基于多模态特征融合的隧道渗水异常检测方法  

Tunnel water leakage anomaly detection method based on multi-modal feature fusion

作　　者：朱厚喜 曹伍富[1] 李克飞 伊建峰 刘少华 李洪珏 ZHU Houxi;CAO Wufu;LI Kefei;YI Jianfeng;LIU Shaohua;LI Hongjue(Beijing MTR Construction Administration Company Limited,Beijing 100068,China;Beijing Infrastructure Investment Company Limited,Beijing 100101,China;Beijing Metro Consultancy Corporation Limited,Beijing 100068,China;National Engineering Laboratory for Urban Rail Transit System Safety Assurance Technology,Beijing 100068,China;Beijing Key Laboratory for Urban Rail Transit Fully Automated Operation System and Safety Monitoring,Beijing 100068,China;School of Astronautics,Beihang University,Beijing 102206,China)

机构地区：[1]北京市轨道交通建设管理有限公司,北京100068 [2]北京市基础设施投资有限公司,北京100101 [3]北京城市轨道交通咨询有限公司,北京100068 [4]城市轨道交通系统安全保障技术国家工程实验室,北京100068 [5]城市轨道交通全自动运行系统与安全监控北京市重点实验室,北京100068 [6]北京航空航天大学宇航学院,北京102206

出　　处：《计算机应用》2023年第S02期276-284,共9页journal of Computer Applications

基　　金：北京市基础设施投资有限公司科研项目(2021-18-02-11)。

摘　　要：地铁隧道渗水对隧道施工安全造成严重威胁,目前隧道渗水异常检测面临以下问题:单一模态渗水异常检测方法抗干扰能力差;忽略了隧道渗水的动态变化特性;隧道渗水检测数据集获取难度大。为此,提出一种基于多模态观测数据的隧道渗水异常检测(MFF-WLAD)方法。为降低单一模态下干扰因素对渗水异常检测的影响,该方法融合可见光模态和红外热成像模态的观测数据,感知并学习潜在渗水区跨模态差异化成像特征;同时,构建渗水区时空演化先验指导的显著特征抽取模块,并与长短期记忆(LSTM)相结合,以提升MFF-WLAD方法对视频序列中渗水区域动态演化现象的识别能力。为了解决公开数据集缺乏的问题,建造隧道模拟场景并搭建多模态数据采集系统,构建多模态隧道渗水异常检测数据集。实验结果表明,在测试集的所有渗水异常事件上,MFF-WLAD方法实现了100%的检测成功率,虚警率为2%,与近年优秀对比方法AP-Model、ASTNet和LGN-Net相比,MFF-WLAD的有效检测率至少提高了27个百分点,同时虚警率也低于对比方法至少0.4个百分点,验证了MFF-WLAD方法的有效性。Metro tunnel water leakage poses a serious threat to tunnel construction safety.Currently,there are three problems in detection of abnormal tunnel water leakage:poor anti-interference ability of single-modal water leakage detection methods,neglect of the dynamic characteristics of tunnel water leakage,and difficulty in obtaining tunnel water detection datasets.To address these issues,a Multimodal Feature Fusion-based Water Leakage Anomaly Detection(MFF-WLAD)method was proposed.To reduce the influence of interference factors on water leakage detection in single-modal scenarios,by integrating visible light and infrared thermal imaging modalities,the cross-modal differential imaging features of potential water leakage regions were perceived and learned;meanwhile,a spatiotemporally evolved prior-guided salient feature extraction module was onstructed for water leakage regions and combined with Long Short-Term Memory(LSTM)to improve the ability for recognizing the dynamic evolution phenomena of water leakage regions in video sequences.To overcome the lackness of publicly available datasets,a tunnel simulation scenario was constructed,and a multi-modal data acquisition system was built to create a dataset for multi-modal tunnel water leakage detection.Experimental results show that MFFWLAD achieves a 100% detection rate with a false alarm rate of 2%for all water leakage events in test set.Compared to recent excellent comparative methods including AP-Model,ASTNet and LGN-Net,the successful detection rate of MFFWLAD has been improved by at least 27 percentage points,and the false alarm rate is also lower by at least 0.4 percentage points,verifying the effectiveness of the proposed multi-modal tunnel water leakage detection method.

关 键 词：渗漏水检测 多模态学习 红外热成像 视频异常检测 循环神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]