基于迁移学习图像识别的桥梁监测数据异常检测方法
Bridge Health Monitoring Data Anomaly Detection Method Based on Transfer Learning and Image Recognition
殷鹏程 1谭曼丽莎 2曹阳梅 1单德山2
作者信息
- 1. 中铁第四勘察设计院集团有限公司,湖北 武汉 430063;中国铁建股份有限公司桥梁工程实验室,湖北 武汉 430063
- 2. 西南交通大学 土木工程学院,四川 成都 610031
- 折叠
摘要
为了改善桥梁结构健康监测数据多模式异常检测难以全覆盖的问题,提出了基于迁移学习图像识别技术的数据异常检测方法.首先,绘制结构应变和温度数据时程图,统计分析数据日周期变化规律,并标记异常类型作为训练集;通过图像仿射变换实现数据增强,减少网络对不平衡数据集的过拟合.其次,以AlexNet预训练网络为基础,选择迁移学习策略,建立多个对比模型用于分析不同优化算法、初始学习率、Dropout取值和结构轻量化对网络识别结果的影响;以模型分类准确率、训练速率和内存需求作为评价指标,获得性能优越的监测数据异常检测深度网络模型,模型对试验数据的异常分类准确率可达95.5％,体量可缩减94.7％.最后,基于优化模型编译数据异常检测模块,并以某斜拉桥一个月的监测数据为例进行验证,验证集准确率为98.0％.所提方法准确率高、计算速度快、内存需求小,方便应用于桥梁健康监测数据预处理平台.
Abstract
To address the issue of incomplete coverage in multi-mode anomaly detection of bridge health mo-nitoring data,a method based on transfer learning and image recognition is proposed.Firstly,the daily period-ic variation rules of bridge structural strain and temperature monitoring data are statistically analyzed,and the time-history graphs are labeled with abnormal types as the training set.Data augmentation is achieved through image affine transformation to reduce overfitting of the network to unbalanced datasets.Subsequently,multiple comparative models are constructed based on the pre-trained AlexNet network,employing a transfer learning strategy to analyze the effects of various optimization algorithms,initial learning rates,dropout values,and network lightweight structures on recognition results.The model classification accuracy,training rate,and memory requirement are evaluated to obtain a high-performance monitoring data anomaly detection deep net-work model.The model achieved an impressive accuracy of 95.5％ for anomaly classification on experiment data,with a reduction in volume of 94.7％.Finally,the data anomaly detection module is compiled and vali-dated using one-month monitoring data of a cable-stayed bridge based on the optimized model.The proposed method exhibits high accuracy,fast calculation speed and low memory requirements with convenience for ap-plication in bridge health monitoring data preprocessing platforms.
关键词
结构健康监测/数据异常检测/迁移学习/图像识别/AlexNetKey words
structural health monitoring (SHM)/data anomaly detection/transfer learning/image recognition/AlexNet引用本文复制引用
基金项目
国家自然科学基金项目(51978577)
中铁第四勘察设计院集团有限公司科技研究开发计划项目(2022K086)
出版年
2024
NETL
NSTL
万方数据