基于可解释性深度学习的传感器异常数据诊断  

Diagnosis of abnormal sensor data based on interpretable deep learning

作　　者：童浩[1,2] 阮先虎 林峰 刘朵 TONG Hao;RUAN Xianhu;LIN Feng;LIU Duo(College of Transportation,Southeast University,Nanjing 211189,China;The Expressway Operation&Administration Center of Jiangsu Province,Nanjing 210009,China;State Key Laboratory of Safety,Durability and Healthy Operation of Long Span Bridges,JSTI Group Co Ltd,Nanjing 211112,China)

机构地区：[1]东南大学交通学院,江苏南京211189 [2]江苏省高速公路经营管理中心,江苏南京210009 [3]长大桥梁安全长寿与健康运维全国重点实验室苏交科集团股份有限公司,江苏南京211112

出　　处：《传感器与微系统》2024年第12期41-44,共4页Transducer and Microsystem Technologies

基　　金：江苏交控科研项目基金资助项目(2023JKY20)。

摘　　要：针对桥梁健康监测中多源传感器数据异常诊断可解释性较差和效率低的问题,提出了一种基于特征可视化的可解释性卷积神经网络(CNN)数据异常检测方法。充分考虑异常模式和多源传感器类型的完整性,结合数据扩充方法,构建了多源监测数据异常模式库,同时,基于CNN展开异常特征提取,利用特征和类激活图(CAM)可视化的方法,深入分析异常类型特征,从而实现对深度学习网络的解释分析。实验结果表明:考虑多源传感器类型可以充分挖掘数据中的有效信息,模型的整体准确率达到了99.37%,所有异常模式的查全率与查准率均超过96%。该方法还能够捕捉时间序列深度学习分类模型的特征学习过程,为桥梁结构连续性监测数据的分析和预警提供了先决条件。Aiming at the problem of poor interpretability and low efficiency in multisource sensor data anomaly diagnosis for bridge health monitoring,an interpretable convolutional neural network(CNN)data anomaly detection method based on feature visualization is proposed.Fully considering the completeness of anomaly patterns and multisource sensor types,a library of multisource sensor monitoring data anomaly patterns is constructed by integrating data augmentation methods.Concurrently,anomaly features are extracted using a CNN and feature and class activation map(CAM)visualization method is utilized to deeply analyze the features of anomaly types,thereby achieving an interpretive analysis of the deep learning network.Experimental results indicate that considering multisource sensor types can fully exploit the effective information in the data,the overall accuracy of the model reaches 99.37%,and both the recall and precision rates of all anomaly patterns exceed 96%.This method can also capture the feature learning process of timeseries deep learning classification models,providing prerequisites for the analysis and early warning of continuous monitoring data of bridge structures.

关 键 词：深度学习 卷积神经网络 特征可视化 异常诊断 健康监测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U447[自动化与计算机技术—计算机科学与技术]