基于SCN与贝叶斯优化LSTM的结构健康监测异常振动数据半监督诊断方法OA
Semi-supervised diagnostic method for abnormal vibration data in SHM via SCNs and Bayesian-optimized LSTM
结构健康监测(SHM)系统通常在恶劣环境条件中运行,不可避免地会出现离群值、偏差和非线性漂移等数据异常.因此,本文提出了基于长短期记忆网络(LSTM)的异常数据半监督诊断方法.该方法利用改进的基于密度的空间聚类方法(DBSCAN),自动为加速度数据分配标签,解决了传统监督学习中标注成本高的问题.在此基础上,采用散射卷积网络(SCN)对原始加速度信号进行多尺度分解,利用前3阶小波散射系数作为训练数据,大幅降低了冗余信息(数据压缩率达97.01%),提升了LSTM的计算效率与分类精度.最后,通过苏州科技馆(SSTM)的振动监测数据验证了该方法的有效性.结果表明,当改进的DBSCAN算法的最小包含点数M和扫描半径ε分别设置为5和0.1时,可有效实现振动数据的高精度标注.此外,利用贝叶斯优化对LSTM模型的超参数进行调优后,漏检率降低至0%,误检率降低至1.4%,诊断准确率提升至96%.该方法在保证诊断精度的同时显著提升了计算效率,为结构健康监测数据异常诊断提供了有效参考.
Structural health monitoring(SHM)systems fre-quently operate under harsh environmental conditions,in-evitably leading to data anomalies such as outliers,devia-tions,and drifts.This study develops a semi-supervised learning framework leveraging long short-term memory(LSTM)networks to systematically diagnose and catego-rize abnormal acceleration signals.The proposed method in-tegrates an enhanced density-based spatial clustering of ap-plications with noise(DBSCAN)algorithm to automati-cally assign labels to acceleration data,substantially reduc-ing the need for labor-intensive manual annotation of raw da-tasets.A scattering convolution network(SCN)is em-ployed to decompose raw acceleration signals.The first three orders of wavelet scattering coefficients are extracted as the training input features for the subsequent model.This approach achieves a 97.01%reduction in feature dimension-ality compared with the original signal,thereby optimizing the computational efficiency and diagnostic accuracy of the LSTM model.The effectiveness and applicability of the methodology are rigorously validated through a case study involving the Suzhou Science and Technology Museum(SSTM).Experimental results confirm that when the mini-mum points threshold(M)and scanning radius(ε)of the en-hanced DBSCAN algorithm are configured at 5 and 0.1,re-spectively,the unsupervised labeling module attains high data annotation accuracy.Furthermore,Bayesian optimiza-tion is utilized to fine-tune the hyperparameters of the LSTM model,yielding a missed detection rate of 0%,a false detection rate of 1.4%,and an overall diagnostic accu-racy of 96%.These findings demonstrate the proposed method significantly enhances computational efficiency while preserving diagnostic accuracy,providing an effec-tive reference for anomaly diagnosis in measurement data.
王亚伟;祝青鑫;王浩;傅文炜
东南大学混凝土及预应力混凝土结构教育部重点实验室,南京 211189东南大学混凝土及预应力混凝土结构教育部重点实验室,南京 211189||上海理工大学环境与建筑学院,上海 200093东南大学混凝土及预应力混凝土结构教育部重点实验室,南京 211189苏州科技大学土木工程学院,苏州 215011
建筑与水利
半监督诊断方法散射卷积网络长短期记忆网络基于密度的空间聚类方法结构健康监测
semi-supervised diagnostic methodscattering convolution networklong short-term memorydensity-based spatial clusteringstructural health monitoring
《东南大学学报(英文版)》 2026 (3)
298-305,8
The National Natural Science Foundation of China(No.52408342,52378291),Shanghai Pujiang Program(No.23PJ1410000).
评论