Abnormal Data Detection And Reconstruction Of Structural Health Monitoring Based On Deep Learnin

Complete and high quality structural health monitoring(SHM)data is very important for an effective assessment of structural safety condition.However,due to the long-term harsh working environment,SHM’s sensing system fails frequently,resulting in a large number of missing and abnormal values in the measurement data,and the resulting incomplete measurement information seriously affects the validity of structural condition assessment.At this stage,due to the huge volume of monitoring data,manual identification of abnormal data has little effect,making data problems hinder the development and application of SHM for a long time.For this reason,with the wave of development of artificial intelligence,it is significant to improve data integrity and the reliability of structural state assessment results by adopting intelligent algorithms to learn the characteristics of anomalous data,identify anomalous values efficiently,accurately and automatically from a large amount of long-term monitoring data,and then reconstruct the real measured values of anomalous time periods based on incomplete measurement information.In this paper,based on the existing research,the following research work is conducted around the low recognition accuracy due to the imbalance of training data in intelligent data anomaly detection,and the low utilization of relevance of time dimension in data reconstruction:(1)Firstly,the relevant networks and principles of deep learning network architectures for anomaly detection and reconstruction of SHM data are introduced,including convolutional neural networks,recurrent neural networks and self-attention mechanism.Convolutional neural network mainly consists of convolutional layer and pooling layer.The stacked convolution layer continuously extracts high-dimensional representative features of the input data through successive convolution operations;while the pooling layer can effectively reduce computation,filter key features and smooth data features by downsampling the feature map.Long short-term memory network is a modified model of recurrent neural network.Long and short term memory networks rely on the design of neuron ’gate’ structures to overcome the gradient explosion or disappearance and the major long term dependency problems often encountered in recurrent neural networks,allowing the network to learn temporal relationships over large spans of data.The self-attention mechanism is an important component of the anomaly data detection network in this paper,whose autocorrelation-like computational operation takes a global view and allows the network to assign greater computational weight to important features that are highly correlated.(2)A reasonable network structure can effectively improve the network performance and training efficiency.In this chapter,TDNet(Transformer Enhanced Densely Connected Neural Network)and bi-directional long short-term memory network are specially designed to solve the problem of anomaly data detection and data reconstruction,which are one-dimensional time-series data and imply structural vibration characteristics of SHM monitoring data.TDNet is embedded with Dense Net with high-density information connectivity,which has powerful one-dimensional signal feature extraction capability to quickly and accurately capture highdimensional data features of structural acceleration response while also filtering out measurement noise interference using bottleneck structures.Considering that the anomalous data appear for a short time but with obvious features,TDNet designs a Transformer-based encoder,aiming to strengthen the network’s attention to the local features of the data with the help of the attention mechanism.Previous methods of reconstructing data only rely on the transfer relationship between structural responses in different acquisition channels,but ignore the vibration data timing pattern.To this end,this study makes use of the long-and short-term memory network which is good at processing and extracting data timing relations,and designs a data reconstruction model based on the long-and short-term memory network and attention mechanism based on the existing network architecture based on spatial relations.(3)The small volume of anomalous data compared with normal data and the large difference in the frequency of various types of anomalous data are typical sample imbalance problems for machine learning.Therefore,this paper proposes an anomaly detection method for civil engineering structural SHM data based on data-level and deep learning techniques.The method includes two stages of data set category balancing and anomaly detection.In the first stage,anomalous data samples are generated on the basis of normal data based on the features of the anomalous data to balance the training dataset.The balanced dataset is used to train the specially designed network TDNet so that the network learns high-level abstract features of different anomalous data and establishes the mapping relationship between features and corresponding anomaly classes.In this section,numerical simulations based on the steel frame model are conducted to demonstrate the excellent feasibility and interference resistance of the data anomaly detection method.Subsequently,this paper verifies the applicability of the method in practical data anomaly classification by taking the actual measured data of Guangzhou Tower as an example.(4)Eliminating abnormal data segments and reconstructing the real data based on the correct identification of abnormal data can effectively compensate the influence of incomplete structural information on structural health state assessment.In this paper,in response to the low efficiency of existing data reconstruction methods in utilizing the correlation of vibration data in time dimension,a two-way long and short term memory network based on reconstructing the real measurement data is adopted.This study validates the method with a nonlinear finite element model using the structural response under seismic excitation.The validation results indicate that the reconstructed response of the network can be accurately matched with the real measurements in both time and frequency domains,with only small differences in the vibration amplitudes of the time series.In addition,the training data of the network is a nonlinear,smallsample dataset generated by only four earthquakes,indicating that the data reconstruction network has a powerful ability to extract features and predict nonlinear structural responses.