基于性能与计算效率的平衡性验证,确立RoBERTa为意图分类基座模型。实验表明,所提框架在实体抽取与意图识别任务中均显著优于传统基准模型,其双任务协同机制为民机运维知识图谱构建与智能问答服务提供了数据支撑,具有较高的应用潜力和实用价值。
Abstract
With the rapid development of the civil aviation industry, the in-depth analysis of civil aircraft O&M (operation and maintenance) data has become a critical pathway to enhance flight safety and maintenance efficiency. To address the dual demands of entity recognition from unstructured O&M texts and user intent understanding, a multi-task collaborative analysis framework was proposed. First, semantic information was captured through pre-training with the RoBERTa (robustly optimized BERT approach) model. Subsequently, BiLSTM (bidirectional long short-term memory) networks were employed to extract temporal dependency features. The CRF (conditional random field) was then applied to optimize label transition constraints and generate the final label sequence, achieving high-precision recognition of entities such as components, faults, and operations in maintenance records. Second, to meet the semantic understanding requirements of intelligent Q&A (question and answering) systems, a comparative analysis was conducted on the adaptability of the BERT, RoBERTa, and Qwen-Max pre-trained language models in the task of operation and maintenance intent recognition. Through performance-efficiency trade-off validation, RoBERTa is selected as the backbone model for intent classification. Experimental results demonstrate that the proposed framework significantly outperforms traditional baseline models in both entity extraction and intent recognition tasks. Its dual-task collaboration mechanism provides structured data support for constructing civil aircraft O&M knowledge graphs and intelligent Q&A services, exhibiting substantial application potential and practical value.
MA Chao, ZOU Zi-yang, FAN Jie.
Identification of Corpus Elements for Civil Aircraft Operation and Maintenance Data Based on RoBERTa Model[J].
Science Technology and Engineering, 2026, 26(2): 701-707 DOI:10.12404/j.issn.1671-1815.2501446
RoBERTa是一种基于Transformer的预训练语言模型[11],具有出色的上下文建模能力,特别是在长文本的语义理解方面表现优异。其能够有效捕捉文本中的复杂语义关系,并在多个自然语言处理(nature language processing,NLP)任务中提供丰富的特征表示,在对语境的理解上表现更加稳健[12]。Transformer双向编码器用于编码文本字符和缩略语,通过计算词符间的跨维度关联权重,动态捕捉语言单元间的语义关联强度与结构依存特性,以获取词汇之间的相互关系和内部结构,从而对样本句子进行编码,其计算式为
ZhaoChendi, SunYi, YangXiufeng. Fault diagnosis of aero-piston engine based on condition monitoring data[J]. Science Technology and Engineering, 2023, 23(36): 15732-15738.
XiaCunjiang, ZhanYuyou. Research status of gas path fault diagnosis for aeroengine basedon data-driven[J]. Science Technology and Engineering, 2022, 22(5): 1741-1750.
NiuWei, ChengJuan, ZhaoJianping. Data-driven collaborative analysis of aviation equipment and intelligent support decision-making method[J]. Science Technology and Engineering, 2021, 21(24): 10531-10535.
[7]
ErcanK, GülfemT, ÖzalpV. Safety management system and risk-based approach in aviation maintenance: a systematic literature review[M]. Berlin: Springer, 2024.
[8]
StantonI, MunirK, IkramA, et al. Predictive maintenance analy-tics and implementation for aircraft: challenges and opportunities[J]. Systems Engineering, 2023, 26: 216-237.
[9]
XiaoX, ShengB, FuG, et al. Construction of knowledge graph for air compressor fault diagnosis based on a feature-fusion RoBERTa-BiLSTM-CRF model[J]. Actuators, 2024, 13.DOI: 10.3390/act13090339.
XuJian, ZhangJing, SongLingyu, et al. Fine-grained sentiment analysis of online reviews based on RoBERTa-BiLSTM-CRF[J]. Systems Engineering: Theory & Practice, 2023, 43(12): 3519-3536.
MiaoJiachi, ChenYing, ShengLong, et al. Chinese address resolution method based on RoBERTa-BiLSTM-Self Attention-CRF[J]. Journal of the Hebei Academy of Sciences, 2024, 41(6): 25-34.
XiangQian, CaiZangtai, LiCuo. Study on identification of Tibetan news element based on RoBERTa-BiLSTM-CRF[J]. Plateau Science Research, 2024, 8(4): 108-114.
HaoKuangong, DongBing, WuYue, et al. Airport class based on BERT-BILSTM-CRF model Chinese navigation notice element entity recognition[J]. Science Technology and Engineering, 2024, 24(10): 4182-4188.
TangWeiguang, ChenYong, YaoJian. Text retrieval sorting method based on RoBERTa model[J]. Computer and Network, 2024, 50(5): 448-455.
[20]
FengP, ZhangB, YangL, et al. Entity-alignment interaction model based on Chinese RoBERTa[J]. Applied Sciences, 2024, 14(14): 61-62.
[21]
LiaoW, ZengB, YinX, et al. An improved aspect-category sentiment analysis model for text sentiment analysis based on RoBERTa[J]. Applied Intelligence, 2020, 15(51): 3522-3533.
[22]
LiaoJ, ShiH. Research on joint extraction model of financial product opinion and entities based on RoBERTa[J]. Electronics, 2022, 11(9): 13-45.
[23]
LakA J, BoostaniR, AleniziF A, et al. RoBERTa, ResNeXt and BiLSTM with self-attention: the ultimate trio for customer sentiment analysis[J]. Applied Soft Computing, 2024, 164. DOI: 10.1016/j.asoc.2024.112018.
[24]
HochreiterS, SchmidhuberJ. Long short-termmemory[J]. Neural Computation, 1997, 9(8): 1735-1780.
[25]
YangG, XuH. A residual BiLSTM model for named entity recognition[J]. IEEE Access, 2020(99): 1-11.
[26]
HameedZ, Garcia-ZapirainB. Sentiment classification using a single-layered BiLSTM model[J]. IEEE Access, 2020, 8(99): 73992-74001.
[27]
LeY, JunL, MingC, et al. Memristor-based circuit design of BiLSTM network[J]. Neural Networks, 2025, 181.DOI: 10.1016/j.neunet.2024.106780.
ZhaoYitao, LiZhao, LiuXinglong, et al. Load recognition method based on convolutional neural network and attention mechanism[J]. Electric Power Engineering Technology, 2025, 44(1): 227-235.
WangZichen, QuYouli. A partition inverted index compression algorithm based on CRF[J]. Computer and Modernization, 2024(2): 36-42, 55.
[32]
XiangJ, XuG, MaC, et al. End-to-end learning deep CRF models for multi-object tracking deep CRF models[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2021(1). DOI: 10.1109/tcsvt.2020.2975842.