Data Anomaly Diagnosis And Reconstruction Based On Deep Learning For Structural Health Monitoring

In recent years,structural health monitoring technology has been rapidly developed and applied worldwide.Large-scale health monitoring systems have been installed on many large bridges in China.These monitoring systems have accumulated a large amount of monitoring data.However,the service environment of large bridges is harsh,and most of the modules of their health monitoring systems are working outdoors.It is difficult to avoid anomalous data due to hardware and software failures.The large-scale analysis of real bridges’monitoring data shows that there are numerous anomalous data in the monitoring system.These anomalous data are randomly distributed in the monitoring data,leading to false warnings of the monitoring systems,and also seriously affecting the effectiveness of data analysis and health diagnosis accuracy.How to effectively diagnose,identify and reconstruct the anomalous data is a scientific problem that needs urgent research.The main contents includes:Anomaly data diagnosis method based on stacked autoencoder deep neural network.The main idea of the method is to imitate the human bio-visual information acquisition and logical thinking process.Time series signals are segmented into grayscale images,modeled as an image classification problem.Subsequently,a deep neural network is constructed and trained using a greedy layer-wise training method based on a labeled dataset consisting of randomly selected images and their manual labels.The feasibility of the method and the classification accuracy of the neural network are verified using the annual acceleration data of a long-span cable-stayed bridge structural health monitoring system.A time-and frequency-information fused anomaly data diagnosis method based on convolutional neural network.The raw time series data are segmented by windowing and visualized in the time and frequency domains separately.Two-channel images are constructed using time-domain response and frequency-domain response,and manually labeled according to the two-channel image features.Subsequently,a convolutional neural network is constructed for data anomaly classification.The strong imbalance between different types of data anomalies in real-world engineering is considered to construct balanced and imbalanced training sets,and the influence of training set size on data anomaly diagnosis is considered.Acceleration monitoring data of a long-span cable-stayed bridge is used for validation and comparative analysis with a single information source approach based on computer vision and deep neural networks.Machine learning based compressive sensing method for data reconstruction.The computation process of compressive sensing is modeled as a standard supervised machine learning regression task.The basis matrix and the compressive sensing sampled signal are set as the input and output of the network as prior knowledge,respectively,and the basis coefficient matrix is embedded as the parameters of a layer of the network,and the conventional objective function of compressive sensing is set as the loss function of the neural network.These basis coefficients are l₁ regularized and optimized by a general back-propagation optimization algorithm.Based on the artificially generated sinusoidal data and the field test wireless data of long-span suspension bridges,the data reconstruction capability of the method is investigated and compared with the classical l₁method.A continuous data missing recovery method based on a group sparsity-aware convolutional neural network.The data recovery task is described as a matrix completion optimization problem,and subsequently the group sparsity optimization problem for data reconstruction is established and the feedforward computation of the optimization problem is established in a convolutional neural network.The convolutional kernels,ie,the basis coefficients,is optimized by gradient-based back-propagation algorithms for minimizing the regression error and group sparsity.The reconstruction is readily obtained after optimization.The missing data recovery capability of the method is tested using simulated data and real bridges’data,respectively.Also the method was compared with the convolutional neural network based on the l₁regularization.